Mouse vs. AI: A Neuroethological Benchmark for Visual Robustness and Neural Alignment.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computer Vision

Background:

• Modern reinforcement learning (RL) agents struggle with visual robustness in real-world conditions.
• Biological systems like mice exhibit remarkable resilience to environmental changes and degraded visual input.

Purpose of the Study:

• Introduce the Mouse vs. AI: Robust Foraging Competition, a bioinspired benchmark for RL agent generalization.
• Evaluate RL agents' ability to generalize to unseen, ecologically realistic visual perturbations in a naturalistic 3D environment.
• Facilitate the development of robust, generalizable, and biologically inspired AI by integrating RL, computer vision, and neuroscience.

Main Methods:

• Participants train RL agents for a visually guided foraging task in a 3D Unity environment.
• Agents are evaluated on generalization to novel visual perturbations (Track 1: Visual Robustness).
• Agent internal representations are assessed for predicting mouse visual cortical activity using linear readout (Track 2: Neural Alignment).
• A fog-perturbed training condition is provided for validation, alongside baseline proximal policy optimization (PPO) agents.

Main Results:

• The competition provides a platform for benchmarking RL agents against real mouse behavior and neural data.
• It enables the assessment of generalization capabilities across diverse visual perturbations.
• It allows for the evaluation of neural alignment between AI representations and biological systems.

Conclusions:

• The Mouse vs. AI competition advances AI development by grounding it in biological principles and shared tasks.
• It fosters the creation of more robust and generalizable AI systems capable of real-world performance.
• This bioinspired approach bridges AI, computer vision, and neuroscience, paving the way for future research.