Natasha Jaques
Natasha Jaques
Awards
Press
Featured
Publications
Topics
Talks
Communities
Light
Dark
Automatic
3
Concept-based Understanding of Emergent Multi-Agent Behavior
Interpreting whether multi-agent reinforcement learning (MARL) agents have successfully learned to coordinate with each other, versus finding some other way to exploit the reward function, is a longstanding problem. We develop a novel interpretability method for MARL based on concept bottlenecks, which enables detecting which agents are truly coordinating, which environments require coordination, and identifying lazy agents.
N. Grupen
,
Natasha Jaques
,
B. Kim
,
S. Omidshafiei
2022
In
Preprint
Cite
Moral Foundations of Large Language Models
Moral Foundations theory decomposes human moral reasoning into five factors, which vary reliably across different human populations and political affiliations. We use moral foundations to analyze large language models like GPT-3 to determine what, if any, consistent moral values it brings to conversations, whether these can be deliberately manipulated, and whether holding a particular moral stance affects downstream tasks.
M. Abdulhai
,
C. Crepy
,
D. Valter
,
J. Canny
,
S. Levine
,
Natasha Jaques
2022
In
Preprint
Cite
In the ZONE: Measuring difficulty and progression in curriculum generation
Past work on curriculum generation in RL has focused on training a teacher agent to generate tasks for a student agent that accelerate student learning and improve generalization. In this work, we create a mathematical framework that formalizes these concepts and subsumes prior work, taking inspiration from the psychological concept of the Zone of Proximal Development. We propose two new techniques based on rejection sampling and maximizing the student’s gradient norm that improve curriculum learning.
R. E. Wang
,
J. Mu
,
D. Arumugam
,
Natasha Jaques
,
N. Goodman
2022
In
Preprint
Cite
Multi-Agent Reinforcement Learning for Hardware Architecture Search: A Case Study on Domain-Specific DRAM Memory Controller Design
Reinforement Learning can potentially be a powerful tool for solving complex combinatorial optimization problems, such as microprocessor desgin. Here, we show that a multi-agent RL approach outperforms past work using single agent RL, since the problem can easily be decomposed into designing independent sub-systems.
S. Krishnan
,
Natasha Jaques
,
S. Omidshafiei
,
D. Zhang
,
I. Gur
,
V. J. Reddi
,
S. Faust
2022
In
Preprint
Cite
Basis for Intentions: Efficient Inverse Reinforcement Learning using Past Experience
Using inverse reinforcement learning to infer human preferences is challenging, because it is an underspecified problem. We use multi-task RL pre-training and successor features to learn a strong prior over the space of reasonable goals in an environment—which we call a
basis
—that enables rapidly inferring an expert’s reward function in only 100 samples.
M. Abdulhai
,
Natasha Jaques
,
S. Levine
2022
In
Preprint
PDF
Cite
Code
Project
A Comparison of Random Forests and Dropout Nets for Sign Language Recognition with the Kinect
We conduct a study in which participants form American Sign Language hand signs while being recorded with a Microsoft Kinect. The resulting infra-red distance data are used to train both neural networks with dropout (dropout NN) and Random Forests; dropout NN perform significantly better.
Natasha Jaques
,
J. Nutini
2013
In
Unpublished manuscript
PDF
Cite
Emotionally Adaptive Intelligent Tutoring Systems using POMDPs
An emerging field in user-adaptive systems is affect adaptivity: modeling and responding to an estimation of the user’s emotional state. Prior work used Dynamic Bayesian Networks to obtain adaptivity, but in this paper we represent the problem as a Partially Observable Markov Decision Process (POMDP) and find solutions that compute a plan of interventions for an Intelligent Tutoring System to take given an estimation of the user’s mood and goals.
Natasha Jaques
2013
In
Unpublished manuscript
PDF
Cite
Fast Johnson–Lindenstrauss transform for classification of high dimensional data
This paper investigates the utility of using the Fast Johnson-Lindenstrauss Transform to produce a low-dimensional random projection of eye-tracking data features that can be used for classifying emotion in an Intelligent Tutoring System. Interestingly, the FJLT provides similar or superior performance to more computationally expensive techniques.
Natasha Jaques
2013
In
Unpublished manuscript
PDF
Cite
Cite
×