Volume 105

Number 1

Spring 2019

Journal of the

WASHINGTON

ACADEMY OF SCIENCES

WM ASRS EME ANIMIST ES! SF OVI CL recast ce- ences teconeFeschansténsetbsasuucocesonaisassndasecodasgesseastensdeshonssdtondeststebinsbes il

arched RM ENN RUE SPNERES SE CLLL CONS 92001 oo 20 ooo co scan ath ocsuasscoiaiescasotosshedsobsSasucdsscasitnscbevedlsagiuvsseaviedobosvinses iil

Meme VASEROTIORAY, PIGHIOSG. BYLC. cencsecccscossecsssiscnssPisoencsnssvssoacnsssoscessvostssstssactenssssconcnenseiesseecuosconssie 1

The Antigenic Shift or Drift of the Influenza Virus J. PQULONIS. .0......eecseecseessessseeeseesteesieeees 7

In Vitro Antibacterial Activity of Garlic and Tea Tree Oil S. Godinez et al. ov... 13

Contextual Label Smoothing with a Phylogenetic Tree ™. J. Trammell et al............... 23

NaN MET MT SES ESMR eA ENOONN roof essere eco eh avere toned sch vcotanscovest deena vaceistnsavtnnatsRen@eeceseecepessalies 47

Se amet RMN AMIE PM EES ENGN ESS ccc Pci yo seek avs cS ccc anc ancy sos ysansectezssebenessit deecdanstimeviavancstine 48

pO ENYE SLD SEE aT ee oe ae or en 49

PTR AEE SOCIETIES ANG DELO ATCS 0. isn ec aass de csaccesnscssesasesseneecnscsdugsisudososecebecnbeccssanasieaipnantbostosasecdons

ISSN 0043-0439 Issued Quarterly at Washington DC

Washington Academy of Sciences

Founded in 1898

BOARD OF MANAGERS

Elected Officers

President

Judy Staveley

President Elect

Mina Izadjoo

Treasurer

Ronald Hietala

Secretary

Poorva Dharkar

Vice President, Administration

Terry Longstreth

Vice President, Membership

Ram Sriram

Vice President, Junior Academy

Paul Arveson

Vice President, Affiliated Societies

Gene Williams

Members at Large

Joanne Horn

David Torain

Barbara Ransom

Noriko Behling

Lisa Frehill

Mike Cohen

Past President

Mina Izadjoo

AFFILIATED SOCIETY DELEGATES

Shown on back cover

Editor of the Journal

Sethanne Howard

Journal of the Washington Academy of

Sciences (ISSN 0043-0439)

Published by the Washington Academy of

Sciences

email: wasjournal@washacadsci.org

website: www.washacadsci.org

The Journal of the Washington Academy

of Sciences

The Journal is the official organ of the

Academy. It publishes articles on science

policy, the history of science, critical reviews,

original science research, proceedings of

scholarly meetings of its Affiliated Societies,

and other items of interest to its members. It

is published quarterly. The last issue of the

year contains a directory of the current

membership of the Academy.

Subscription Rates

Members, fellows, and life members in good

standing receive the Journal free of charge.

Subscriptions are available on a calendar year

basis, payable in advance. Payment must be

made in US currency at the following rates.

US and Canada $30.00

Other Countries $35.00

Single Copies (when available) $15.00

Claims for Missing Issues

Claims must be received within 65 days of

mailing. Claims will not be allowed if non-

delivery was the result of failure to notify the

Academy of a change of address.

Notification of Change of Address

Address changes should be sent promptly to

the Academy Office. Notification should

contain both old and new addresses and zip

codes.

Postmaster:

Send address changes to WAS, Rm GL117,

1200 New York Ave. NW

Washington, DC 20005

Academy Office

Washington Academy of Sciences

Room GL117

1200 New York Ave. NW

Washington, DC 20005

Phone: (202) 326-8975

MCZ !

LIBRARY

JUL 10 2079

HARVARD

UNIVERSITY

Volume 105

Number 1

Spring 2019

Journal of the

WASHINGTON

ACADEMY OF SCIENCES

Editor's Comments S. Howard

Rea CMCHT MRIC IYSUINI EG) CROSS 0.50 chee Wid seta aactsis aod es ve, Mirena cherie uae trate dances lil

TW EAENN Gee PRURCOTSOUINY, EDITING EV i Bo coc gatas sanded vena ae taseanns on ces psn aa coed 1

The Antigenic Shift or Drift of the Influenza Virus J. PQULONUS. ......c..cccceccesccstesscsteetesteeesseenes 7

In Vitro Antibacterial Activity of Garlic and Tea Tree Oil S. Godinez et al. .....c.cccce. is

Contextual Label Smoothing with a Phylogenetic Tree M. J. Trammell et al. ......0..... 23

(Ero lil, anya oo TCC o1 fC |) Seannne Set ete eens eae ese santTe Sree amu NT DONC Renner ener eT 47

RE TSSUN MLC RMCUENF CCE POCA CREOES: sess feces nsesk tO cans caseva hv ascii pein ines Oar babe A Dts ed alenes AT eogecm ea NR 48

POF AMNACELESCE NETSU BLUR UNCEINES Zoo cpsccte sore ea rea ants ets eseadb das ios neice scan g A roa N RCT Relea cea Al eee ere e 49

POPPE RTECS Se CIE CIS CATICE TO GNECIACES: yes eres ccssrs sccts sivas cies scares davende Solereeraess ate nape tee cotter ca

ISSN 0043-0439 Issued Quarterly at Washington DC

Spring 2018

EDITOR’S COMMENTS

Presenting the 2019 Spring issue of the Journal of the Washington Academy

of Sciences.

I encourage people to write letters to the editor. Please send email

(wasjournal@washacadsci.org) comments on papers, suggestions for

articles, and ideas for what you would like to see in the Journal. I also

encourage student papers and will help the student learn about writing a

scientific paper.

First up are two tactile astronomy demos. These are especially

useful for students who learn through tactile means. Just how many stars

are in the Milky Way? A mere number is difficult to comprehend. This

paper addresses that issue.

To follow is a short description of the flu virus and how it can adapt

and change. Flu pandemics have killed millions of people. This paper was

accepted some months ago when the flu session was in full swing.

Next up is a student paper from Frederick Community College. It

discusses the medical uses for garlic and tea tree oil.

Finally a multi-author paper on contextual label smoothing.

The Journal is the official organ of the Academy. Please consider

sending in technical papers, review studies, announcements, and book

reviews.

We are a peer reviewed journal and need volunteer reviewers. If you

would like to be on our reviewer list please send email to the above address

and include your specialty.

Sethanne Howard

Washington Academy of Sciences

Journal of the Washington Academy of Sciences

Editor Sethanne Howard showard@washacadsci.org

Board of Discipline Editors

The Journal of the Washington Academy of Sciences has a twelve member

Board of Discipline Editors representing many scientific and technical

fields. The members of the Board of Discipline Editors are affiliated with a

variety of scientific institutions in the Washington area and beyond —

government agencies such as the National Institute of Standards and

Technology (NIST); universities such as Georgetown; and professional

associations such as the Institute of Electrical and Electronics Engineers

(IEEE).

Anthropology Emanuela Appetiti eappetiti@hotmail.com

Astronomy Sethanne Howard sethanneh@msn.com

Behavioral and Social

Sciences Carlos Sluzki esluzki@gmu.edu

Biology Poorva Dharkar poorvadharkar@gmail.com

Botany Mark Holland maholland@salisbury.edu_

Chemistry Deana Jaber djaber@marymount.edu

Environmental Natural

Sciences Terrell Erickson terrellerickson|@wde.nsda.gov

Health Robin Stombler rstombler@auburnstrat.com

History of Medicine

Operations Research

Science Education

Systems Science

Alain Touwaide

Michael Katehakis

Jim Egenrieder

Elizabeth Corona

atouwaide@hotmail.com

mnk(@reirutgers.edu_

jim(@deepwater.org

elizabethcorona@gmail.com_

Spring 2018

Washington Academy of Sciences

Tactile Astronomy Demos:

Milky Way “Stars like Grains of Sugar” plus

Ball and Sun Lunar Phases

Gene Byrd

University of Alabama

Abstract

Indoor and outdoor astronomical size/distance demonstrations are well- known.

Here we discuss two tactile demos showing nos sizes but astronomical number

and shape. In the first even elementary students appreciate the immense number

of Milky Way stars using a 5 |b. bag of fine-grained sugar. Using the approximate

size of a grain, a typical bag would be about 1000x1000x1000 grains in length,

width and depth thus containing about a billion grains. When the bag is

theatrically poured slowly into a container, students can see and, afterward, feel

the "multitude" of sugar stars in just one bag. The roughly 100 billion stars in the

disk of our Milky Way are comparable to the number of grains in a hundred bags

of sugar, far too many to bring to class! Sand can be used if convenient. The

second demo dramatically shows the shape and origin of the phases of the

Moon.as illuminated by the Sun. Both must be visible on a clear sunny morning

or afternoon. Holding a small ball with thumb and forefinger in the Moon's

direction magically creates on a “microscopic” scale the same phase for the ball

(crescent, half or gibbous) as for the much larger and more distant Moon “beside”

the ball.

Introduction

INDOOR AND OUTDOOR ASTRONOMICAL size/distance demonstrations are

well-known, e.g., of the huge ratio of the Sun’s size versus planets, and the

separations of the Sun and planets versus their sizes. The excellent NASA

After School Universe program and site:

https://imagine.gsfe.nasa.gov/educators/programs/au/ contains exercises

along these lines, most notably a paper plate scale model of the Milky Way.

Here we discuss a visual and tactile demonstration showing not sizes but the

“astronomical” number of stars in our Milky Way. We also discuss a tactile

demonstration of lunar phases on a micro and macro astronomical scale.

Spring 2019

in)

The Number of Stars in the Milky Way

In the first demonstration we used grains of sugar to help students

appreciate the immense number of Milky Way stars. While this concept Is

probably not totally new, for this author, this demo was triggered by

Archimedes’ work: The Sand Reckoner. With only a few planets and only a

few hundred cataloged stars known at that time, Archimedes estimated the

number of grains to fill an enlarged universe as necessitated by Aristarchus’

heliocentric theory. Today, an immensely larger number of stars in just our

Milky Way Galaxy is inferred from modern estimates of the mass of the

Galactic disk and bulge.

For an elementary school class, we bought a 5 lb. bag of fine-grained

sugar. See Figure |. The size of a grain is about 0.1 mm so a 10x10x10 cm

bag would be about 1000x1000x1000 grains in length, width and depth.

Multiplying, together, our bag had about a billion grains. The teacher

theatrically poured the bag's grains slowly into a container letting the

students see and, afterward, feel the “multitude” of sugar stars from just one

bag. There are about 100 billion stars in the disk of our Milky Way. This

huge number is comparable to the grains in a hundred bags of sugar. This is

far too many to bring to class! Sand can be used if available in a

conveniently sized or shaped bag.

