Mamba Spike Behavioral Decoding

---
name: mamba-spike-behavioral-decoding
description: >
  Mamba forecaster methodology for implicit behavioral decoding from next-step
  spike forecasts at population scale. A single sequence model trained only on
  next-step Poisson rate prediction produces predicted firing rates that decode
  animal behavior better than raw spike counts under matched temporal context.
  Enables closed-loop BCI without separate behavioral decoding networks.
category: neuroscience
tags: [spike-forecasting, mamba, bci, neural-population, neuromorphic, behavioral-decoding, state-space-model]
related_skills:
  - autoregressive-flow-matching-neural-dynamics
  - neural-population-dynamics
  - eeg-ieeg-bridge-bci
  - mind2drive-eeg-driver-intention
activation_keywords:
  - mamba forecaster
  - spike forecast behavioral decoding
  - implicit behavioral decoding
  - neural population rate prediction
  - closed-loop bci decoding
  - poission rate forecasting bci
  - next-step spike prediction
  - neuropsychixels behavioral readout
---

# Mamba Spike-Based Behavioral Decoding

**Paper**: *Implicit Behavioral Decoding from Next-Step Spike Forecasts at Population Scale*
**Authors**: John R. Minnick, Jesus Gonzalez-Ferrer, Kamran Hussain, Jinghui Geng, Ash Robbins, Mohammed A. Mostajo-Radji, David Haussler, Jason Eshraghian, Mircea Teodorescu
**Institution**: UC Santa Cruz (ECE, Genomics Institute, Biomolecular Engineering)
**arXiv**: 2605.12999 (May 13, 2026)
**Category**: q-bio.NC, cs.LG

## Overview

Closed-loop brain-computer interfaces (BCIs) require two capabilities: forecasting neural population dynamics ahead of time, and decoding the subject's behavioral state. This paper demonstrates that a single Mamba sequence model, trained only to predict next-step population firing rates (Poisson NLL), implicitly encodes behavioral information in its predicted rates -- eliminating the need for a separate behavioral decoding network.

## Core Methodology

### Problem Formulation

Given a sliding window of H population spike-count vectors, predict the next time bin:

```
X_t = {x(t-H+1), ..., x(t)} -> lambda_hat(t+1) in R^M
```

- **M**: number of simultaneously recorded neurons (up to 1,998)
- **delta_t**: 50 ms bin width
- **H = 10** bins (500 ms context)
- Model outputs predicted firing rates via softplus activation
- Trained with **Poisson negative log-likelihood (NLL)** on spike counts only
- **Behavioral labels are NEVER used during training**

### Architecture

**Mamba (Selective State-Space Model)**:
- Selective SSM with causal recurrence and content-aware gating
- 1.95M parameters
- Multi-session training with per-sample channel masking (padded to M_max=1,998)
- Well-matched to spike-count autoregression on long session-length sequences

**Architecture Controls Tested**:
- Transformer (causal self-attention, 2.22M params)
- LRU (linear recurrent unit, 1.23M params)
- NDT2-style bidirectional masked-attention variant

### Key Discovery: Implicit Behavioral Decoding

The predicted firing rates from the forecaster serve as behaviorally informative compression of the recent population state. A lightweight per-session linear readout over these predicted rates:

1. **Decodes behavior WITHOUT any behavioral training** in the forecaster
2. **Outperforms matched-context raw spike count baselines** by +4-6 pp
3. Works across Transformer, LRU, and NDT2 architectures (not Mamba-specific)

**Mechanism**: The Poisson NLL training objective forces the model to integrate 500ms of population history into a smoothed rate estimate. Single-bin raw counts at 50ms are dominated by Poisson noise; matched-context summed raw counts ignore population structure. Mamba's predicted rate lies at the intersection -- same temporal context, plus cross-neuron dependency exploitation.

## Results

### Dataset
- **Steinmetz visual-discrimination benchmark**: 39 sessions, ~27,000 neurons, 1,994 held-out trials
- **IBL Repeated Site**: 66 sessions, ~63K neurons (cross-laboratory extension)
- 105 sessions total, 89,768 real channels across 42 Allen CCF regions

### Behavioral Decoding Performance

| Decoder | Response 3-class (trial vote) | Stim Side 3-class (trial vote) |
|---------|-------------------------------|-------------------------------|
| Linear / 1-bin raw | 72.1% | 49.8% |
| Linear / H=10 sum | 69.6% | 60.5% |
| Ridge / H=10 flat | 71.3% | 61.5% |
| **Mamba (3-seed)** | **75.7% ± 0.2** | **66.1% ± 0.6** |
| Chance | 33.3% | 33.3% |

- Mamba decodes mouse choice at **2.3x chance level**
- Mamba wins at trial vote by **+4-6 pp** on both response and stimulus side
- Architecture controls (Transformer, LRU) reproduce gains within ~1-3 pp

### Cross-Neuron Coupling Verification
Population shuffle test: shuffling each neuron's time series independently drops mean per-neuron r by **48.4%** (median 50.7%), with 38/39 sessions showing >25% degradation. Direct evidence that Mamba exploits cross-neuron temporal coupling, not single-neuron autocorrelation.

