Eva Dyer | What Will Happen Next? Predicting the Brain's Future @ Vision Weekend Puerto Rico 2026
Why It Matters
Accurate neural forecasting could transform brain‑computer interfaces and enable foundation models that accelerate neuroscience research, with direct clinical and AI benefits.
Key Takeaways
- •Predicting neural activity is crucial for brain‑machine interfaces.
- •Neural forecasting faces data inconsistency and variable channel counts.
- •Transformer models struggle; wavelet multiscale representation improves forecasting performance.
- •SCRIER merges wavelets with cross‑attention, achieving state‑of‑the‑art accuracy.
- •Cross‑species forecasts enable foundation models for future neuroscience.
Summary
Eva Dyer opened her Vision Weekend Puerto Rico 2026 talk by framing neural forecasting as a linchpin for next‑generation brain‑machine interfaces, deep‑brain stimulation timing, and AI systems aligned with human values. She argued that, much like large language models predict future tokens, a pre‑training task that predicts upcoming neural activity could serve as a powerful foundation for neuroscience models.
Dyer highlighted two fundamental obstacles: the ever‑changing set of recorded neurons across sessions and experiments, and the predominance of univariate time‑series models that ignore the rich multichannel structure of neural data. To overcome these, her team built SCRIER, a hybrid architecture that first applies wavelet transforms to capture hierarchical, multiscale temporal patterns, then uses cross‑attention to fuse information across thousands of neuronal channels.
The speaker presented benchmark results on zebrafish, mouse, and sea‑urchin datasets, showing lower mean‑squared and mean‑absolute errors than prior state‑of‑the‑art methods. More importantly, SCRIER generated novel neural responses rather than merely extrapolating constant decay, a distinction she illustrated with side‑by‑side plots of predicted versus actual spiking activity. Audience questions underscored the potential of incorporating connectomics data to further refine forecasts.
If neural activity can be reliably projected forward, the implications span clinical neuromodulation, real‑time prosthetic control, and the creation of universal foundation models for brain science. Dyer’s work suggests that integrating classic signal‑processing techniques with modern attention mechanisms may be the key to unlocking these applications.
Comments
Want to join the conversation?
Loading comments...