Figure 1: A billion grains visually and tactilely displayed

Washington Academy of Sciences

we

Moon Phases with a Ball

Again tactilely and visually, the second demonstration dramatically

shows the shape and origin of phases of the Moon. For this demonstration,

the Sun and Moon must both be visible in a clear sky. The morning sky

shortly after sunrise is usually best. The teacher or student has to be alert

for good observing conditions and the time a given phase is in the sky.

Holding a golf or tennis ball with thumb and forefinger in the Moon’s

direction magically creates on a “microscopic” scale the same phase for the

ball (crescent, half or gibbous) as for the much larger and more distant

Moon seen “beside” the ball. See Figure 2 for the arm, ball, Moon, and

observer orientation.

Figure 2: Holding the ball in a line almost between the eye and the sun.

This is a clear morning with both the sun and the waning gibbous moon in

the sky.

Spring 2019

Figure 3 shows a close-up of a golf ball on a push pin in the 3" quarter

position relative to the Sun. If you look carefully, you can see the 3™ quarter

moon directly above the golf ball! Note that the “day-night line” terminator

orientation matches that of the actual Moon. This is a simple photo taken

with a cell phone camera held at the observer’s eye. The camera lens must

be as close as possible to the eye/golf ball/moon line, not off to one side.

The golf ball provides ready-made “craters” which are best seen along the

terminator of the ball as on the moon itself through a small telescope or

binoculars.

Figure 3: Holding a golf a ball on a stickpin in sunlight to generate phases

of the Moon. 3" quarter is created for both the ball and the Moon is seen

above it. The same phase results because of the same Sun-Observer-

Moon/ball shape and orientation on a micro and macro size/distance scale.

Washington Academy of Sciences

Conclusions

We have explored two simple tactile astronomical demonstrations.

The first gives a striking visual and tactile “feel” for the billions of stars in

the disk of our Milky Way Galaxy. An abstract factor in the Drake Equation

for the number of currently existing life and civilizations in our Galaxy is

thus made real.

When illuminated by the Sun, we have seen how a hand-held golf

ball “magically” shows the same phase as the more distant Moon in the

same direction. This provides a strong tactile and visual feel beyond simply

looking at a diagram or just using an artificial light and ball alone.

Acknowledgements

| acknowledge Ms Lavender's Tuscaloosa Capitol School 4th grade and

University of Alabama online Astronomy lab course students for serving as

test subjects. See https://www.researchgate.net/profile/Gene_Byrd2 for

fm

photos efc. on this and other educational and research topics

References

Archimedes’ work 287-212 BCE. The Sand Reckoner, Paupitns,

http://www.numericana.com/answer/archimedes.htm annotated and

translated by Gerard Michon 2002-2015, is a work by Archimedes in

which he set out to determine an upper bound for the number of grains

of sand that fit into the Universe.

NASA After School Universe program and site.

https://imagine.gsfc.nasa.gov/educators/programs/au/ Click on

Afterschool Universe Program Leader’s Manual and go to Session 9 on

the Milky Way.

Drake Equation. The Drake equation 1s a probability argument established

by Dr. Frank Drake and used to estimate the number of active,

communicative extraterrestrial civilizations in the Milky Way galaxy.

The equation summarizes the main concepts which scientists must

contemplate when considering the question of other radio-

communicative life. See https://www.seti.org/drake-equation-index

Spring 2019

Bio

Gene G. Byrd is a Professor Emeritus of Astronomy at the University of

Alabama in Tuscaloosa, Alabama.

Washington Academy of Sciences

The Antigenic Shift or Drift of the Influenza Virus

John J. Paulonis

Abstract

Although there is no antigenic shift for this year, the process is quite

interesting. I trace the history of the flu and describe antigenic drift and

antigenic shift.

THE FLU WAS FIRST IDENTIFIED by Hippocrates around 410 BCE,

describing a highly contagious illness found in northern Greece. It wasn’t

until 1357 CE that the term ‘influenza’ was derived. The word originated

from the Italian ‘influenza di freddo’ (cold influence) named for an

epidemic in Florence, Italy where the people identified that this illness was

demonstrated during the colder weather.

The flu was first thought to be a bacterium, but it wasn’t until 1931

that a virus in pigs was discovered to be the cause of the flu (in humans, in

1933).

The most infamous pandemic (occurring over a large geographic

area, either in a country or the world) was the Spanish Flu of 1918. It has

been said that more U.S. soldiers had died from the flu during WWI than

from battle itself. (https://www.history.com/topics/inventions/flu )

Influenza viruses have distinct nomenclature depending upon the

genetic make-up of the virus. The various particular strains are given

nomenclature such as HIN1I, more commonly referred to as the “Swine

Flu”. (The H is an abbreviation for hemagglutinin while the N is an

abbreviation for neuraminidase. HA, meaning hemagglutinin antigen, and

NA meaning neuraminidase antigen).

We are currently experiencing the 2018 — 2019 Flu Season. In

general the influenza virus can undergo a number of changes and may

become virulent, even though a person has received an influenza vaccine.

This year’s influenza activity are listed in Figure 1.

Winter 2018

17.7 million — 20.4 million 214,000 - 256,000 13,600 - 22,300

flu illnesses flu hospitalizations flu deaths

?

“These estimates are preliminary and based on data from CDC’s weekly influenza surveillance reports summarizing key influenza activity indicators.

Figure | Influenza activity

(Retrieved Feb 2019 from https://www.cde.gov/flu/index.htm)

According to the CDC, the dominant Influenza A strain which has

been predominantly testing positive is (HIN1)pdm09, with one quarter of

specimens testing positive for H3N2. Vaccine effectiveness was estimated

to be 46% (30%—58%) against illness caused by influenza A(H1N1)pdm09

viruses. (Office of the Associate Director for Communication, Digital

Media Branch, Division of Public Affairs. (2019, Feb. 22))

Antigenic drift are small changes in the genes of influenza viruses

that happen continually over time as the virus replicates. As antigenic

changes accumulate, the antibodies created against the older viruses no

longer recognize the “newer” virus, and the person can get sick again. See

Figure 3.

“Antigenic shift is an abrupt, major change in the influenza A

viruses, resulting in new hemagglutinin (HA refers to glycoproteins on the

surface of influenza viruses which cause red blood cells to agglutinate. The

red blood cells clump. HA attaches to cell receptors and initiates the

process of virus entry into cells.)' and/or new hemagglutinin and

neuraminidase (NA).” The function of the NA protein is to remove sialic

acid from glycoproteins. It is the cell receptor to which the influenza virus

attaches via the HA protein. HA and NA are proteins in influenza viruses

' http://www. virology.ws/2013/11/05/the-neuraminidase-of-influenza-virus/

Washington Academy of Sciences

9

that infect humans. While influenza viruses are changing by antigenic drift

all the time, antigenic shift happens only occasionally.* See Figure 2.

Avian influenza refers to the disease caused by infection with avian

(bird) influenza (flu) Type A viruses. These viruses occur naturally among

wild aquatic birds worldwide and can infect domestic poultry and other

bird and animal species. Avian flu viruses do not normally infect humans.

However, sporadic human infections with avian flu viruses have occurred.

(Centers for Disease Control and Prevention, National Center for

Immunization and Respiratory Diseases (NCIRD) (2017, Apr. 13))

Such a “shift” occurred in the spring of 2009, when an HIN1 virus

with a new combination of genes emerged to infect people and quickly

spread, causing a pandemic. When shift happens, most people have little

or no protection against the new virus.

Bio

J. Paulonis has a Master’s of Science in Natural Sciences from the

Roswell Park Cancer Institute Graduate Division of the State University

of New York at Buffalo and a Master’s of International Management from

the Thunderbird School of Global Management.

fe ees

? Sep 27, 2017 ( https:/Awww .cde.gov/flu/about/viruses/change.htm)

Winter 2018

10

The genetic change that enables a flu strain to jump from

one animal species to another, including humans, is called “ANTIGENIC SHIFT.”

Antigenic shift can happen in three ways:

The new strain

may further

evolve to spread

from person to

person. If so, a

flu pandemic

could arise.

© without

undergoing

Bird influenza A strain | genetic change,

a bird strain of

influenza A can

jump directly

from a duck

or other aquatic

bird to

humans.

HA

antigen

4 T A-1 } A duck or other

aquatic bird passes a bird

strain of influenza A to

an intermediate host

such as a chicken or pig.

antigen

Co

Without

undergoing

genetic change,

a bird strain of

influenza A

can jump

directly from a

duck or other

aquatic bird to

an intermediate

animal host and

then to humans.

r A-2 } A person passes a

human strain of

influenza A to the

same chicken or pig. (Note that reassortment can

occur in a person who is infected with two flu strains.)

antigen

A-3 | When the viruses infect the same cell,

the genes from the bird strain mix

with genes from the human

strain to yield a new strain.

|

5

Viral entry

intermediate host cell

The new strain

can spread

from the

intermediate

host to

humans.

Intermediate

host cell

Genetie mixing

Link Studio for NIAID

Intermediate

host (pig)

Figure 2 antigenic shift

Washington Academy of Sciences

1) Each year’s flu vaccine contains three flu strains —

two A strains and one B strain — that can change from year to year.

C2) After vaccination, your body produces infection-fighting antibodies

against the three flu strains in the vaccine.

3) If you are exposed to any of the three flu strains during

the flu season, the antibodies will latch onto the virus’s

HA antigens, preventing the flu virus from attaching to

healthy cells and infecting them.

r 4) Influenza virus genes, made of RNA,

eee ~ are more prone to mutations than

genes made of DNA.

y Mutation

<a Antibody

\

\

HA

antigen

Link Studio for NIAID

5 if the HA gene changes, so can the

antigen that it encodes, causing

it to change shape.

HA gene

HA antigen

6 ) If the HA antigen changes shape, antibodies that 7

normally would match up to it no longer can, allowing oY ab

the newly mutated virus to infect the body’s cells.

This type of genetic mutation is called “ANTIGENIC DRIFT.”

https://www.verywellhealth.com/what-are-antigenic-drift-and-shift-

770400

Figure 3 antigenic drift

Winter 2018

References

History.com Editors. (2018, Aug. 21) Influenza. Retrieved from

https://www.history.com/topics/inventions/flu

Office of the Associate Director for Communication, Digital Media

Branch, Division of Public Affairs. (2019, Feb. 22) Weekly U.S.

Influenza Surveillance Report. Retrieved from

https://www.cdc.gov/flu/weekly/index.htm#whomap

Racaniello, V. (2013, Nov. 5) The neuraminidase of influenza virus.

Retrieved from http://www. virology.ws/2013/1 1/05/the-neuraminidase-

of-influenza-virus/

Office of the Associate Director for Communication, Digital Media

Branch, Division of Public Affairs. (2017, Sept. 27) How the Flu Virus

Can Change: “Drift” and “Shift”. Retrieved from

https://www.cdc.gov/flu/about/viruses/change.htm

Duda, K. (2018, Dec. 19) Antigenic Drift and Shift With the Flu Virus.