### Forecasting Accuracy
- Per-neuron Pearson r = 0.176 (modest due to Poisson noise at 50ms bins)
- Population-rate r = 0.783
- Population-cosine = 0.648
- DTW alignment reduces average error by 41% vs naive bin-to-bin

### Calibration Budget
- **100-150 trials** (~5-8 minutes) of session-start calibration brings per-session readout within 1-2 pp of asymptotic accuracy
- Response decoding asymptotes at ~120 trials; stimulus side at ~140

### Deployment Latency
- Single 50ms-bin Mamba forward pass: ≤6.4 ms per batch of 512 windows
- RTX 5000 Ada, M=1,240 neurons, 152 MB peak VRAM
- Per-session linear head: sub-millisecond
- Total: well within 50 ms bin budget

## Implementation Patterns

### Pipeline Architecture
```
Spike-sorted Neuropixels (M neurons, 50ms bins)
  -> Window into X_t in R^(MxH) (H=10, 500ms context)
    -> Mamba SSM: h_t = A_bar * h_{t-1} + B_bar * x_t, y_t = C * h_t
      -> Predicted next-bin firing rates lambda_hat_{t+1} in R^M
        -> Per-session multinomial LR: y_hat = softmax(W * lambda_hat + b)
          -> Decoded: mouse response / stimulus side (3-class each)
```

### Training Configuration
- Poisson NLL loss on spike counts only
- AdamW with cosine LR + warmup, 50 epochs
- Per-session temporal 70/15/15% train/val/test split
- 20% trial-level holdout for behavioral evaluation

### Matched-Context Baselines
Three non-forecasting reference baselines for fair comparison:
1. **Linear / 1-bin**: Single 50ms count vector
2. **Linear / H=10 sum**: Sum past 10 bins, same M-dim features
3. **Ridge / H=10 flat**: Flatten past 10 bins (M*H ≈ 12,000 features), more capacity

## Deployment Considerations

### Closed-Loop BCI Pipeline
1. **Fixed forecaster**: Single Mamba checkpoint, trained once on combined substrate
2. **Per-session calibration**: 100-150 trials to fit linear readout
3. **Online operation**: 500ms history -> rate prediction -> class probabilities per bin
4. **Hardware**: Workstation-class external GPU (on-implant not claimed)

### Limitations
1. H=1 horizon; greedy rollout regresses to session mean over ~3-5 bins
2. 50ms bins miss <20ms dynamics relevant to some motor BCIs
3. Cross-laboratory extension is partial (stimulus side gain survives, response/contrast don't on IBL)
4. Per-session linear head does not transfer across sessions
5. Modest per-neuron r=0.18 at 50ms bins is the binding constraint

## Applications

1. **Closed-loop BCI**: Single model replaces separate forecasting + decoding networks
2. **Neural population analysis**: Rate predictions as behaviorally-informed latent space
3. **Neuromorphic computing**: Efficient SSM-based spike processing at Neuropixels scale
4. **Neuroscience research**: Cross-neuron coupling as computational substrate

## Comparison to Related Work

| Method | Behavioral Labels | Spike Prediction | Matched-Context Baseline |
|--------|-------------------|------------------|--------------------------|
| NEDS (Zhang 2025) | Yes (joint) | Yes (Poisson) | No |
| LFADS (Pandarinath 2018) | No | Smoothed rates | No |
| NDT2 (Ye 2023) | Yes (masked) | No | No |
| CEBRA (Schneider 2023) | Yes | No | No |
| **This Work** | **No** | **Yes (Poisson)** | **Yes (3 baselines)** |

## Key Equations

**Mamba SSM recurrence**:
```
h_t = A_bar * h_{t-1} + B_bar * x_t
y_t = C * h_t
lambda_hat = softplus(y_t)
Loss = -sum(x * log(lambda_hat) - lambda_hat)  # Poisson NLL
```

**Biological decoding**:
```
y_hat = softmax(W * lambda_hat + b)  # Per-session multinomial LR
```

## Pitfalls

1. **Matched-context is critical**: Comparing against baselines with same temporal context window is essential to isolate the forecasting contribution from raw temporal integration
2. **Per-neuron r is misleading**: Single-neuron Pearson r at 50ms bins underestimates fidelity due to Poisson noise; population-level metrics (cosine similarity, DTW) are more informative
3. **Cross-laboratory transfer**: Gains may not fully transfer across different task structures (binary forced-choice vs. no-go designs)
4. **Calibration is per-session**: Linear readout must be re-fit for each recording session

## References

- Gu & Dao (2023): Mamba selective state-space model
- Steinmetz et al. (2019): Visual discrimination dataset
- IBL et al. (2025): Repeated Site release
- Pandarinath et al. (2018): LFADS
- Ye & Pandarinath (2021): NDT
- Linderman et al. (2017): rSLDS
- Schneider et al. (2023): CEBRA