@maryamshanechi
Sawchuk Chair & Prof at USC Viterbi School of Engineering | Founding Director of USC Center for Neurotech | Developing AI/ML methods & neurotech to decode the brain & treat its conditions 🧠🤖💻 https://nseip.usc.edu/
🚀 Overall, MRINE can advance #BCI by:
🔸Nonlinearly fusing neural modalities
🔸Enabling real-time decoding
🔸Handling varied timescales & distributions
📍 Poster Today: Hall C,D,E #2100 | Thurs 4 Dec | 4:30-7:30 PM @neuripsconf.bsky.social
📃 openreview.net/forum?id=jOH...
🖥️ github.com/ShanechiLab/...
On two public spike-LFP motor cortex datasets during reaching, MRINE:
✅ Improves real-time behavior decoding via multimodal fusion
✅ Outperforms recent multimodal neural models
It also:
✅ Generalizes to a high-dim dataset with Neuropixels spikes and calcium imaging
MRINE tackles these challenges with a multiscale encoder that:
🔸 Models each modality at its own timescale & forward-predicts its dynamics to infer fast latent factors
🔸 Nonlinearly fuses modality-specific factors
🔸 Enables real-time inference via its linear state-space model (SSM) backbone
Brain activity is recorded through modalities like spikes and LFPs. Fusing them can unlock richer neural representations & improve BCI performance and robustness.
But fusion is hard: modalities differ in timescales & distributions, and BCIs require real-time inference.
Can we achieve nonlinear multimodal neural fusion while also enabling real-time recursive decoding for #BCI?
In our third paper at #NeurIPS2025, we present MRINE, which does exactly that — improving decoding even for modalities w/ distinct timescales & distributions.
👏 Eray Erturk
🧵 Paper Code ⬇️
🚀BaRISTA shows how to flexibly encode spatial information toward future foundation models of intracranial brain activity.
👏L Oganesian, S Hashemi
📍Poster: Wed Dec 3 4:30-7:30pm, Exhibit Hall C,D,E 2011 @neuripsconf.bsky.social
Paper: openreview.net/forum?id=LDj...
Code: github.com/ShanechiLab/...
We also show that BaRISTA:
✅ Scales with increased pretraining data
✅ Generalizes to held-out subjects
✅ Can use spatial scales larger than channel level without sacrificing channel reconstruction performance
On a public iEEG dataset (Brain Treebank):
✅ BaRISTA shows that spatial encoding at scales larger than individual channels improves downstream decoding of auditory or visual features
✅ BaRISTA outperforms state-of-the-art iEEG models given its flexible spatial encoding
To address how to encode spatial information, BaRISTA:
🔹 Flexibly incorporates multiple spatial scales in a transformer model trained with masked latent reconstruction
🔹 Separates the choice of spatial encoding scale from the masking scale to isolate their effects
At what scale should spatial information be encoded toward future foundation models of intracranial brain activity?
In our second paper at #NeurIPS2025, we present BaRISTA ☕ — a self-supervised multi-subject model that enables flexible spatial encoding & boosts downstream decoding.
🧵 Paper Code ⬇️
🚀Overall, our cross-modal distillation can enable high-accuracy and stable LFP models for #BCIs.
👏Eray Erturk, Saba Hashemi
See Eray at #NeurIPS2025
📍Poster Session 1, Hall C,D,E #2115 | Wed 3 Dec | 11AM - 2PM @neuripsconf.bsky.social
📃 openreview.net/forum?id=hT7...
🖥️ github.com/ShanechiLab/...
Our framework achieves:
✅ Substantial decoding performance improvement over several LFP-only baselines
✅ Consistent improvements in unsupervised, supervised & multi-session distillation setups
✅ Generalization to unseen sessions without additional distillation
✅ Spike-aligned LFP latent structure
Our framework:
1️⃣ Pretrains a multi-session spike model
2️⃣ Fine-tunes the multi-session spike model on new spike signals
3️⃣ Trains the Distilled LFP model via cross-modal representation alignment
🔥 This produces spike-informed LFP models with significantly improved decoding.
Why this matters:
🧠 LFPs are used in many #BCIs and have high stability, but often underperform for behavior decoding compared with spikes.
❓ We ask: Can we transfer representational knowledge from spike models → LFP models?
✅ Answer: Yes — and the gains are substantial!
🎉 New in #NeurIPS2025, we present “Cross-Modal Representational Knowledge Distillation for Enhanced Spike-Informed LFP Modeling”.