Retrieved from https://www.verywellhealth.com/what-are-antigenic-

drift-and-shift-770400

Centers for Disease Control and Prevention, National Center for

Immunization and Respiratory Diseases (NCIRD) (2017, Apr. 13)

Information on Avian Influenza. Retrieved from

https://www.cdc.gov/flu/avianflu/index.htm

Washington Academy of Sciences

In Vitro Antibacterial Activity of Garlic and Tea Tree

Oil

Silvia Godinez, Godfrey Ssenyonga, Judy Staveley

Frederick Community College

Abstract

To evaluate antibacterial activity of tea tree oil and fresh pure garlic against

infectious bacteria preparations of each were combined at different

concentrations with cultures of bacteria. The selected essential oil and fresh

crushed garlic were screened against one gram-negative bacteria (Escherichia

coli) and five gram-potentially positive bacteria (Bacillus cereus,

Staphylococcus epidermidis, Bacillus subtilis, and Micrococcus luteus).

Different concentrations (1:1, 1:25, 1:50) were tested using the disc diffusion

method. Tea tree essential oil and fresh crushed garlic showed antibacterial

activity against one or more bacterial strains. The different concentrations were

used to test for differences in antibacterial activity employing the disc diffusion

method. The 100% tea tree essential oil and fresh crushed garlic preparations

exhibited significant inhibitory effects against the tested bacterial strains. Tea

tree oil and the fresh crushed garlic showed promising inhibitory activity even at

low concentrations. In conclusion, tea tree oil and crushed fresh garlic showed

antibacterial activity against several tested bacterial strains. These findings

support the inference that preparations of 100% tea tree oil and of garlic could

play a role in inhibiting infection by some gram negative and gram positive

bacteria.

Background

THE SPREAD OF ANTIMICROBIAL RESISTANT PATHOGENS is one of the

most serious threats to efficacious treatment of microbial diseases. Essential

oils and other food plant extracts such as garlic have been used as alternative

medical treatments. Many such remedies have been investigated for

potentially possible use against a variety of communicable diseases (Zaika,

1988).

Medicinal plants like garlic are used extensively today in food

products and in culinary dishes. Fresh garlic has been used for many

centuries around the world, especially in the United States, Mexico, Africa,

and the Far East. It is scientifically proven that garlic is effectively used

against bacterial, viral, mycotic and parasitic infections (Gulsen & Erol,

2010). There is evidence that the garlic plant has immunological properties

Spring 2019

14

that include enhancing the immune system against malignancy and

disorders of immune functioning. In this research the potential

antibacterial properties of crushed garlic (A//ium sativum) and its use of

antimicrobial potency were investigated against six strains of bacteria. The

antibacterial activity was determined using the disc diffusion method.

Essential oils have been shown in many research articles to possess

antibacterial, antifungal, antiviral insecticidal and antioxidant properties

(Burt, 2004). Tea tree oil has been used for over 100 years as a healing

treatment in different countries, particularly for skin conditions. Tea tree oil

is best known for its antibacterial activity although it has other likely

medicinal properties. To evaluate specifically the antibacterial activity of

Tea Tree Oil (Melaleuca alternifolia) preparations of different

concentrations of the oil were tested against six strains of bacteria. Again

the level of antibacterial activity was determined using the disc diffusion

method.

Methods

Microorganisms

Microorganisms were obtained from the Department of

Biotechnology, Frederick Community College, Frederick, MD. Six strains

of bacteria were used (Table 1). The cultures of bacteria were maintained in

their appropriate agar slants at 4°C throughout the study and used as stock

cultures. The selected essential oil was screened against one gram-negative

bacteria (Escherichia coli) and five gram-positive bacteria (Bacillus cereus,

Staphylococcus epidermidis, Serratia marcences, Bacillus subtilis, and

Micrococcus luteus).

The three different concentrations of fresh pure garlic (A//ium

sativum) and Tea Tree Oil (Melaleuca alternifolia) (1:1, 1:25, and 1:50)

were prepared using the disc diffusion method.

Washington Academy of Sciences

Table 1

6 Strains of bacteria Type of bacteria

Bacillus subtilis ATCC 6633

Staphylococcus Gram positive ATCC 12228

epidermis

Escherichia coli ATCC 75922

Essential oils

100% concentration tea tree oil was obtained and was used in this

study (Table 2). This essential oil was selected based on previous literature

in which it has been used in alternative medical practices and

experimentation.

Fresh Crushed Garlic

Fresh chopped garlic was obtained from a local grocery store, and

used this study (Table 2). This fresh garlic was selected based on previous

literature used in alternative medical experiments.

Antibacterial Assay

Screening of the tea tree oil and crushed garlic was conducted to

estimate antibacterial activity. The antibacterial assay was conducted with

the disk diffusion method. This process is normally used as a preliminary

check. The antibacterial assay was performed by using a 45 h culture at

37°C incubation. Five hundred microliters of the suspensions were spread

over the plates containing BBL nutrient agar using a sterile inoculating loop

in order to get a uniform microbial growth on both control and test plates.

Spring 2019

16

The tea tree and fresh crushed garlic were dissolved in an aqueous solution

of water and dimethylsulfoxide (DMSO).

Table 2

Essential Oils Botanical Name Properties

Tea Tree Oil Species: M. alternifolia Antiseptic,

Kingdom: Plantae antibacterial,

Clade: Angiosperms, antiviral, antifungal,

Eudicots and anti-

Family: Myrtaceae inflammatory agent.

Genus: Melaleuca

Fresh Garlic Species: A. sativum Antiseptic,

Kingdom: Plantae antibacterial,

Clade: Angiosperms, antiviral, antifungal,

Monocots and anti-

Family: Amaryllidaceae inflammatory agent

Subfamily: Allioideae

Genus: Allium

Under aseptic conditions empty sterilized discs (5S, 6 mm diameter)

were infused with different concentrations (1:1, 1:25, and 1:50) of the

respective tea tree oil and fresh crushed garlic. They were placed on the

BBL nutrient agar surface. The paper disc was saturated with aqueous

concentrations of the tea tree oil and fresh crushed garlic. DMSO was mixed

in a microcentrifuge with different concentrations of tea tree oil and fresh

garlic. The standard disc was saturated with mixed concentrations and

placed on the petri dish. A standard disc containing DMSO was used as

reference control for every species of bacterium. All petri dishes were

sealed with sterile laboratory tape to avoid evaporation of the test samples.

The plates were left for 30 min at room temperature to allow for the

diffusion of oil, and then they were incubated at 37°C for 45 h. After the

incubation period, the zone of inhibition was measured in centimeters with

a caliper and data were recorded. Studies and data were collected over a

Washington Academy of Sciences

series of months.

Results

Antimicrobial activity of Tea Tree oil and garlic oils

We tested the effects of tea oil and garlic against six types of bacteria in

three different concentrations. The Tree Tea Oil showed a greater inhibitory effect

on cereus and E. coli and the smallest effect was observed on the S. epidermis

(Graph 1A). While fresh garlic extract was more effectively inhibits M. /uteus and

E.coli (Graph 1B), both effects are clearly observed at the high and medium

concentrations of essential oils. The values were compared against negative

control.

A concentration vs response

2.0 =

E -@ Bacillus subtilis

£& 45 +4 Bacillus cereus

=F

= -*- Staphylococcus epidermis

2 4.0 -@- Escherichia coli

= -e Micrococcus luteus

io) :

w 0.5 -e Serratia marcencens

6

N

0.0

1.00 1.25 1.50

Concentration

B concentration vs response

2.0 -O- Bacillus subtilis

= +4 Bacillus cereus

= Ae5

ie) -«- Staphylococcus epidermis

fe

2 1.0 ->- Escherichia coli

£€ .

= -e Micrococcus luteus

= 0.5 -e Serratia marcencens

=

(o)

N

0.0

1.00 1.25 1.50

Concentration

Graph 1.Antimicrobial activity graphics. A) Concentration vs response

graph of the inhibition from Tree tea oil. B) Concentration vs response

graph of the inhibition from Fresh garlic.

Spring 2019

18

Tea Tree Oil and Fresh Garlic Extract showed a synergic effect

We tested the inhibition of both compounds (Tree tea oil and Fresh

garlic) at the 1.25 concentration with two bacteria (Bacillus cereus and

Escherichia coli). The bacteria showed inhibition in the presence of both

extracts. The results indicated that the effect of inhibition of these two

extracts together was more effective than the activity of each one. Figure |

shows the plates where the inhibition when the tea and garlic were mixed

and Graph 2 shows the corresponding data.

Tree tea oil and Fresh

garlic

Tree tea oil “= Fresh garlic

— —s

So a

Zone of inhibition (cm)

So

a

0.0

Graph 2 and Figure 1. Synergic effect. Right panel. Graph of inhibition

with Tree tea oil, Fresh garlic and mixture. Left panel. Upper plate,

inhibition of B. cereus by tree tea oil, fresh garlic and mixture. Lower

plate, inhibition of £. coli by tree tea oil, fresh garlic and mixture.

Dilution 1:25 was shown to have antibacterial activity against E.coli

and B. cereus. The mix of Tea Tree Oil and Fresh Garlic Extract showed a

synergistic effect.

Preliminary results — antibacterial Allicin Identification

Several reports have described that the main component of the

antibacterial activity of garlic is allicin. Obtaining this biologically active

component compound is difficult. We assessed the presence of allicin from

a commercial product by analyzing it with Infrared Fourier Transform

Spectroscopy. Figure 2 shows the observations and comparisons between

Washington Academy of Sciences

19

the allicin and fresh garlic extract. The spectra are demonstrated by the main

peaks.

0]

: 4 ‘i f { _

s / \ ° "ae .

oi / 1634.1 cm-W aa

8 / 2978 crn-1 | \ I

7 1424 cm-1 | Ne

920 cm

1087.0 crn-1

3000 2500 2000 1500 1000 500

Wavenumber cm-1

ewe od

\

\

a) / | /\

2 iy A j / |

/ \(2848.93 cm-1 ri La

=e \ a s em-1ern-1 t] Y{!

5 ' 1638.15 zal 32718 om-1 a

5 \ | | \

c © |} | i

s 1418.98 cm-1| fi

13666cm-1 | |

3B |

3285.35 crn-1

3 \f \

| { My,

ze 10392.69cm-1 |

~~ — = 1 ot poe

3500 3000 2500 2000 1500 1000 500

Wavenumber cm-1

Figure 2. Spectra of the Infrared Fourier Transform. Top panel, Spectrum of

allicin from capsules. Bottom panel, spectrum of the bulb extract of the garlic.

Peaks identified, 988 cm’! prob. Flex (6) R-CH=CH2; 1087 cm! S=O; 1424 cm’!

5 CH; 1634. 1 cm! C=C; -1 v sim CH: and v asim CH2,

Spring 2019

20

Preliminary Results Antibacterial activity- Allicin)

The results showed the characteristic peaks of allicin were presented

in the fresh garlic extract. We evaluated the activity of the allicin from

capsules and used seem conditions by identifying if it had a synergistic

effect when mixed the Allicin and Tree tea oil. Figure 3 showed that allicin

inhibited the growth of bacteria; however, the inhibition of the mixture did

not show a synergistic effect.

