Author Correction: Foundation model of neural activity predicts response to new stimulus types

2026-04-08 · Source: Machine learning : nature.com subject feeds · Field: Science & Research — Life Sciences & Biology, Mathematics & Computational Sciences · Depth: Expert, short

Summary

An author correction published in Nature on April 8, 2026, clarifies architectural details for the Conv-LSTM and CvT-LSTM models described in an original article from April 9, 2025, titled "Foundation model of neural activity predicts response to new stimulus types." These clarifications, which do not alter the original results or conclusions, address discrepancies in the Methods section. Specifically, the pupil-position MLP in CvT-LSTM models uses a 16-dimensional hidden representation, not 8-dimensional. The architecture is a four-head ensemble, with modulation, core, and readout modules independently parameterized. The modulation network in CvT-LSTM models uses only treadmill velocity and pupil radius, not their derivative. LSTM hidden and cell states are 6-dimensional in Conv-LSTM and 16-dimensional in CvT-LSTM. The feedforward component uses ELU in Conv-LSTM and GELU in CvT-LSTM. Some Conv-LSTM variants also receive explicit spatial information, and typographical errors in core module equations have been corrected.

Key takeaway

For AI Scientists and Research Scientists working with neural activity models, carefully review published architectural details against actual implementations. Your understanding of model behavior and ability to reproduce results hinges on precise documentation of components like hidden layer dimensions, ensemble configurations, and specific activation functions. Always cross-reference method descriptions with supplementary information or code when available to avoid misinterpretations.

Key insights

Accurate documentation of model architectures is critical for reproducibility in neural activity prediction.

Principles

• Model documentation must precisely reflect implementation.
• Architectural details impact model interpretation.

Method

The corrected model uses a four-head ensemble, averaging standardized log-responses across heads for predictions, with shared perspective transform and readout grid.

In practice

• Verify hidden representation dimensions.
• Confirm activation functions per model variant.
• Document all behavioral inputs used.

Topics

• Foundation Model of Neural Activity
• Conv-LSTM Architecture
• CvT-LSTM Architecture
• Neural Network Ensembles
• Model Architectural Details

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Emerging Flexible Designs for Geospatial Multimodal Foundation Models — Apples-to-apples comparison of geospatial foundation model architectures reveals design trade-offs in flexibility, alignment, and performance.
• Position: Don't Just "Fix it in Post": A Science of AI Must Study Training Dynamics — A scientific understanding of AI requires studying training dynamics, not just post-training analysis, to predict and control model behaviors.
• Cross-Architecture Model Diffing with Crosscoders: Unsupervised Discovery of Differences Between LLMs — Cross-architecture model diffing with DFCs can unsupervisedly uncover specific behavioral alignments in LLMs.
• Pattern Selectivity is Not Task-Causal Structure: A Cross-Architecture Mechanistic Study of Composed-Task Circuits in 1B-Class Language Models — Specific mechanistic circuits for composed tasks do not transfer across 1B-class language models.
• Quick Paper Review: "There Will Be a Scientific Theory of Deep Learning" — Learning mechanics, a theory of deep learning training dynamics, is proposed as an emerging scientific framework.
• Why Your AI Assistant Confidently Lies — And Why It’s Not the Data’s Fault — LLM hallucinations stem from three architectural mechanisms, not just bad training data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.