Machine Learning and Data Science (June 2016—Present)
Explainable Artificial Intelligence
Using Bayesian teaching, our goal is to build a machine that uses examples to explain opaque models, where the examples are very small subsets of the original training data. We demonstrate the efficacy of this approach to model explanation through human experimentation via Amazon Mechanical Turk.
For our first milestone, I have implemented and tested inference for a probabilistic linear discriminant analysis model, previously used for facial recognition. My implementation of the model, demo, and corresponding unit and inference tests can be found here. In addition to writing a teaching model to generate examples that explain learned parameters, I also implemented various emotion classification and discrimination tasks for model simulations and cross validation, using the Google Faces and Child Affective Facial Expressions datasets.
The repository containing my teaching model code, derivations, and implementation details will remain private until all experimental work is completed and submitted for publication — I am currently analyzing data obtained from our human experiments.
Topological Data Analysis of Video
From empirical insights about the human visual system and the statistics of naturalistic input, we are representing video data on significantly lower dimensional manifolds for scalable, unsupervised learning of features across spatiotemporal frequencies. In addition to implications for compression and generative models of video, this approach addresses important limitations of procedures relying on predefined Gabor filter banks and Fourier transformations for characterizing visual input.
This project began as a journal club meeting (reading), for which I derived the mathematics and intuitions required to reproduce some of the main results. The code I wrote to replicate those results can be found here. After extending this approach to video, I used a Dirichlet Process mixture model to show that this unsupervised approach also discovers features one expects to generate signals in existing approaches (e.g. translation of oriented lines).
I am currently leading a group of Master’s students and undergraduate research assistants to optimize and scale this approach to work on very small spatiotemporal frequencies (i.e. large video patches).
We are analyzing the long run consequences of human-algorithm interactions (e.g. Google Search, Amazon Recommendation, and Facebook News Feed). The idea here is that these services make assumptions about the distributions of data and feedback signals (e.g. clickthroughs), introducing non-optimal inductive biases into the machine inference process. Because these systems then provide recommendations to billions of users, inducing changes in consumer behavior that becomes further input into these systems, it is important to understand the limiting behavior of such interactions.
Using a concept learning learning framework and Markov chain theory, I am showing that in the present day setup of human-algorithm interactions, both recommendation quality and feedback for the algorithm can decline as a function of market share, without setting off obvious feedback signals (e.g. in clickthroughs, purchases, and “Likes”).
Neuroscience (January 2012—June 2016)
Striatal and Hippocampal Effects of Aging
Collecting behavioral and single-nucleotide polymorphism (genetic) data, we studied the effects of healthy cognitive aging on learning that is dependent on striatal and hippocampal neural structures.
This was my first foray into academic research. Over four years, I personally collected data from hundreds of participants, seeing the project from start to completion. During this time, I contributed to all aspects of the research pipeline and was leading data and journal club meetings. Before moving onto my present work, I was fitting reinforcement learning models to behavioral data.
I presented my research contributions to this project at various conferences (SfN, CogSci, and CNS) and have a manuscript that was recently accepted for publication in Neurobiology of Aging.
Striatal and Hippocampal Effects of Exercise and Lifestyle Factors
To study the effects of cardiovascular fitness on striatal and hippocampal dependent learning, we employed several brain imaging techniques (including fMRI), evaluated physical fitness, and conducted exercise interventions, in addition to collecting behavioral and genetic data.
This project started about a year before I left to pursue my current research and is still in the process of collecting data from human experiments (exercise interventions operate on the scale of 1-6 months). However, I presented some work at an annual meeting of the Cognitive Neuroscience Society that hinted at potential benefits of medicating hypertension for striatal based learning. This is interesting because previous work on cardiovascular fitness and intervention has primarily focused on changes in structures such as the frontal lobe and hippocampus. Results showing brain changes in the striatum as a function of cardiovascular fitness would shed an interesting light on reinforcement learning and cognitive aging.
Service and Outreach (January 2012—June 2016)
African American Brain Health Initiative
The objectives of this project are to increase brain health awareness and research participation in predominantly African American communities, especially in Newark and the Greater Newark Area. We increased participation in research experiments, stimulated conversations about research that ought to be conducted in local communities, and convinced and trained young (high school and college) students of color to pursue careers in neuroscience.
During my final year, my contribution to this project included ~30 talks to audiences of ~15-100 people at churches, senior centers, and publicly assisted housing units in Newark, NJ. The content of these talks included sharing information about increased risk for brain disease in African American communities, lifestyle habits that are known to predict better brain health outcomes, and the need for more research participation from minority communities in both experiments and the scientific process. I dedicated a significant amount of time during these events for Q&A to address questions and concerns about research participation, ethics, rights of participants, and handling of collected data.
Over four years, I handled logistics of larger events by managing teams of students, and for smaller audiences, I handled the logistics myself. I also helped post-doctoral fellows train students for research careers in psychology and neuroscience.