Tree tea oil “= Allicin @m Tree tea oil and Allicin

°o

for)

=

for)

Zone of inhibition (cm)

—) =)

ip ~~

=

ro)

Graph 3 and Figure 3. Effects of allicine. Right panel. Graph of inhibition

of Tree tea oil, allicin and mixture. Left panel. Upper plate, inhibition of

B. cereus by tree tea oil, allicine and mixture. Lower plate, inhibition of E.

coli by tree tea oil, allicin and mixture.

Conclusion

The 100% Tea Tree essential oil preparation (Melaleuca

alternifolia), and fresh crushed garlic (Adlium sativum) exhibited significant

inhibitory effects against the tested bacterial strains. Tea Tree oil

(Melaleuca alternifolia), and the crushed fresh garlic (A//ium sativum)

showed promising inhibitory activity even at low concentrations. In general,

E. coli, M. luteus and B. cereus were the most susceptible. Therefore, the

Tea Tree oil and crushed fresh garlic both showed significant antibacterial

activity against the tested strains. The combination of tea tree oil and

crushed fresh garlic exhibited a degree of antibacterial activity that was

more than additive. Both tea tree oil and fresh crushed garlic separately and

in combination may have potential for use in suppressing the growth of

Washington Academy of Sciences

21

ys

pathogenic bacteria and could be used to develop a dose dependent practical

application as antibacterial agents. Further research is warranted.

References

Burt, S.A. (2004). Essential oils: their antibacterial properties and

potential applications in foods: a review. /nternational Journal of

Food Microbiology. 94: 223-253.

10.1016/j.1jfoodmicro.2004.03.022.

Gulsen, G. and Erol, A. (2010). Recent Patents on Anti-Infective Drug

Discovery. 5: 91. https://doi.org/10.2174/157489 110790112536

Prabuseenivasan, S., Jayakumar, M., & Ignacimuthu, S. (2006). /n

vitro antibacterial activity of some plant essential oils. BMC

Complementary and Alternative Medicine, 6, 39.

http://do1.org/10.1186/1472-6882-6-39

Staveley, J. and Ramos, M. (2018). Antimicrobial Properties of Four

Essential Oils. The Incubator Journal Frederick Community

College. Voll, Issue 1.

Zaika, L. (1988). Spices and herbs: their antibacterial activity and its

determination. J Food Safety, 23:97-118.

Spring 2019

Bio

Dr. Silvia Godinez received her doctoral degree four years ago. She spent

one year as a post-doc in Mexico, one year in Dr. Staveley’s laboratory at

Frederick Community College, and currently she is working in a second

post-doctoral position at the CRAG in Barcelona, Spain.

Washington Academy of Sciences

23

~ Contextual Label Smoothing with a Phylogenetic Tree

on the iNaturalist 2018 Challenge Dataset

Michael J. Trammell', Priyanka Oberoi', James Egenrieder’,

John Kaufhold!

Val eae ho ea = 5 . . > . aie

General Dynamics Mission Systems’ Deep Learning Analytics Center ?Virginia Tech

Abstract

Recognition of fine-grained visual categories (FGVC) in the natural world is

a long-tailed problem, meaning recognizers must accurately recognize a large

diversity of categories and most of those categories will naturally have limited

training data, increasing the likelihood of overfitting in these many limited

training data categories. The iNaturalist 2018 Challenge aimed to benchmark

the state-of-the-art performance on species identification from a photo, where

the long-tailed aspect of training is compounded by the visual similarity of

many species. We demonstrate a new state of the art on the iNaturalist 2018

Challenge with Contextual Label Smoothing (CLS). CLS extends label

smoothing to narrow the list of categories smoothed to only those within the

same branch of a phylogenetic tree. CLS regularization improves performance

significantly—the best publicly reported Top3 error reported on the 1Naturalist

2018 Challenge was approximately 13%, which we improve to 12% with an

ensemble of CLS networks trained with dynamic minibatching and additional

inference windows. We present evidence that a 1% improvement on the FGVC

iNaturalist 2018 Challenge test score (public score) represents over a 5 sigma

improvement (test score stdev = 0.17 %) over the former state of the art.

1. Introduction

THE PROBLEM OF FINE-GRAINED VISUAL CATEGORIZATION (FGVC) has

been studied across many domains with many image datasets, including

FGVC-Aircraft [1], Stanford Cars [2], motorcycles [3] and shoes [4],

among others. Many FGVC datasets of the natural world collect plant and

animal species [5], birds [6], vegetables and fruits [7], plants [8], and dog

breeds [9] to identify, among others. One of the largest and most imbalanced

public datasets of natural imagery with these long-tailed FGVC challenges

is the iNaturalist 2017 Challenge dataset, which the iNaturalist 2018

Challenge dataset made even larger and more imbalanced [10]. The

iNaturalist 2018 Challenge training and validation data was made available

by iNaturalist [11] and the competition was hosted on kaggle [12], which

scored submissions on an unseen test set. Organizers of the iNaturalist 2018

Challenge aimed to:

Spring 2019

push the state of the art in automatic image classification for real

world data that features a large number of fine-grained

categories with high class imbalance. ... The dataset features

many visually similar species, captured in a wide variety of

situations, from all over the world. [12]

1.1. iNaturalist 2018’s Long Tails

We call the most represented training categories in the iNaturalist

2018 Challenge data the “head” and the least represented categories the

“tail” of the distribution (as in [13]). Recent work [13] has highlighted key

properties of FGVC of long-tailed distributions: (1) there are many

categories (2) most of the categories have limited training data (the tail

categories) (3) error rates improve only when more labeled data is made

available for the tail categories and (4) additional training data for the head

categories does not appreciably improve overall performance (i.e. the

network does not transfer learn from the head categories to the tail

categories). On the iNaturalist 2018 Challenge data, approximately 10% of

the categories (~800) comprise the head of the distribution, where each

category has between 100 and 1000 training examples, and 75% of the

categories (~6000) comprise the tail categories, where each category has

between 2 and 30 training examples.

The prohibitive cost curve associated with generating sufficient

training data for long-tailed FGVC applications to reach a threshold

accuracy is sketched in [13]:

Collecting the eBird dataset took a few thousand motivated birders

about | year. Increasing its size to the point that its top 2000 species

contained at least 10* images would take 100 years.

Washington Academy of Sciences

Nw

Nn

n, : number of categories

Ny : number of categories at level ¢

in x,'s branch of phylogenetic tree i Le = »

phylog Vix = Hy (x)

U : constant over all n, categories — :

:

u(4) : 1 for all categories at level ¢ a k Network I

in x's branch of phylogenetic tree -

0 otherwise

K networks

Ps in learned

ensemble

YiLabsmooth =

(1-G)Y\ 1 por + AU/N,

Constant ove:

oll categores SISSIES SS SS Se

Contextual Label Smoothed

1-@:------- oo

. || Yicus c(u(g))+u(f)))

species ee ee o (ay, in (Np +i)

genus

Y learned

Figure 1: Contextual Label Smoothing (CLS) label form compared to related label smoothing

forms: |1-hot encodings are sparse labels (top left). For example, for xi, only one nonzero value in

Yi.l-hot 1s the target category and all others are Os. |-hot labels incorporate no regularization (either via

a prior or learned post hoc from ensembling). Label smoothing (middle left), contextual label

smoothing (bottom left), and distillation (right) all incorporate into their full label vectors some

degree of regularization. In label smoothing, the regularizer is very weak but effective—yi.Labsmooth

spreads out a small constant residual contribution of 0/ne to every category (where nc is the number

of categories and u is a constant over all categories). In distillation, K classifiers are first trained with

the 1-hot labels—the temperature-relaxed logits from the output layers of these K classifiers are then

combined into a learned regularization term that is scaled and added to the 1-hot target category to

form yi. The distilled version’s regularized yi. has dense structure reflecting similarities among

categories learned from the ensemble. Our method, contextual label smoothing (CLS), requires no

learning as distillation does, and encodes label similarity from a phylogenetic tree into yicis. The

number of categories shared at the genus and family level are ng and np, respectively. The notation

u(é\) takes the value | for all categories shared at the ¢ level with the target category for xi.

1.2. Label-efficient Approaches to Long Tails

For this reason, we seek more label-efficient approaches that

incorporate context to address long-tailed FGVC challenges. Our aim is to

efficiently encode in the labels, themselves, information that mitigates the

performance degradation to tail categories stemming from limited training

data. In the spirit of [14], in our proposed Contextual Label Smoothing

(CLS), we allow tail categories to learn from training data pooled from

similar categories as defined on a hierarchy (a phylogenetic tree) with label

vector encodings (i.e. soft targets). This judicious form of label smoothing

encodes information about which other categories are (likely to be) most

similar, but unlike [14], we do not /earn these relationships (which incurs a

computational cost), but encode them directly with a portion of the

phylogenetic tree [15] as the prior. The labels in the CLS approach are

Spring 2019

26

diagrammed in contradistinction to 1-hot encoding, label smoothing and

distillation.

While 1-hot label encodings (where one category is assigned a | and

all others Os) of categories have become common in mainstream object

recognition [16]-[18], we argue these 1-hot independent category labels are

label-inefficient—they do not effectively share informative training

examples across similar labels; they are also overconfident—they make

deep networks more susceptible to overfitting, especially on categories with

limited training data.

Two simple relaxations of the 1-hot label encoding to better calibrate

confidences in FGVC have been shown to improve (A) the robustness of

the learned networks [19] and (B) the ability to learn more accurate tail

categories post hoc from ensembles with limited training data [20]. In both

label smoothing and distillation, the training labels are not 1-hot, but full,

and retain some nonzero dot product from label vector to label vector.

Inspired by both label smoothing and distillation, we demonstrate that

contextual label smoothing (CLS), like hierarchical semantic encoding

(HSE), can improve recognition rates on long-tailed FGVC problems.

1.3. CLS is Hierarchical Label Smoothing

Uniform label smoothing is an a priori decision to spread

contributions from a target label over all other labels uniformly, which has

the effect of penalizing overconfident predictions [19]. Intuitively, label

smoothing allows a// other categories to contribute training data to a target

category, and spreading over a// categories may spread the label

information too thinly to efficiently transfer learn (as observed in [13]). In

this work, we extend label smoothing to spread contributions from a label

only within a branch of a phylogenetic tree provided a priori, not smooth

over all other categories. Briefly, CLS exploits the phylogenetic tree to be

more judicious about the label smoothing prior. Practically, we do not label

smooth a training example of a humpback whale to have a nonzero

contribution to learning a monarch butterfly category, but we do label

smooth a training example of a gluphisia moth to have a nonzero

contribution to learning the monarch butterfly category. While branches of

phylogenetic trees are not always indicative of visual similarity, we

empirically demonstrate that enough are to justify use of this prior.

