Nathan Tsoi

I am a computer scientist working at the intersection of machine learning and robotics. I explore how to design and train neural networks to improve the effectiveness and robustness of robotic systems that interact with humans, with the ultimate goal of making robotic platforms more useful in the real world. Currently I am a PhD student at Yale University in robotics and a member of the Interactive Machines Group advised by Marynel Vázquez. Previously, I have done research at Stanford University in the Stanford Vision and Learning Lab under Silvio Savarese and have worked on machine learning and data engineering at Sequoia. For fun, I enjoy designing hardware and embedded systems programming.

Research

A Heaviside Function Approximation for Neural Network Binary Classification
Nathan Tsoi, Yofti Milkessa, Marynel Vázquez
Neural network binary classifiers are often evaluated on metrics like accuracy and $F_1$-Score, which are based on confusion matrix values (True Positives, False Positives, False Negatives, and True Negatives). However, these classifiers are commonly trained with a different loss, e.g. log loss. While it is preferable to perform training on the same loss as the evaluation metric, this is difficult in the case of confusion matrix based metrics because set membership is a step function without a derivative useful for backpropagation. To address this challenge, we propose an approximation of the step function that adheres to the properties necessary for effective training of binary networks using confusion matrix based metrics. This approach allows for end-to-end training of binary deep neural classifiers via batch gradient descent. We demonstrate the flexibility of this approach in several applications with varying levels of class imbalance. We also demonstrate how the approximation allows balancing between precision and recall in the appropriate ratio for the task at hand.
SEAN: Social Environment for Autonomous Navigation
Nathan Tsoi, Mohamed Hussein, Jeacy Espinoza, Xavier Ruiz, Marynel Vázquez
Proceedings of the 8th International Conference on Human-Agent Interaction
Social navigation research is performed on a variety of robotic platforms, scenarios, and environments. Making comparisons between navigation algorithms is challenging because of the effort involved in building these systems and the diversity of platforms used by the community; nonetheless, evaluation is critical to understanding progress in the field. In a step towards reproducible evaluation of social navigation algorithms, we propose the Social Environment for Autonomous Navigation (SEAN). SEAN is a high visual fidelity, open source, and extensible social navigation simulation platform which includes a toolkit for evaluation of navigation algorithms. We demonstrate SEAN and its evaluation toolkit in two environments with dynamic pedestrians and using two different robots.
Improving Social Awareness Through DANTE: Deep Affinity Network for Clustering Conversational Interactants
Mason Swofford, John Peruzzi, Nathan Tsoi, Sydney Thompson, Roberto Martín-Martín, Silvio Savarese, Marynel Vázquez
Proceedings of the ACM on Human-Computer Interaction
We propose a data-driven approach to detect conversational groups by identifying spatial arrangements typical of these focused social encounters. Our approach uses a novel Deep Affinity Networkto predict the likelihood that two individuals in a scene are part of the same conversational group, considering their social context. The predicted pair-wise affinities are then used in a graph clustering framework to identify both small (e.g., dyads) and large groups. The results from our evaluation on multiple, established benchmarks suggest that combining powerful deep learning methods with classical clustering techniques can improve the detection of conversational groups in comparison to prior approaches. Finally, we demonstrate the practicality of our approach in a human-robot interaction scenario. Our efforts show that our work advances group detection not only in theory, but also in practice.
Prompting Prosocial Human Interventions in Response to Robot Mistreatment
Joe Connolly, Viola Mocz, Nicole Salomons, Joseph Valdez, Nathan Tsoi, Brian Scassellati, Marynel Vázquez
ACM/IEEE International Conference on Human-Robot Interaction (HRI) 2020
Inspired by the benefits of human prosocial behavior, we explore whether prosocial behavior can be extended to a Human-Robot Interaction (HRI) context. More specifically, we study whether robots can induce prosocial behavior in humans through a 1x2 between-subjects user study ($N=30$) in which a confederate abused a robot. Through this study, we investigated whether the emotional reactions of a group of bystander robots could motivate a human to intervene in response to robot abuse. Our results show that participants were more likely to prosocially intervene when the bystander robots expressed sadness in response to the abuse as opposed to when they ignored these events, despite participants reporting similar perception of robot mistreatment and levels of empathy for the abused robot. Our findings demonstrate possible effects of group social influence through emotional cues by robots in human-robot interaction. They reveal a need for further research regarding human prosocial behavior within HRI.
Generalized Intersection over Union
Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, Silvio Savarese
Computer Vision and Pattern Recognition (CVPR) 2019
Intersection over Union (IoU) is the most popular evaluation metric used in the object detection benchmarks. However, there is a gap between optimizing the commonly used distance losses for regressing the parameters of a bounding box and maximizing this metric value. The optimal objective for a metric is the metric itself. In the case of axis-aligned 2D bounding boxes, it can be shown that $IoU$ can be directly used as a regression loss. However, $IoU$ has a plateau making it infeasible to optimize in the case of non-overlapping bounding boxes. In this paper, we address the weaknesses of $IoU$ by introducing a generalized version as both a new loss and a new metric. By incorporating this generalized $IoU$ ($GIoU$) as a loss into the state-of-the art object detection frameworks, we show a consistent improvement on their performance using both the standard, $IoU$ based, and new, $GIoU$ based, performance measures on popular object detection benchmarks such as PASCAL VOC and MS COCO.

Projects

Robots For Good: Fighting Social Isolation with Robots
Robotic telepresence for elementary-age children during social distancing.
Multi-sourced 2D and 3D Sensor Fusion and Person Tracking Pipeline
For research in imitation learning, creating motion policies for social navigation
Darkboard
A Tensorboard-like visual interface for Darknet, available as part of g-darknet

Awards

Alan J. Perlis Graduate Fellowship Recipient
This fellowship was established at Yale in 2006 through generous gifts from various donors in honor of Professor Alan J. Perlis (1922–1990), a pioneer of programming language research, the first winner of the Association for Computing Machinery’s (ACM) Turing Award, and the founding chair of Yale’s Computer Science Department.

Service and Activities

YHACK 2019 Final Judge