We show that high-fidelity spike transformer models can teach LFP models to substantially enhance LFP decoding. #BCI
👏 Eray Erturk
🧵 Paper, Code ⬇️
Excited for SBIND to support neural image modalities, thus expanding our prior neural-behavioral models:
PSID & DPAD (Nat Neuro 2021 & 2024), IPSID (PNAS 2024), PGLDM (NeurIPS 2024), BRAID (ICLR 2025)
📜 Paper: openreview.net/pdf?id=k4KVh...
💻Code: github.com/shanechiLab/...
Also on public data (🙏to Churchland, Andersen, and Shapiro labs)
✅ Self-attention improves neural-behavior predictions by learning long-range patterns while convolutions learn local ones
✅ Two-stage learning improves behavior prediction by disentangling behaviorally relevant dynamics
On public widefield calcium (Churchland lab) and functional ultrasound (Andersen and Shapiro labs) neural imaging data, SBIND outperforms other neural-behavioral models in decoding continuous and categorical behaviors in visual decision-making and memory-guided saccade tasks.
SBIND:
✅ Operates directly on raw images & avoids preprocessing.
✅ Combines self-attention and convolutional layers to model both global and local patterns.
✅ Uses two-stage learning of convolutional RNNs (ConvRNNs) to disentangle behaviorally relevant and other neural dynamics.
🎉 New in #ICML2025, we develop SBIND for modeling neural imaging modalities and disentangling their behaviorally relevant dynamics.
SBIND learns local and global spatiotemporal patterns in raw widefield calcium and functional ultrasound neural images.
👏M Hoseini
🧵Paper, Code⬇️
Excited for BRAID to expand our neural-behavioral models: PSID & DPAD (Nat Neurosci 2020 & 2024), PGLDM (NeurIPS 2024), IPSID (PNAS 2024)!
See Parsa Vahidi at #ICLR2025!
📍Poster Session 5, Hall 3+Hall 2B #57 | Sat 4/26 | 10AM - 12:30PM
📜 openreview.net/forum?id=3us...
💻 github.com/ShanechiLab/...
On public motor cortex data during reaching from the Sabes lab, BRAID outperformed several baselines in neural-behavioral predictions by capturing nonlinearity, modeling sensory task instructions as input, and disentangling intrinsic behaviorally relevant neural dynamics.
In nonlinear simulations, BRAID accurately disentangled intrinsic neural-behavioral dynamics from input dynamics. In terms of learning the intrinsic dynamics and decoding behavior, BRAID outperformed prior neural-behavioral models, which either don’t include input or are linear.
BRAID
✅ Disentangles intrinsic behaviorally relevant neural dynamics from input, neural-specific & behavior-specific dynamics
✅ Captures nonlinearity
It is a multi-stage RNN: each stage learns a subtype of dynamics & combines a predictor network w/ a generative network to learn intrinsic dynamics.
🎉 New in #ICLR2025, we present BRAID for input-driven nonlinear dynamical modeling of neural-behavioral data.
BRAID disentangles the intrinsic dynamics shared between modalities from input dynamics and modality-specific dynamics.
👏 Parsa Vahidi & Omid Sani
@iclr-conf.bsky.social
🧵, Paper & Code ⬇️
You can see our poster at #ICLR2025!
📍 Poster Session 1, Hall 3 + Hall 2B, #68 | Thu, Apr 24 | 10 AM - 12:30 PM
Poster: iclr.cc/virtual/2025...
📜 💻Paper and code: openreview.net/pdf?id=mkDam...
On public neural data from the mice head direction circuit from Buzsáki lab, PGPCA outperforms baselines across all state dimensions. Also, interestingly, the geometric coordinate outperforms the Euclidean one, showing that the noise around the manifold also follows the same geometry.
In simulations, PGPCA recovers the true data distribution and distinguishes between different coordinates (geometric vs. Euclidean) regardless of the manifold state distribution p(z). Also, PGPCA outperforms Probabilistic PCA (PPCA) in modeling data around a nonlinear manifold.
PGPCA decomposes the data distribution p(y) into a state distribution on a nonlinear manifold p(z) plus a deviation from the manifold captured by the distribution coordinate K(z). K(z) can be Euclidean or geometric, as we derive. A new algorithm learns the model parameters.
New in our #ICLR2025 spotlight ✨, we introduce PGPCA, a Probabilistic Geometric method that enables modeling and dimensionality reduction for data distributed around nonlinear manifolds. We also show PGPCA’s application to brain data.
👏 Han-Lin Hsieh
@iclr-conf.bsky.social
Paper, Code, 🧵⬇️