1.4. CLS is Distillation with a Prior

Where distillation is an empirical post hoc approach to encode

similarity into label vectors [20], our CLS work can be viewed as a form of

Washington Academy of Sciences

27

a priori distillation (Figure 2). Specifically, in distillation, an ensemble of

classifiers are trained (from 1-hot labels). After learning, the (temperature-

relaxed) logits of this ensemble empirically develop higher values for both

the true category and visually similar categories. These post-hoc logits from

this ensemble are added to the true 1-hot (hard targets) label for every

training example in a downstream distillation of the ensemble. Intuitively,

if only a handful of other classes are visually similar to the true class, when

downstream training occurs with these distilled label vectors (soft targets),

every one of those visually similar categories will contribute non negligibly

to the training set for the original |-hot target label. In this way, distillation

reuses training examples from other categories to train to recognize the

target categories most visually similar to 1t—this makes distillation a more

label-efficient strategy than 1-hot encoding (Figure 2). CLS is an a priori

version of distillation, encoding similarity as shared parentage on a

phylogenetic tree provided without any downstream ensemble training (as

are /earned in either distillation or HSE).

1.5. Fine-Tuning with more Balanced Categories

On similar FGVC tasks [21], better performance was obtained by

further fine-tuning on a more balanced subset of FGVC validation data with

a small learning rate. Improvements on head categories with >100 training

images were relatively small compared to tail categories with <100 training

images. This provides an empirical rationale for fine-tuning on validation

data more uniformly distributed over categories to improve performance on

underrepresented tail categories. We incorporate this type of fine-tuning

into CLS.

1.6. Contributions

We make a number of original contributions in this work:

e Contribution 1: New State-of-the-Art on the iNaturalist 2018

Challenge. We demonstrate a new state of the art result on the long-

tailed FGVC iNaturalist 2018 Challenge Data [11]. We estimate

through a prediction set that this new state-of-the-art outperforms the

prior state-of-the-art by greater than 5 o on the unseen test data. We

estimate the confidence interval of the score estimator for the unseen

test data empirically via a Monte Carlo method. Specifically, we

estimate the best fit line to the score computed by kaggle on the

unseen test labels as a function of the score on the test score prediction

set labels we do see to estimate the standard deviation of the estimator

(see Figure 7 and Section 5 Test Score Prediction Analysis for

Spring 2019

details).

Contribution 2: CLS works best with uniform sampling over

categories. In contradistinction to natural sampling advocated in [13],

CLS benefits from uniform sampling of categories in training.

Contribution 3: CLS improves ensemble performance more per

marginal network than other methods. Given a choice between

adding a network trained with some other technique to increase model

diversity in an ensemble, adding another CLS-trained network is a

better choice. This clarity can reduce the significant hyperparameter

search and tuning costs over an ensemble.

Contribution 4: Larger Input Images Improve Performance.

While this is not a novel claim, we confirm empirically that larger

input size images, which have recently been shown to improve

performance on the same task without CLS [21], also improves

performance of CLS.

Washington Academy of Sciences

1-hot

I Label smoothing CLS | mY

] | Loans +3)

(x, yi) ! (xX), y;+au) (x, ytoy,(x))) 1 > =

1 | wes

1 u v(x) ‘ y eA

‘ a r 2s

! | £2

Xi Vi | xX Vi xX Vi 1 cA &

l 1 -

No 1 Label, y, regularization Label, y, regularization 1 2 8

regularization | with unlearned flat with unlearned ; & S

I _ _prior hierarchical prior | —

eeeeneaseeee

(x;+F(x;), yi)

ml

Regularizes gradient Label, y, regularization

flow by learning on learned from data, x, ©

on one semantic level

SUC OS ESS Se SEE ESE Eee Eee

Learned regularizers J]

Figure 2: Residual connection blocks regularize data and labels: Five deep learning-based

conceptual regularizer “blocks” to remedy overfitting and vanishing/noisy gradient issues of |-hot

label encodings (top left) are diagrammed. Across the top row are methods that only incorporate

unlearned regularizers (i.e. priors only). Across the bottom row are methods which incorporate

learned regularizers. On the bottom right, HSE incorporates both learned and unlearned regularizers.

The well-known ResNet architecture ({22] bottom left) adds copies of the data, x, to regularize

gradients—this architectural change is common to many of the other methods (both the trunk and

branch networks of HSE [14] implement ResNet models, e.g.). Label smoothing ([19], top middle)

can be viewed as a residual connection between a | -hot yi, and an unlearned uniform prior. This same

strategy inspires this work on CLS (top right), but we use an unlearned hierarchical prior in the form

of a phylogenetic tree. Distillation (bottom middle) can be viewed as a residual connection between

a |-hot yi and a /earned soft target (the posterior distribution from learning an ensemble was used in

[20]). The most general form of these combinations we have found is the very recent work on HSE

(bottom right), which incorporates residual connections /earned within trunk and branch networks,

learns to update soft target priors based on an unlearned hierarchical prior, and combines these with

residual connections at each level of the hierarchy.

2. Related Work

2.1. Deep Learning from 1-hot Labels

Since 2012 [17], deep networks have dominated the state of the art

in object recognition on images, maturing year over year to include new

network architectures [18], [22] until the performance of deep networks was

on par with or better than human performance on a standard benchmark

[23]. While significant attention has been paid to data augmentation [17],

transfer learning [24], and new architectures [18], [22], less work has been

Spring 2019

30

devoted to improving the |-hot labels [19], [20], themselves, for training

data. This work addresses improvements to the design of labels, themselves.

2.2. Label Vector Benefits

Work on improved label vector engineering includes label

smoothing [19] and distillation [20], among others (Figure 2). Label

smoothing is a simple method that incorporates a prior to drive deep

networks to solutions with higher posterior entropy. Distillation, while

originally proposed as a method to make networks smaller (in memory and

computational cost of inference), has also demonstrated regularization and

adversarial example defense properties.

Work on Hierarchical Semantic Embedding is most similar in spirit

to this work, but achieves its goals of incorporating category similarity

through a trunk and branches architecture over a collection of 1-hot label

vectors at various semantic levels (from coarse to fine) [14]. Similar to

distillation, it adds a predicted category score vector (i.e. a soft target) from

a coarser level to the 1-hot label vector at the next finer level. FGVC results

on three natural datasets, CUB [6], butterflies [14], and VegFru [7],

demonstrate the value of HSE. HSE outperforms 17 other state of the art

methods on CUB. The strategies employed in HSE appear to be more

general than the simpler unlearned CLS prior proposed here (Figure 2), but

HSE benefits have not yet been demonstrated on as large a dataset as that

of the iNaturalist 2018 Challenge, which has >25x more fine-grained

categories and >100x larger category imbalance, which are critically

relevant aspects of long-tailed FGVC challenges [13].

Importantly, none of the datasets used to demonstrate HSE has more

than 292 fine grained categories (compared to 8,142 for the iNaturalist

Challenge 2018 data), with CUB’s 200 categories separated into 122

genera, 37 families, and 13 orders, where 75% of CUB categories fall into

the head category with 60 training images/category, and where all

categories have at least 41 training images, for a max class imbalance of 1.5

(compared to 500 for the iNaturalist 2018 Challenge). The authors’ new

butterfly dataset also only contains 200 categories. This smaller scale of the

FGVC challenges addressed by nascent exploration of HSE is encouraging,

but qualitatively smaller scope than evaluation on iNaturalist Challenge

2018 data, which is an open dataset and more comprehensive than those

datasets HSE authors chose to evaluate on.

Interestingly, HSE training develops learned attention mechanisms,

Washington Academy of Sciences

3]

making a convincing case that without specifically labeled parts, HSE can

learn features that exploit part-based attention to discriminate in FGVC, as

was demonstrated to be critical for natural FGVC in other work [25]. The

critical difference between the label vectors in HSE and our CLS work is

that all of our label hierarchy information is encoded in label vectors

without branches. CLS is a de facto flat prior that is not learned and is

modularly separable from the architecture—i.e. there is only one label

vector for each example in CLS, whereas HSE requires different label

vectors at different levels in the architecture, increasing hyperparameter

search costs.

2.3. Long-tailed FGVC Implications

The properties and implications of long-tailed distributions in

FGVC have been summarized with convincing evidence [13] that (1)

statistics of natural image categories are long-tailed, (2) more training data

for head categories does not improve performance on tail categories, and

(3) natural sampling of categories in training minibatches outperforms

uniform sampling over categories. In [13], authors used standard 1|-hot label

encodings and sampled “naturally” (as opposed to uniformly) during

training. The argument for natural over uniform sampling was empirical—

results demonstrated both head and tail category performances both

improved more with natural sampling. In contrast, we argue that the

thoughtful vector encoding of labels with CLS overturns that guidance on

sampling method (Contribution 2). Choosing training minibatches from

CLS with uniform sampling over categories outperforms natural sampling.

Authors conclude: “As a community we need to face up to the long-tailed

challenge and start developing algorithms for image collections that mirror

real-world statistics” which outlines the core motivation for this work [13].

2.4. Prior State of the Art iNaturalist Performance

The iNaturalist 2017 Challenge was won by Google (GMI, for

Google Mountain View, on the leaderboard) with a TopS error rate of less

than 5% with an ensemble of InceptionV3 and InceptionV4 models trained

at both 299x299 and 560x560 input image sizes, and subsequently fine-

tuned on a balanced subset of the data left out of the test set [21]. The fine-

tuning on balanced data boosts performance on tail categories of the dataset

[1] and during inference 12 crops outperformed inference on a single

prediction for the entire image.

Compared to the iNaturalist 2017 Challenge, the iNaturalist 2018

Challenge reduced the number of training images provided from 675,170 to

Spring 2019

ioe)

i)

461,939, increased the number of classes from 5,089 to 8,142, and perhaps

most significantly, provided a complete taxonomy for each class. A team

from Dalian University won the 2018 challenge with a Top3 error rate of

13% [12]. Their winning ensemble consisted of 12 ResNet-152 models

trained at both 320x320 and 392x392 input image sizes, six of which used

matrix power normalized covariance pooling of the last layer of

convolutional features [2].

3. Training Methodology

3.1. Training and Validation Data Set Splits

The iNaturalist 2018 Challenge data includes three mutually

exclusive data sets: training, validation, and test data, each containing

photos drawn from one of 8,142 species categories distributed over 4412

genera. The training data distribution is imbalanced, with the most

represented species, Branta canadensis the “Canada goose”, having 1000

training examples, whereas the least represented species in the training data

is the Spatula clypeata, the “Northern shoveler duck,” with only two

training examples. The validation set is uniformly distributed over species,

with three validation images per species. The test set labels are not provided

to entrants, but entrants can submit Top3 label lists for each of the 149k test

images to be scored on a Top3 error rate that is blind to which examples

were marked correctly or incorrectly. In the development that follows, 2/3

of the validation data (two photos per species) is used for validation fine-

tuning and 1/3 of the validation data (one photo per species) is used as the

test score prediction set. In “vanilla” label smoothing [19], we assign the

target label 0.8 and distribute the remaining 0.2 of that example to all other

8,141 categories in the label vector.

