[Paper Review] NeuroAI for AI Safety

Patrick Mineault, Niccolò Zanichelli+13 more|arXiv (Cornell University)|Nov 27, 2024

Fault Detection and Control Systems4 citations

TL;DR

This paper proposes NeuroAI—integrating neuroscience insights into AI safety by emulating brain architectures, learning algorithms, and neural representations to enhance robustness, interpretability, and alignment in artificial intelligence. It demonstrates that brain-inspired designs can mitigate risks in agentic AI, with key results showing improved generalization and safety through neuroscience-informed architectures and data-driven training.

ABSTRACT

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.

Motivation & Objective

• Address long-term AI safety concerns related to future agentic and general-purpose AI systems.
• Explore how neuroscience can inform technical solutions to AI safety beyond current prosaic AI limitations.
• Identify and evaluate brain-inspired approaches to improve AI robustness, cooperation, and goal alignment.
• Provide actionable recommendations for integrating neuroscience into AI safety research and development.

Proposed method

• Adapt DeepMind’s 2018 framework for AI safety to map neuroscience-inspired solutions across robustness, interpretability, and alignment.
• Emulate brain-like representations and information processing through cognitively-inspired neural architectures.
• Fine-tune AI models using brain data (e.g., electrophysiology, calcium imaging) to improve generalization and safety.
• Apply neuroscience-based interpretability tools such as representational similarity analysis and shared variance component analysis (SVCA) to probe model internals.
• Scale up brain-inspired architectures using large-scale neural datasets and analyze their dimensionality and scalability.
• Use Bayesian linear regression and log-log scaling laws to model how neural data dimensionality scales with recording size.

Experimental results

Research questions

• RQ1 How can brain architecture and learning algorithms inform the design of safer, more robust AI systems?
• RQ2 To what extent can fine-tuning AI models on neural data improve their out-of-distribution generalization and safety?
• RQ3 Can neuroscience-based interpretability methods enhance transparency and alignment in artificial intelligence?
• RQ4 What is the scaling behavior of neural data dimensionality across species and recording modalities?
• RQ5 How can freely available neural datasets be leveraged to train and evaluate safer AI systems?

Key findings

• Neural data dimensionality scales sublinearly with the number of recorded neurons, following a power law with exponent β ≈ 0.5–0.7 across species.
• Shared Variance Component Analysis (SVCA) reliably estimates neural dimensionality by measuring variance explained in test sets after orthogonal decomposition.
• The total amount of freely available neural data across repositories like DANDI, OpenNeuro, and iEEG.org exceeds 100,000 hours of recording, with major contributions from Human Connectome Project and UK Biobank.
• Vision models on the BrainScore leaderboard show a quadratic improvement trend in neural similarity over time, indicating accelerating alignment with biological visual processing.
• Brain-inspired architectures trained on neural data exhibit enhanced robustness to distributional shift and improved interpretability via representational similarity to cortical areas.
• Neural data from diverse modalities—electrophysiology, calcium imaging, fMRI—can be standardized and used to train AI systems with safer, more human-like inductive biases.