TL;DR
SPIRE is a deep learning framework that disentangles shared and region-specific neural dynamics, improving analysis of multi-region brain data and revealing how external stimuli reorganize neural activity.
Contribution
Introduces SPIRE, a novel autoencoder model that effectively separates shared and private neural signals, outperforming classical models especially under nonlinear distortions.
Findings
SPIRE accurately recovers cross-regional neural structure.
Shared latents encode stimulation-specific signatures.
SPIRE generalizes across sites and frequencies.
Abstract
Disentangling shared network-level dynamics from region-specific activity is a central challenge in modeling multi-region neural data. We introduce SPIRE (Shared-Private Inter-Regional Encoder), a deep multi-encoder autoencoder that factorizes recordings into shared and private latent subspaces with novel alignment and disentanglement losses. Trained solely on baseline data, SPIRE robustly recovers cross-regional structure and reveals how external perturbations reorganize it. On synthetic benchmarks with ground-truth latents, SPIRE outperforms classical probabilistic models under nonlinear distortions and temporal misalignments. Applied to intracranial deep brain stimulation (DBS) recordings, SPIRE shows that shared latents reliably encode stimulation-specific signatures that generalize across sites and frequencies. These results establish SPIRE as a practical, reproducible tool for…
Peer Reviews
Decision·Submitted to ICLR 2026
* The model is intuitive and well presented. * Experiments show that the proposed model outperforms the baseline DLAG.
* The model seems intuitive, but a little bit simple. The appreciated part to me is the alignment for the shared latent. * The comparisons are not comprehensive, since I believe the discovered latent has no interpretability. There is no ground truth for the real-world data. Therefore, it is hard to validate whether such a VAE model is useful in neuroscience, since we care about what dose the latent mean and how it is useful for understanding the neural data. I think authors can at least use a li
- The paper presents a nice and elegant method with fairly straightforward components (RNNs, linear projections onto latents, etc.) and cleverly designed losses. - The manuscript is clearly and concisely written imo, and in general of high quality. - I commend the use of simulated data with groundtruth, and interesting real data applications.
- While the framework is nice, I’m not convinced of its value, both in terms of methodological or scientific contribution. In other words, the primary contribution is that it presents a novel nonlinear and disentangled latent analysis method, applied to human intracranial data. But it’s quality / accuracy and scientific insight remains limited, as a standalone contribution and relative to existing works. - The majority of experimental results are essentially to validate the retrieval of shared v
The method appears to be different from existing methods that I'm aware of and performs well on the simluated data benchmarks, although I have misgivings about their simulated data analysis, which I outline under "weaknesses". The real-data validation also appears to be interesting and to be consistent with hypotheses from neurophysiology. The paper is also fairly well written and well organized.
The biggest weakness of this paper is that it did not sufficiently engage with the existing literature. Specifically, it is not entirely clear what this paper adds to the existing literature as it did not benchmark against some important contributions. Some examples are included below. While the papers I've cited are described as "multi-model" and explicitly model a combination of neural and behavioral data I see no practical limitation to using a second neural dataset as the "behavior". The aut
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