3.2. Initialization with Pretrained Networks

Closely following the winning GMV entrant from the iNaturalist

2017 Challenge, we start from an IRV2 and [V4 pretrained on ImageNet

[18], [22]. These two network architectures are the starting points for

training across all input sizes (299x299 and 598x598) and label smoothing

methods (1-hot, vanilla label smoothing, and CLS). As in GMV, for each

network in an ensemble, we strip the final layer of ImageNet-1K classes

from the pretrained network and replace it with the iNaturalist 2017 output

layer of 5,089 categories and sample minibatches of 32 images per

minibatch without replacement from all training examples (we trained on 4

GPUs in parallel for an effective minibatch size of 128 for the IRV2 model

and 6 GPUs in parallel for an effective minibatch size of 192 for the [V4

Washington Academy of Sciences

LoS)

ww

model). We fine-tuned on the iNaturalist 2017 training data for {80, 84}

epochs for {IRV2, 1V4!. We then fine-tuned on 90% of the iNaturalist 2017

validation data for {30, 14} epochs for {IRV2, V4} using {8, 4! GPUs for

effective minibatch sizes of {256, 128!. We used SGD with an initial

learning rate of 0.018 and momentum=0.9 in the first round of training for

the IRV2 model, reducing the learning rate by 10% every {8,6} epochs for

(IRV2, 1V4}. We used RMSProp for all other training. The second round

of training began with learning rates of 0.002 for the IRV2 model and 0.001

for the 1V4 model, and the training rate was multiplied by 0.9 every 10

epochs. Note that all minibatches in this pretraining were sampled naturally

(as opposed to uniformly with replacement).

3.3. Base Fine-Tuning on iNaturalist 2018 Challenge Data

We strip the final layer of iNaturalist Challenge 2017 categories

from each pretrained network and replace it with the iNaturalist 2018

Challenge output layer with 8,142 categories. When training, we sample

minibatches uniformly over categories with replacement (i.e. we sample

uniformly); this produces minibatches with approximately equal

contributions from all 8,142 categories. We train for 1M-1.4M iterations

using RMSprop with a base learning rate of 0.0045 in base fine-tuning. We

use a batch size of 32. We retain only the model with the highest

performance on the validation set, as assessed every SOk iterations.

3.4. Validation Fine-Tuning on iNaturalist 2018 Challenge Data

We fine-tune on the validation fine-tuning set only. The validation

fine-tuning regime is identical to the base fine-tuning regime with the

exception that training begins with a base learning rate of 0.0002, and

continues for only 25k iterations.

Spring 2019

Portrait oriented photo

Figure 3: Additional inference windows on a photo. The “standard” twelve inference

windows (six with the original image, the same six with the image flipped horizontally) are

shown on the left of each orientation. For portrait-oriented photos, a second set of inferences

is made on twelve more windows biased toward the top of the photo; for landscape-oriented

photos, the second set of inferences is made on twelve more windows biased toward the

left of the photo.

3.5. Ensembling

We compute unweighted model average ensemble results from

multiple label smoothing methods to conduct a post hoc ablation study via

ensemble composition. We rank the performance boosts from different

components of the ensembles to assess the benefits of individual

components of each ensemble. Ensemble components vary in input image

size, network type, and label smoothing type.

3.6. Test Performance Error Analysis

Additional inference windows: When scoring, we include the

standard middle, whole image, and four corner inference windows (with LR

reflections). As an approximation to attention, we also include additional

inference windows favoring the sides and top of the image calculated based

on the aspect ratio of each image, under the assumption that this is where

photographers are more likely to include the subject of the photo.

Test score prediction error rates: Nominally small (<0.5%)

differences in Top3 error rates on leaderboards can be difficult to assess the

relative merits of. By estimating a test score from the test score prediction

data on many model outputs, we estimate a practical error bar on our test

performances.

4. Results

The results collected here represent approximately 20,000 total GPU

hours across a mix of NVIDIA GTX® 1080s, V100s and Titan® Xs.

Washington Academy of Sciences

35

For practical perspective, training a single one of our models through

to final scoring on 2 GPUs requires approximately 10 days of compute on

299x299 input image sizes and 20 days on 598x598 input image sizes. Note

that due to the size of our images and batches, only V100s can be used to

train some of our models at our largest image sizes.

c

~ o

Qa

tS) 8 x

ae x

w

= “

ro) 3

¥ 9.83 ™

re

o *

=

0.82

° + CLS IRV2 and Iv4

+ CLS IRV2 or IV4

0.81 -@- 1-hot IRV2 and Iv4

o 1-hot IRV2 or Iv4

>< LabSmooth IRV2 and Iv4

x LabSmooth IRV2 or Iv4

3 4 6

Number of Models in Ensemble

Figure 4: CLS vs. label smoothing vs. 1-hot encodings. CLS networks and ensembles of CLS

networks outperform label smoothing and no label smoothing for both IRV2 and 1V4 architectures

assessed. The iNaturalist 2018 Challenge test scores returned from kaggle for the unseen test set is

plotted vs. the number of models ensembled for each label smoothing method. A second-degree

spline fit is plotted through the mean score of each set of IRV2 and 1V4 ensembles for visual clarity.

0.8797

0.88 4 % —— = x

0.87

0.86 ; 0.8583

La) +

2 ; +

ic v » ay

vu 4 i

4

°

0 0.85

"

w

va)

vo

| |

|

0.84 +

+

:

0.83: —} IRV2 and IV4 @598

Vv IRV2 or |\V4 @598

+ IRV2 and IV4 @299

+ IRV2 or IV4 @299

3 4 6

Number of Models in Ensemble

Figure 5: CLS input size comparison. We find that CLS on larger input image sizes (598x598)

consistently outperforms CLS on smaller input image sizes (299x299). A second-degree spline fit is

plotted through the mean score of each set of IRV2 and IV4 ensembles for visual clarity.

Spring 2019

4.1. Label Smoothing Method Comparison

We show final iNaturalist 2018 Challenge test score results from

kaggle on 299x299 pixel resolution images for the three label smoothing

methods: |-hot (i.e. no label smoothing), vanilla label smoothing (with 0.2

redistributed across all non-target classes), and CLS (with 0.2 redistributed

across non-target classes in the same branch of the phylogenetic tree).

Results of 3 runs each of {IRV2,IV4} and their ensembles demonstrate CLS

outperforms both label smoothing and no label smoothing (i.e. 1|-hot)

encodings (Figure 4).

0.87

oS

&

a

Test score Top3

oO

0.84

|— CLS@598

1 -A- Includes CLS@598

| | | | += Includes CLS@299

| a —<— Includes LabSmooth |

|

| --&- Includes 1-hot

|

a, ua 3 i 4 tian a % 6 i 7 a 8 ,

Number of Models in Ensemble

Figure 6: Ensemble Ablation: Only including 598x598 CLS networks in an ensemble with many

networks provides state of the art performance with significantly reduced training and

hyperparameter search and tuning costs compared to training a larger ensemble with a diversity of

networks. Combining CLS networks trained with smaller input image sizes or networks not trained

with CLS does not improve performance per network as much as adding another 598x598 CLS

network (top curve).

4.2. Image Size Ensemble Ablation

We trained ensembles of CLS on both smaller (299x299) and larger

(598x598) image input sizes into both IRV2 and IV4. The CLS performance

on larger images consistently outperforms CLS trained on smaller images,

whether on specific network types or ensembles of the same or different

network types (Figure 5).

4.3. CLS Ensemble Ablation

Throughout testing, we find that additional CLS networks trained on

larger input image sizes (598x598) improve ensembled results the most per

Washington Academy of Sciences

37

additional network in the ensemble. We find that unweighted network type

diversity (including networks trained with and without label-smoothing, i.e.

I-hot, IRV2 and IV4 architectures, and smaller input image sizes) do not

improve ensemble performance per additional network as much as adding

a CLS-trained network at a 598x598 input image size, indicating that CLS

with large imagery dominates the potential expected benefit of model

diversity in these ensembles. When ensembles contain four or more

networks, we observe that adding networks trained with either |-hot or

vanilla label smoothing label vectors can hurt performance.

Best Ensemble of CLS@598

| Final Public Score of 0.8805

0.38 ~

>

iNaturalist 2018 Winner vy’ Vv

' Final Public Score of 0.8693 aes

0.8 iin

eae oe ep ae er ee wee - - -

FS | Best Ensemble of CLS@299

2.0.86} Final Public Score of 0.8583 ae

2 Pune ndsechsessnctseshouresresasecauabensshescagencestesearaasacthned Barua eee eeseeeeeceeceneresecerescecersneessensnnetecesneess

3) ;

5 a

°

oO |

Ww

w 0.85

Ww .

Vv SS

a ae ie

’ a ee

0.84 ; y 4

EDS *

Pre A

> <a V_ IRV2 and Iv4 @598

0.83 pa y Vv IRV2 or IV4 @598

"% + IRV2 and 1V4 @299

+ IRV2 or IV4 @299

| o=0.0017

0.82! _ . —— ee eee ee ee

0.90 0.91 0.92 0.93 0.94

Test score prediction Top3

Figure 7: Test Score Error Analysis: By predicting the test error rate on the unseen test data based

on a test score prediction subset (1/3) of the validation data we can observe, we develop confidence

+/- 1-6 and 2-6 band estimates on the Test scores returned by the kaggle server on the unseen test

data. The iNaturalist 2018 Challenge final Test score winner as reported on the iNaturalist 2018

Challenge leaderboard [12] at 13% Top3 error is shown as a dashed line.

4.4. Test Performance Error Analysis

Using an empirical Monte Carlo approach we develop a Test score

predictor by fitting a line to the Test score as a function of the Test score

prediction and from this we estimate that our new CLS state of the art result

on iNaturalist 2018 Challenge test score has a +/- 0.17% 6 error (

Spring 2019

Best Ensemble of CLS@598

Final Public Score of 0.8805

0.88

iNaturalist 2018 Winner v ey V

Final Public Score of 0.8693 rf

OE a me a emia ra i Se eee == ——-—-—————=—-4

v

wey

wr

a Best Ensemble of CLS@299 lad

2.0.86 Final Public Score of 0.8583 a

fo)

=

—

4 0.85 ther

7) PS 3

cy) e

ld tig

tC.

+

0.84 ee

Pe

wie es 3 |

ie ‘J VV IRV2 and |lvV4 @598 |

oe

0.83 a Vv IRV2 or IV4 @598 |

be + IRV2 and |v4 @299 |

a IRV2 or IV4 @299

o = 0.0017

Oh, ee ooo a +

0.90 0.91 0.93 0.94

Test score prediction Top3

Figure 7). Our 1.0% improvement over the former state of the art

represents a greater than 5 6 improvement over the best prior reported

public test score of 0.8693 (compared to our 0.8805) with this estimate of

score variability.

