Research map / Probabilistic and scientific modeling / Dynamical systems
A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Spikachu, a causal, scalable spiking-neural-network (SNN) transformer hybrid for multi-session, multi-subject neural decoding of NHP cursor velocity, with energy estimates from FLOPs.
Problem
Neural decoders for BCIs must be causal, energy-efficient, scalable across sessions and subjects, and generalize from few calibration examples, but existing models satisfy only some of these constraints.
Approach
Combines a shared-latent ANN harmonizer (Perceiver cross-attention over learnable unit embeddings) with multi-timescale spiking MLPs, spiking self-attention, and a membrane-potential observer readout, so most computation is spiking and causal, and multi-session pretraining transfers to new sessions, subjects, and tasks.
Key results
- Single-session Spikachu reaches R^2 0.8398 (CO) and 0.6761 (RT) at 5.14 uJ per inference, versus POYO at 0.8937/0.6785 and 2151.65 uJ (418.81x more energy).
- Spikachu-mp finetuned to individual sessions improves R^2 over scratch training (0.88 vs 0.84 CO, 0.69 vs 0.68 RT), with 1.91% and 2.01% energy savings.
- Transfer to new subject monkey T improves CO R^2 from 0.76 to 0.78 and converges about 3x faster than scratch training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Spikachu instead of GRU | Single-session neural decoding · CO | Energy (μJ): 2.26× less energy | Helped |
| Use Spikachu instead of POYO | Single-session neural decoding · CO and RT | Energy (μJ): 418.81× less | Helped |
| Train on 99 sessions instead of a single session | Spikachu · CO and RT | R²: 0.84 and 0.68 -> 0.80 and 0.57 | Mixed |
| Finetune pretrained Spikachu-mp on an individual session instead of training from scratch | Spikachu · CO | R²: ΔR²_CO = 3.75% | Helped |
| Finetune pretrained Spikachu-mp on an individual session instead of training from scratch | Spikachu · RT | R²: ΔR²_RT = 1.71% | Helped |
| Finetune from Spikachu-mp rather than initialize randomly | Spikachu · CO and RT | Energy per inference: saving 1.91% and 2.01% | Helped |
| Transfer pretrained Spikachu-mp to monkey T rather than train from scratch | Spikachu · Monkey T, CO and RT | R²: 0.76 -> 0.78; 0.66 -> 0.68 | Helped |
| Transfer pretrained Spikachu-mp rather than train from scratch | Spikachu · Monkey T, CO and RT | Energy per inference: reduction ... of 3.71% and 3.63% | Helped |
| Pretrain Spikachu-mp on more sessions | Spikachu-mp · Seen sessions, new sessions, and new subjects | R² | Helped |
| Pretrain Spikachu-mp on more sessions | Spikachu-mp · Seen sessions, new sessions, and new subjects | Energy per inference | Helped |
| Transfer pretrained weights rather than train from scratch | Spikachu-mp · MC-RTT and MC-Maze | R²: 0.57 -> 0.56; 0.79 -> 0.78 | No clear effect |
| Transfer pretrained weights rather than train from scratch | Spikachu-mp · MC-RTT and MC-Maze | Convergence speed: 2.33× speedup for the MC-RTT and 2.66× for MC-Maze | Helped |
Surprising findings
- Pretraining did not significantly improve transfer accuracy on MC-RTT/MC-Maze, though it accelerated convergence (2.33x and 2.66x).
- Spikachu is not the top decoder in accuracy: non-causal POYO scores higher (CO R^2 0.8937 vs 0.8398), so the advantage is energy efficiency.
Limitations
- Energy is estimated from FLOPs and standard per-operation energy constants rather than measured on neuromorphic hardware.
- The ANN harmonizer (cross-attention) consumes a disproportionate share of energy despite being under 15% of parameters.
- Performance gap remains with non-causal state-of-the-art POYO; no online or real-time closed-loop validation.
Experimental setup
Datasets: DANDI multi-session NHP motor cortex dataset (monkeys C, J, M, T) Neural Latents Benchmark MC-RTT Neural Latents Benchmark MC-Maze
Benchmarks: DANDI multi-session NHP motor cortex dataset MC-RTT MC-Maze
Metrics: R^2 energy per inference (microjoules, FLOP-based)
Baselines: MLP GRU LSTM POYO POYO-causal
Builds on: POYO Perceiver Spikformer SpikingJelly
Closest papers by meaning
- Leveraging unlabelled data for generalizable neural population decoding (NeurIPS 2026 · similarity 0.89)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.89)
- Towards a "Universal Translator" for Neural Dynamics at Single-Cell, Single-Spike Resolution (NeurIPS 2024 · similarity 0.89)
- SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding (NeurIPS 2025 · similarity 0.88)
- Causal pieces: analysing and improving spiking neural networks piece by piece (NeurIPS 2026 · similarity 0.88)
- SpikingSSMs: Learning Long Sequences with Sparse and Parallel Spiking State Space Models (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.