Reverse predictivity for bidirectional comparison of neural networks and biological brains

2026-03-25 · Source: Nature Machine Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computational Neuroscience, Data Science & Analytics · Depth: Expert, short

Summary

A study published in Nature Machine Intelligence on March 25, 2026, introduces "reverse predictivity," a new diagnostic metric for comparing artificial neural networks (ANNs) with biological brains. Unlike traditional forward predictivity, which assesses how well model features predict neural responses, reverse predictivity quantifies how accurately macaque inferior temporal cortex responses predict ANN unit activations. The research reveals a significant asymmetry: ANNs with high forward predictivity (explaining approximately 50% variance) often contain units that are unpredictable from neural activity, indicating biologically implausible dimensions. In contrast, monkey-to-monkey brain mappings exhibit symmetry, serving as a biological reference. This framework identifies "common" ANN units that align with biological brains, are behaviorally relevant, and generalize across species, as well as "unique" units lacking such alignment. Reverse predictivity is influenced by factors like feature dimensionality, training objectives, and adversarial robustness, positioning it as a conservative tool to guide the development of next-generation ANNs towards both enhanced task performance and greater biological plausibility.

Key takeaway

For AI researchers and neuroscientists developing brain-inspired models, integrating reverse predictivity into your evaluation pipeline is crucial. This metric provides a more stringent test of biological alignment than forward predictivity alone, helping you identify and refine ANN architectures that truly mimic neural representations. Focus on optimizing models for high reverse predictivity to achieve both robust task performance and greater biological realism.

Key insights

Reverse predictivity offers a bidirectional metric to assess biological plausibility in artificial neural networks.

Principles

• Forward predictivity alone is insufficient for biological plausibility.
• Biological brain mappings exhibit symmetric predictivity.
• ANN unit predictability from neural activity indicates biological alignment.

Method

The method involves quantifying how well macaque inferior temporal cortex responses predict ANN unit activations, contrasting this with traditional forward predictivity, and identifying "common" versus "unique" ANN units.

In practice

• Use reverse predictivity to evaluate ANN biological plausibility.
• Prioritize training objectives that enhance reverse predictivity.
• Analyze "common" units for cross-species generalization.

Topics

• Reverse Predictivity
• Artificial Neural Networks
• Systems Neuroscience
• Biological Plausibility
• Inferior Temporal Cortex

Code references

• vital-kolab/reverse_pred

Best for: AI Researcher, AI Scientist, Research Scientist

Related on AIssential

• The Predictive Mind: How Human Brains and AI Speak the Same Language — Human brains and large language models both operate on predictive processing, continuously generating expectations and updating based on new information.
• What’s wrong with modeling? — Bottom-up neuroscience modeling can hinder functional understanding by pre-selecting data and overlooking crucial biological mechanisms.
• Why Every Brain Metaphor in History Has Been Wrong [SPECIAL EDITION] — Scientific models are useful simplifications, not necessarily literal truths about underlying reality.
• A Shared Valence Axis Across Modern LLMs and Human EEG: The Saturation Regularity — A shared emotional valence axis emerges independently in large language models and human brain activity, but direct alignment is counterproductive.
• Harnessing cortical geometry, wiring, and function as inductive biases for recurrent neural networks — Cortical geometry, wiring, and function serve as powerful inductive biases for more effective recurrent neural networks.
• Synaptic Activation and Dual Liquid Dynamics for Interpretable Bio-Inspired Models — Dual liquid dynamics and synaptic activation significantly enhance RNN interpretability in bio-inspired models.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Nature Machine Intelligence.