5. Discussion

CLS shares training data among categories: By encoding non-

zero values, representative of proximity in the phylogenetic tree, in the label

vectors for categories that are not the true target category, CLS learns from

a more diverse set of examples than only those formally labeled as the

putative target type. In long-tailed FGVC tasks, we expect a number of

benefits from this approach.

In theory, for each target tail category, the relatively few training

examples of that category with their much larger label vector component

(0.8) will anchor the learned latent space of activations for that category

with data from that target category. Without full vector labels of any type

(i.e. 1-hot labels), the deep network could overfit to these relatively few

training examples of the target category (i.e. memorize them), suffering

poor generalization with no other information available to prevent this

overfitting. Relatively fewer categories (but each with more training

examples) from the head of the distribution that share the same branch of

the phylogenetic tree as the target category will also contribute to training

the target category. These examples will bias the learned latent space of

activations for the target tail category to move closer to those related head

Washington Academy of Sciences

39

categories, encouraging transfer learning from the head to the tail.

Relatively more non-target tail categories, each with fewer examples, will

more diffusely contribute to training the target tail category, ensuring that

the network does not overfit to either the relatively fewer training examples

of the target tail category or the more represented contributing head

categories.

In practice, any of these three effects may dominate, and rigorously

calibrating them is left for future work devoted to that detailed analysis to

compare to HSE. In addition to the rich relationships we exploit to improve

discriminative performance of species identification, it is also possible that

this approach could inform related research on ontological views of

relationships between different species. Specifically, the data-rich

categories from the head of the distribution might be used to stabilize,

communicate, and/or extend categorical relationships across hierarchies

(including predicates on the taxonomic relationships).

Focused Ensemble Performance with One Label Smoothing

Method: Since each CLS network at the 598x598 input size added to an

ensemble improves performance more than adding another marginal

network, this CLS benefit also reduces training time by focusing only on

the CLS-trained models. For instance, in our ensemble ablation, we see that

five CLS networks trained at the 598x598 input image size outperforms five

CLS networks with the addition of any other network type that is not CLS

598x598. This clarity allows us to focus computational resources on only

one type of network and not risk losing potentially beneficial diversity in

our ensembles that might accrue from other models with complementary

strengths had we trained them. This is a critical benefit to downstream work

comparing different methods because it guides efficient allocation of

limited compute resources on an already computationally intensive task.

Test Score Prediction Analysis: The scores from the test score

prediction set (part of the validation set, which entrants see) are highly

correlated with the test scores for the same model (network, or ensemble of

networks, e.g) on the unseen test data provided per blinded submission by

kaggle (

Spring 2019

40

Best Ensemble of CLS@598

Final Public Score of 0.8805

0.88

iNaturalist 2018 Winner v oy Vv

Final Public Score of 0.8693 wins |

0.87 ee ewe ween ee a eae eee eee ae He == = - - ee ----------------- 4

9

wr |

v

me

Pa Best Ensemble of CLS@299 ov |

2.0.86 Final Public Score of 0.8583 sun's

rs) 2

i

©

=

°

U

Ww

vw

wn

oO ¥

Ee as

i

} pe

0.84 : a

* “ be |

ae, V_ IRV2 and lv4 @598

0.83 : Vv IRV2 or 1V4 @598

+ IRV2 and IV4 @299

+ IRV2 or IV4 @299

o=0.0017

0:62 4. — ae ee Ne

0.90 0.91 0.94

a ae 0.92 a

Test score prediction Top3

Figure 7). In independent testing, we submitted a number of single

category labels to kaggle to interrogate the 1Naturalist 2018 Challenge test

data and found in each case that the resulting test scores were very close to

each other. This indicated that the mutually exclusive test set, while unseen

and held out from training and validation data, was likely uniformly

distributed over categories, as was the provided validation set. Based on this

insight, we used only a portion of the validation set for validation fine-

tuning (following [21]), leaving out a portion also uniformly distributed

over categories to predict the Test score. We found that a score computed

on this Test score prediction set was highly correlated with the actual Test

score.

We note the interrogation of the test set in this way does not confer

significant benefit on the Test score, as relatively tight bounds can be

estimated [25], and that large numbers of submissions will typically not

improve test scores. To wit, we did not tune, nor overfit to the test set here,

except to establish that it was uniformly distributed over categories.

By predicting the Test score from a presumably identically

distributed (over categories) Test score prediction set, we estimate a

conservative error bar on the Test score—meaning that the actual error bar

is likely smaller than our estimate. Specifically, the error bar fit estimate

degrades with both the Test score variability on the y-axis (the iNaturalist

2018 Challenge test score 6 we seek to estimate) as well as the prediction

test set score variability on the x-axis (which is a nuisance parameter). We

Washington Academy of Sciences

4]

cannot separate out these two sources of variability, but since the test set

has many more examples in it, we anticipate its contribution to the

estimation error, Otest, is smaller than the contribution to the estimation error

of the Test score prediction set, Opredict.

This error analysis helps in two ways. First, it provides a rough

measure of the real performance improvement from method to method

based on an empirically estimated confidence interval. Roughly, for CLS

that translates to slightly larger than an approximately 5 6 improvement

over the former state-of-the-art reported on the iNaturalist 2018 Challenge

[12]. Second, and more important to guide future work, such an estimation

error together with the measured performance improvement per marginal

ensemble network provides a rough means to estimate the expected

performance improvement per additional trained network in an ensemble.

This provides an ensembling stopping criterion to focus compute resources,

which, along with the insight of Contribution 3, that CLS improves

ensemble performance more per marginal network than other methods, is

critical to efficiently allocating compute resources for methodological

comparisons at scale (such as between CLS and HSE, e.g.) in downstream

work.

Improving Tail Category Performance with Fine-Tuning: Prior

work [21] inspired our adoption of fine-tuning on a more uniformly

distributed set of categories. In our case, we used a fraction of the validation

data for this purpose. We see similar gains in this work—i.e. CLS also

benefits from this fine-tuning approach.

6. Conclusion

The long tails of FGVC tasks for natural image corpora present

daunting training data collection requirements to achieve required accuracy

objectives on tail categories with mainstream deep learning methods.

Namely, the tail categories are many, sparse, and similar, making their per-

category accuracies difficult to improve on with |-hot labels that treat them

independently in training. In this work we demonstrate that CLS’

hierarchical prior on vector labels in the form of a phylogenetic tree can

pool training data contributions from many of the tail classes, exploit their

similarities, and thereby improve the accuracy on tail classes compared to

1-hot labels or other less judicious vector label smoothings.

CLS is Encoded by Domain Experts: The benefit of CLS alone is

significant and does not require expertise in deep learning to realize—the

phylogenetic tree prior came directly from a phylogenetic tree curated by

Spring 2019

42

biologists [15]. This is the only change from other methods [21]

benchmarked on this same dataset that we show underperform without CLS

compared to the same methods incorporating CLS.

CLS is Compatible with more Data-Driven Methods: While we

present results only on CLS without a CLS-specific hyperparameter search,

the CLS method proposed is compatible with more empirical distillation

and HSE methods which adjust label vectors based on training. Specifically,

CLS can be incorporated directly into the trunk network of HSE, for

instance. The CLS ensembles can be distilled into a single network to realize

the benefits of distillation, including distillation benefits of adversarial

example defense and compute reduction, e.g.

CLS’s Prior Models can be Extended by Human or Machine:

While we demonstrate a simple CLS approach that exploits an a priori

provided phylogenetic tree, this unlearned prior can very likely be

improved because the phylogenetic tree is not, by design, a guide to visual

similarity, even within a species. For instance, even within species, there

can be further training example pooling with visual similarity as encoded

through latent activation clustering. Among butterflies, for instance, the

within-species separation of chrysalis, caterpillar and butterfly stages may

create separable clusters in an embedding of latent activations (as with t-

SNE, e.g.). Within a bird species, the visual ornamentation of males vs.

females may similarly cluster in an embedding of latent activations.

Similarly, dog breeds may cluster. All of these finer levels may be similarly

encoded into the CLS prior by either machine or human curator. As with all

FGVC tasks, this presents additional challenges as training data fragments

among the categories because categories with very little training data are

split further, dividing the sparse training data among the finer subcategories.

We show that CLS can still effectively pool training data in that scenario at

the genus to species level of granularity and leave for future work the

demonstration of even more fine-grained applications of CLS.

Future Work: Demonstrating and evaluating the combined benefits

of both the a priori hierarchical CLS prior and the post hoc /earned latent

encodings of similarities (as in HSE and distillation, e.g.) together is left for

future work, as is the significant challenge of comparing other methods that

make use of the phylogenetic tree prior (like HSE) to CLS on the scale of

the iNaturalist 2018 dataset. For perspective, even with no CLS

hyperparameter tuning, the present study required >20,000 of GPU compute

time. The GPU compute costs of rigorously comparing HSE to CLS with

the hyperparameter searches required to reach conclusive results are

Washington Academy of Sciences

43

anticipated to be even larger, and may warrant additional AutoML

investigations, further increasing the computational costs.

7. Acknowledgements

This research was developed with funding from the Defense

Advanced Research Projects Agency. The views, opinions and/or findings

expressed are those of the author and should not be interpreted as

representing the official views or policies of the Department of Defense or

the U.S. Government. We thank the iNaturalist organization for providing

the iNaturalist 2018 dataset, phylogenetic tree, and the held out test data for

Kaggle blind scoring. We thank the anonymous reviewers for clarifying

revisions and highlighting important themes we had not sufficiently

emphasized in the original draft.

References

[1] S.Mayi, E. Rahtu, J. Kannala, M. Blaschko, and A. Vedaldi, “Fine-

Grained Visual Classification of Aircraft,” Jun. 2013.

[2] J. Krause, M. Stark, J. Deng, and L. Fei-Fei, “3d object

representations for fine-grained categorization,” in Proceedings of

the IEEE International Conference on Computer Vision Workshops,

2013, pp. 554-561.

[3] A.B. Hillel and D. Weinshall, “Subordinate class recognition using

relational object models,” in Advances in Neural Information

Processing Systems, 2007, pp. 73-80.

[4] T.L. Berg, A. C. Berg, and J. Shih, “Automatic attribute discovery

and characterization from noisy web data,” in European Conference

on Computer Vision, 2010, pp. 663-676.

[5] G. Van Horn et al., “The inaturalist species classification and

detection dataset,” 2018.

[6] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie, “The

caltech-ucsd birds-200-2011 dataset,” 2011.

[7] S. Hou, Y. Feng, and Z. Wang, “Vegfru: A domain-specific dataset

for fine-grained visual categorization,” in Computer Vision (ICCV),

2017 IEEE International Conference on, 2017, pp. 541-549.

[8] H. Goeau, P. Bonnet, and A. Joly, “Plant identification based on

noisy web data: the amazing performance of deep learning

(LifeCLEF 2017),” in CLEF 2017-Conference and Labs of the

Evaluation Forum, 2017, pp. 1-13.

