Welcome back to Beyond the Patterns. So today I have the great pleasure to announce Timothy

Odonga. He holds a master's degree in electrical and computer engineering from Carnegie Mellon

University and two bachelor's degrees in physics and electrical engineering from Gordon

College and the University of Southern California respectively. He has experience on working

on research projects in machine learning at CMU and IBM research. During his time at IBM

research he was an IBM Great Minds scholar and an AI for Good fellow as he worked on

a project on AI fairness in dermatology. The research papers from this work were accepted

and published at the MICHI 2020 conference and at NURIP's FAIR Machine Learning for

Health workshop in 2019. His research interests include machine learning for healthcare focusing

on topics like fairness, explainability and causality. So it's a great pleasure to have

Timothy here and I'm very very much looking forward to his presentation. This is really

an important topic and I think we should consider these topics much more often. So Timothy,

it's a great pleasure to have you here and the title today will be entitled Fairness

of Classifiers across skin tones in dermatology. So Timothy, the stage is yours.

Alright, Andreas, I would like to say thank you for the introduction and as Andreas had

said the title of my presentation is Fairness of Classifiers across skin tones in dermatology.

As Andreas had mentioned this work was featured at MICHI 2020 and at the NURIP FAIR 2019 workshop.

Before I start to present I would like to acknowledge the team of incredible researchers

I worked with at IBM from the labs in Nairobi, Yorktown Heights and Cambridge. So I will

start my talk with the background and make a case for fairness. Then I'll mention the

research questions that motivated our work. Then I'll move on to the approaches we based

our work on. Then I will move on to the datasets and explain our fairness pipeline and lastly

I will share the results from the research work. So as machine learning becomes more

pervasive in our lives it's important to acknowledge the effect of bias. A study by Elamian Group

from 2019 noted that bias is an issue because machine learning in its nature is a form of

statistical discrimination. This discrimination is objectionable if it places certain groups

of people at a systematic disadvantage or advantage. A study by Barocas and Selt from

2016 also acknowledged that machine learning systems may place certain groups of people

at a systematic disadvantage due to dataset bias. Therefore it's important that we recognize

and mitigate unwanted bias when we build machine learning systems so that they are trusted

in the essential domains of deployment. Dermatology like other medical fields has disparities

that exist with respect to ethnicity. Two studies, one from 2017 by Mahendra Rajan Group

and a study from 2011 by Wu and Group found that in the Black population melanoma is often

diagnosed at an advanced stage with deeper tumors. A 2015 study by Maphetian Group found

that the fiber survival rate for acro lentiginous melanoma, ALM, was higher with whites than

in blacks in spite of the fact that there's a lower incidence of hitting blacks. A recent

study by Lester and Group from 2020 noted that the possibility of images of skin manifestations

of COVID-19 in patients with darker skin is a problem because it may make identification

of COVID-19 presenting with cutaneous manifestations more difficult for both dermatologists and

the public. Therefore it's important that the systems we build do not propagate these

disparities. So our research work was motivated by two questions. One, standard dermatology

image data sets used in machine learning tasks biased with respect to skin tone. Can we quantify

this bias? And secondly, if so, does the data set bias lead to an equal performance of downstream

disease classification? So our work was based on research in skin disease diagnosis using

machine learning and research on predictive inequity in computer vision with respect to

skin type. Our work was based on research in skin disease diagnosis, specifically the

benchmark model for melanoma diagnosis that outperformed trained dermatologists in 2016

by Odella and Group. Our work was also based on the challenges held by the International

Skin Imaging Collaboration and specifically the 2018 challenge that involved skin lesion

segmentation and skin disease classification. Our work was based on research in predictive