NeuroAI for AI Safety

As AI systems become increasingly powerful, the need for safe AI has become more pressing. Humans are an attractive model for AI safety: as the only known agents capable of general intelligence, they perform robustly even under conditions that deviate significantly from prior experiences, explore the world safely, understand pragmatics, and can cooperate to meet their intrinsic goals. Intelligence, when coupled with cooperation and safety mechanisms, can drive sustained progress and well-being. These properties are a function of the architecture of the brain and the learning algorithms it implements. Neuroscience may thus hold important keys to technical AI safety that are currently underexplored and underutilized. In this roadmap, we highlight and critically evaluate several paths toward AI safety inspired by neuroscience: emulating the brain's representations, information processing, and architecture; building robust sensory and motor systems from imitating brain data and bodies; fine-tuning AI systems on brain data; advancing interpretability using neuroscience methods; and scaling up cognitively-inspired architectures. We make several concrete recommendations for how neuroscience can positively impact AI safety.

Centaur: a foundation model of human cognition

Establishing a unified theory of cognition has been a major goal of psychology. While there have been previous attempts to instantiate such theories by building computational models, we currently do not have one model that captures the human mind in its entirety. Here we introduce Centaur, a computational model that can predict and simulate human behavior in any experiment expressible in natural language. We derived Centaur by finetuning a state-of-the-art language model on a novel, large-scale data set called Psych-101. Psych-101 reaches an unprecedented scale, covering trial-by-trial data from over 60,000 participants performing over 10,000,000 choices in 160 experiments. Centaur not only captures the behavior of held-out participants better than existing cognitive models, but also generalizes to new cover stories, structural task modifications, and entirely new domains. Furthermore, we find that the model's internal representations become more aligned with human neural activity after finetuning. Taken together, Centaur is the first real candidate for a unified model of human cognition. We anticipate that it will have a disruptive impact on the cognitive sciences, challenging the existing paradigm for developing computational models.

Predecessor Features

Any reinforcement learning system must be able to identify which past events contributed to observed outcomes, a problem known as credit assignment. A common solution to this problem is to use an eligibility trace to assign credit to recency-weighted set of experienced events. However, in many realistic tasks, the set of recently experienced events are only one of the many possible action events that could have preceded the current outcome. This suggests that reinforcement learning can be made more efficient by allowing credit assignment to any viable preceding state, rather than only those most recently experienced. Accordingly, we propose "Predecessor Features", an algorithm that achieves this richer form of credit assignment. By maintaining a representation that approximates the expected sum of past occupancies, our algorithm allows temporal difference (TD) errors to be propagated accurately to a larger number of predecessor states than conventional methods, greatly improving learning speed. Our algorithm can also be naturally extended from tabular state representation to feature representations allowing for increased performance on a wide range of environments. We demonstrate several use cases for Predecessor Features and contrast its performance with other similar approaches.

Improving Experience Replay with Successor Representation

Prioritized experience replay is a reinforcement learning technique shown to speed up learning by allowing agents to replay useful past experiences more frequently. This usefulness is quantified as the expected gain from replaying the experience, and is often approximated as the prediction error (TD-error) observed during the corresponding experience. However, prediction error is only one possible prioritization metric. Recent work in neuroscience suggests that, in biological organisms, replay is prioritized by both gain and need. The need term measures the expected relevance of each experience with respect to the current situation, and more importantly, this term is not currently considered in algorithms such as deep Q-network (DQN). Thus, in this paper we present a new approach for prioritizing experiences for replay that considers both gain and need. We test our approach by considering the need term, quantified as the Successor Representation, into the sampling process of different reinforcement learning algorithms. Our proposed algorithms show a significant increase in performance in benchmarks including the Dyna-Q maze and a selection of Atari games.