Spring 2019

44

[9] J. Liu, A. Kanazawa, D. Jacobs, and P. Belhumeur, “Dog breed

classification using part localization,” in European Conference on

Computer Vision, 2012, pp. 172-185.

[10] “iNaturalist_ Competition FGVC_ 2018.pdf.” [Online]. Available:

https://www.dropbox.com/s/52nz6qce3zcewghoa/1Naturalist_ Competiti

on FGVC_2018.pdf?dl=0. [Accessed: 15-Nov-2018].

[11] iNaturalist competition details. Contribute to visipedia/inat_comp

development by creating an account on GitHub. Visipedia, 2018.

[12] “iNaturalist Challenge at FGVCS.” [Online]. Available:

https://www.kaggle.com/c/inaturalist-2018. [Accessed: 01-Jul-2018].

[13] G. Van Horn and P. Perona, “The Devil is in the Tails: Fine-grained

Classification in the Wild,” ArXiv Prepr. ArXiv170901450, 2017.

[14] ToChen, W. Wu, Y.- Gao, L. Dong) X. to, and [ey lin, “Fine=

grained representation learning and recognition by exploiting

hierarchical semantic embedding,” ArXiv Prepr. ArXiv180804505,

2018.

[15] “GBIF.” [Online]. Available: https://www.gbif.org/. [Accessed: 15-

Nov-2018].

[16] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. A. Alemi, “Inception-v4,

inception-resnet and the impact of residual connections on learning.,”

in AAAT, 2017, vol. 4, p. 12.

[17] A. Krizhevsky, I. Sutskever, and G. Hinton, “Imagenet classification

with deep convolutional neural networks,” in Advances in Neural

Information Processing Systems 25, 2012, pp. 1106-1114.

[18] C. Szegedy, V. Vanhoucke, S. loffe, J. Shlens, and Z. Wojna,

“Rethinking the inception architecture for computer vision,” in

Proceedings of the IEEE conference on computer vision and pattern

recognition, 2016, pp. 2818-2826.

[19] G. Pereyra, G. Tucker, J. Chorowski, \Lukasz Kaiser, and G. Hinton,

“Regularizing neural networks by penalizing confident output

distributions,” ArXiv Prepr. ArXiv170106548, 2017.

[20] G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a

neural network,” ArXiv Prepr. ArXiv150302531, 2015.

[21] Y. Cui, Y. Song, C. Sun, A. Howard, and S. Belongie, “Large Scale

Fine-Grained Categorization and Domain-Specific Transfer

Learning,” in Proceedings of the IEEE Conference on Computer

Vision and Pattern Recognition, 2018, pp. 4109-4118.

[22] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for

image recognition,” ArXiv Prepr. ArXiv151203385, 2015.

Washington Academy of Sciences

45

[23] O. Russakovsky ef al., “Imagenet large scale visual recognition

challenge,” Int. J. Comput. Vis., vol. 115, no. 3, pp. 211-252, 2015.

[24] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How transferable

are features in deep neural networks?,” in Advances in neural

information processing systems, 2014, pp. 3320-3328.

[25] K. Kawaguchi, L. P. Kaelbling, and Y. Bengio, “Generalization in

deep learning,” ArXiv Prepr. ArXiv171005468, 2017.

Bios

Michael Jeremy Trammell is a software engineer at Deep Learning

Analytics and leads the Deep BioThreatID project. He was recognized at

IEEE's CVPR conference in 2018 as part of the team that placed 2nd in

the world in the iNaturalist 2018 Challenge.

Priyanka Oberoi is a data scientist and head of ethics and fairness in

machine learning at Deep Learning Analytics. She received her Masters

degree in Biotechnology and Bioinformatics from Johns Hopkins

University. She was recognized at IEEE's CVPR conference in 2018 as

part of the team that placed 2nd in the world in the iNaturalist 2018

Challenge.

Jim Egenrieder is a fish and wildlife biologist and teaches Biodiversity

Stewardship and Watershed Systems Stewardship at Virginia Tech's

Center for Leadership in Global Sustainability in the National Capital

Region. He is also on the Research Faculty of Virginia Tech's College of

Engineering and is Director of the Virginia Tech Thinkabit

Lab™ teaching engineering and programming of microcontroller and

microprocessor circuits.

John Kaufhold is the Founder of Deep Learning Analytics, a machine

learning startup in the DC Metro Region. He received his Ph.D. in

Biomedical Engineering from Boston University as a Whitaker Fellow,

was named a Technical Fellow of SAIC (Leidos), and was recently named

a Fellow of the Washington Academy of Sciences. He was recognized at

IEEE's CVPR conference in 2018 as part of the team that placed 2nd in

the world in the iNaturalist 2018 Challenge.

Spring 2019

46

Washington Academy of Sciences

Washington Academy of Sciences

1200 New York Avenue

Rm G119

Washington, DC 20005

Please fill in the blanks and send your application to the address above. We will

contact you as soon as your application has been reviewed by the Membership

Committee. Thank you for your interest in the Washington Academy of Sciences.

(Dr. Mrs. Mr. Ms)

Business Address

Home Address

Email

Phone

Cell Phone

preferred mailing address Type of membership

Business Home Regular Student

Schools of Higher Education attended

Present Occupation or Professional Position

Please list memberships in scientific societies — include office held

47

Spring 2019

48

i)

IAI

ep

Instructions to Authors

Deadlines for quarterly submissions are:

Spring — February 1 Fall — August |

Summer — May | Winter — November |

Draft Manuscripts using a word processing program (such as

MSWord), not PDF. We do not accept PDF manuscripts.

Papers should be 6,000 words or fewer. If there are 7 or more graphics,

reduce the number of words by 500 for each graphic.

Include an abstract of 150-200 words.

Include a two to three sentence bio of the authors.

Graphics must be in greytone, and be easily resizable by the editors to

fit the Journal’s page size. Reference the graphic in the text.

Use endnotes or footnotes. The bibliography may be in a style

considered standard for the discipline or professional field represented

by the paper.

Submit papers as email attachments to the editor or to

wasjournal@washacadsci.oreg .

Include the author’s name, affiliation, and contact information —

including postal address. Membership in an Academy-affiliated society

may also be noted. It is not required.

Manuscripts are peer reviewed and become the property of the

Washington Academy of Sciences.

There are no page charges.

Manuscripts can be accepted by any of the Board of Discipline Editors.

Washington Academy of Sciences

Washington Academy of Sciences

Affiliated Institutions

National Institute for Standards & Technology (NIST)

Meadowlark Botanical Gardens

The John W. Kluge Center of the Library of Congress

Potomac Overlook Regional Park

Koshland Science Museum

American Registry of Pathology

Living Oceans Foundation

National Rural Electric Cooperative Association (NRECA)

Spring 2019

49

N

i=)

Delegates to the Washington Academy of Sciences

Representing Affiliated Scientific Societies

Acoustical Society of America

American/International Association of Dental Research

American Assoc. of Physics Teachers, Chesapeake

Section

American Astronomical Society

American Fisheries Society

American Institute of Aeronautics and Astronautics

American Institute of Mining, Metallurgy & Exploration

American Meteorological Society

American Nuclear Society

American Phytopathological Society

American Society for Cybernetics

American Society for Microbiology

American Society of Civil Engineers

American Society of Mechanical Engineers

American Society of Plant Physiology

Anthropological Society of Washington

ASM International

Association for Women in Science

Association for Computing Machinery

Association for Science, Technology, and Innovation

Association of Information Technology Professionals

Biological Society of Washington

Botanical Society of Washington

Capital Area Food Protection Association

Chemical Society of Washington

District of Columbia Institute of Chemists

Eastern Sociological Society

Electrochemical Society

Entomological Society of Washington

Geological Society of Washington

Historical Society of Washington DC

Human Factors and Ergonomics Society

(continued on next page)

Paul Arveson

J. Terrell Hoffeld

Frank R. Haig, S. J.

Sethanne Howard

Lee Benaka

David W. Brandt

E. Lee Bray

Vacant

Charles Martin

Vacant

Stuart Umpleby

Vacant

Vacant

Daniel J. Vavrick

Mark Holland

Vacant

Toni Marechaux

Jodi Wesemann

Vacant

F. Douglas

Witherspoon

Vacant

Vacant

Chris Puttock

Keith Lempel

Vacant

Vacant

Ronald W.

Mandersheid

Vacant

Vacant

Jurate Landwehr

Vacant

Gerald Krueger

Washington Academy of Sciences

Delegates to the Washington Academy of Sciences

Representing Affiliated Scientific Societies

(continued from previous page)

Institute of Electrical and Electronics Engineers, Washington

Section

Institute of Food Technologies, Washington DC Section

Institute of Industrial Engineers, National Capital Chapter

International Association for Dental Research, American

Section

International Society for the Systems Sciences

International Society of Automation, Baltimore Washington

Section

Instrument Society of America

Marine Technology Society

Maryland Native Plant Society

Mathematical Association of America, Maryland-District of

Columbia- Virginia Section

Medical Society of the District of Columbia

National Capital Area Skeptics

National Capital Astronomers

National Geographic Society

Optical Society of America, National Capital Section

Pest Science Society of America

Philosophical Society of Washington

Society for Experimental Biology and Medicine

Society of American Foresters, National Capital Society

Society of American Military Engineers, Washington DC

Post

Society of Manufacturing Engineers, Washington DC

Chapter

Society of Mining, Metallurgy, and Exploration, Inc.,

Washington DC Section

Soil and Water Conservation Society, National Capital

Chapter

Technology Transfer Society, Washington Area Chapter

Virginia Native Plant Society, Potowmack Chapter

Washington DC Chapter of the Institute for Operations

Research and the Management Sciences (WINFORMS)

Washington Evolutionary Systems Society

Washington History of Science Club

Washington Paint Technology Group

Washington Society of Engineers

Washington Society for the History of Medicine

Washington Statistical Society

World Future Society, National Capital Region Chapter

Richard Hill

Taylor Wallace

Neal F. Schmeidler

Christopher Fox

Vacant

Richard

Sommerfield

Hank Hegner

Jake Sobin

Vacant

John Hamman

Julian Craig

Vacant

Jay H. Miller

Vacant

Jim Heaney

Vacant

Larry S. Millstein

Vacant

Marilyn Buford

Vacant

Vacant

E. Lee Bray

Erika Larsen

Richard Leshuk

Alan Ford

Meagan Pitluck-

Schmitt

Vacant

Albert G. Gluckman

Vacant

Alvin Reiner

Alain Touwaide

Michael P. Cohen

Jim Honig

Washington Academy of Sciences NONPROFIT ORG

US POSTAGE PAID

Room GL117 MERRIFIELD VA 22081

1200 New York Ave. NW PERMIT# 888

Washington, DC 20005

Return Postage Guaranteed

Aagnersenis i sabeasaiteacs AUTO"MIXED ADC 207

HARVARD LAW S$ LIB ERSMCZ

LANGDELL HALL 152

1545 MASSACHUSETTS AVE

CAMBRIDGE, MA 02138-2903