DATA SCIENCE ENGINEER, AV BEHAVIOR TESTING at NVIDIA

--Develop algorithms to quantify AV performance.

--Develop tools to automate the testing and evaluation of software changes on AV behavior.

--Use simulation and testing tools to build scenarios, metrics and validate AV performance.



DATA RESEARCH at PERCEPTIVE AUTOMATA

Understanding pedestrian behavior and identifying social driving characteristics for various autonous driving applications.

-- Understanding pedestrian behavior in pick-up drop-off areas for which I analyze and evaluate model predictions for complicated driving scenes

-- Understanding social driving characteristics for which I develop quantifiable metrics and tools for evaluation of psychophysics experiments aimed at understanding pedestrian state of mind and social driving.


I was part of a two member team that won first place in Innovation-week of the first quarter of 2022 for developing tools to investigate performance of our models in crowds.

PhD, BRAIN and COGNITIVE SCIENCES, UNIVERSITY OF ROCHESTER

For my PhD thesis, my broad scientific aim is to understand the mechanisms in the brain underlying formation of visual percepts, perceptual beliefs about the outside world, temporal and spatial perceptual biases, biases in action selection and seeking new information and meta-cognitive confidence judgements.

I studied this in the context of probabilistic inferences in the brain about the outside world by combining sensory information and prior knowledge about the structure of the external world. I design perceptual decision making expeirments for humans with the goal to understand perceptual biases in human behavior as a consequence of sampling-based approximate inference on a hierarchical generative model of the world that the brain has previously learnt. Finally, I aim to investigate if biophysically realistic neurons can implement sampling based inference, bridging all of Marrs three levels, from assuming a computational goal (probabilistic inference over visual inputs) to an algorithm (neural sampling) to neural implementation (network of leaky integrate-and-fire neurons LIF neurons).

I am passionate about working on projects that involve social good such fairness and crime prediction and also enjoy working on machine learning projects investigating theory motivated applications.

Investigating visual perceptual biases as a consequence of approximate inference in a hierarchical generative model of visual decision making tasks.

-- Understanding the primacy and recency biases empirically observed in temporal weighting of evidence in integration tasks. We designed a visual 2-alternative forced choice based discrimination task, developed a hierarchical generative model of the task and simulated sampling based approximate inference to explain and understand various temporal perceptual biases in visual decision making. Paper

-- Using the same paradigm to investigate how spatial and temporal statistics of the stimuli influence the nature and magnitude of temporal perceptual bias in visual decision making. Paper

-- Understanding how perceptual biases and meta-cognitive confidence judgements about correctness of a choice made in decision making tasks are related. We study this especially in the context of predictions from our sampling based approximate inference. Paper

-- Extending the idea of understanding bias in temporal weighting of evidence to understanding biases while seeking new evidence before making a choice. We specifically study whether humans are biased to saccade to evidence favoring their current belief about the correct answer in a trial. Paper


Building a biophysically realistic network of neurons that can implement sampling based inference on a generative model of retinal inputs.

-- Implemented a network of LIF neurons to demonstrate how biophysically realistic neurons can perform Gibbs sampling based inference on a sparse linear Gaussian model of retinal input. Paper

-- Resolved a debate in computational neuroscience, whether neural responses represent sampling code or parametric code by proving their equivalence in a sparse linear Gaussian model of retinal input. Paper

-- Proposed a new paradigm and offered an aletrnative perspective of thinking about and categorizing different models proposed to explain neural responses, aka, Bayesian encoding models and Bayesian decoding models. Paper


Projects for social good: fair rating predictor for public speeches and crime prediction model

-- Developed first of a kind fair rating predictor system for public speaking with respect to speakers race and gender, using counterfactual fairness and causal models on a corpus of TED talk data. Paper

-- Proposed a novel heterogeneity based metric to quantify quality of a speech in multimodal domain (verbal: transcript and non-verbal: facial gesture) and incorporated it into a fair rating prediction for speakers of TED talks. Paper

-- Developed a spatial crime prediction model for crime data from the city of Rochester. Paper


Theory motivated applications of Machine Learning

-- Designed a novel diversity based edge pruning method for feedforward neural networks based on Determinantal Point Processes which outperforms previously known diversity based pruning techniques and theoretically analyzed its performance using generalization error bounds. Paper

-- Introduced a unifying generalization of the Lovasz theta function and the associated geometric embedding. We then showed how the theta function can be interpreted as a measure of diversity in graphs thereby incorporating it in Max-Cut, correlation clustering and document summarization algorithms. Paper