OpenAI Scholars 2020 Final Projects

OpenAI has released the final projects from the 2020 Scholars program, featuring a diverse set of research initiatives ranging from the interpretability of large language models to the application of deep learning in neuroscience. These projects demonstrate the application of AI to solve complex problems in credit assignment, social learning, and adversarial robustness.

Neural Network Interpretability and Language Models

Research in the 2020 cohort focused heavily on how models represent information and how they can be manipulated via adversarial inputs.

Grammatical Representation in GPT-2

Alethea Power investigated how GPT-2 represents English grammar, identifying smaller sub-networks that appear to correspond to specific grammatical structures. This work aims to improve network efficiency and provide insights into both model behavior and human cognition.

Universal Adversarial Perturbations

Pamela Mishkin explored universal adversarial examples in the context of language models. The research involved replicating results to generate universal adversarial triggers for GPT-2 and attacking Natural Language Inference (NLI) models to better understand the vulnerabilities of generative models.

Reinforcement Learning (RL) and Agent Behavior

Several projects addressed fundamental challenges in RL, specifically regarding reward sparsity, multi-agent interaction, and model quantification.

Temporal Reward Transport (TRT)

Cathy Yeh addressed the long-term credit assignment problem—where sparse rewards and long delays between action and effect hinder sample efficiency. By using an attention mechanism to identify significant state-action pairs and augmenting them with rewards from the distant future via "temporal reward transport," Yeh demonstrated improved learning in gridworld experiments using a standard advantage actor-critic algorithm.

Social Learning in Multi-Agent RL

Kamal Ndousse researched the social transfer of expertise among independent RL agents in shared environments. The project investigated whether novice agents could solve hard-exploration tasks by mimicking the behavior of expert agents, which they were unable to master in isolation.

Quantifying Interpretability in Coinrun

Jorge Orbay developed a scalar metric to measure the interpretability of an A2C model trained on Procgen’s Coinrun. This metric was created by combining model attribution with masks of Coinrun’s assets to test the validity of the diversity hypothesis.

Applied AI and Semantic Parsing

Beyond theoretical RL and interpretability, the program produced work on practical interface translation and medical diagnostics.

English-to-GraphQL Semantic Parsing

Andre Carerra developed a system for semantic parsing that translates English prompts (e.g., "How many employees do we have?") into corresponding GraphQL queries. The project included the creation of a custom dataset, model training, and an interaction tool for result verification.

Epileptic Seizure Prediction

Kata Slama applied deep learning to brain recordings to predict epileptic seizures. By framing the problem as an image classification task using spectrogram representations of brain data, Slama identified ResNet18 as the most successful model for this application.