spike-forecast-behavioral-decoding

name

spike-forecast-behavioral-decoding

description

Implicit behavioral decoding from next-step spike forecasts at population scale. Methodology from paper 'Implicit Behavioral Decoding from Next-Step Spike Forecasts at Population Scale' (arXiv: 2605.12999). Demonstrates that a single Mamba forecaster, trained only on next-step spike counts at Neuropixels scale, implicitly learns behavioral representations without behavioral labels. Use when: building closed-loop BCIs, neural population forecasting, spike train prediction, behavioral decoding from neural activity, self-supervised neural representation learning, Mamba/state-space models for neuroscience, Neuropixels data analysis. Activation: spike forecast, behavioral decoding, Mamba neural, implicit neural representation, closed-loop BCI, spike prediction, neural forecasting, Neuropixels decoding.

Spike-Forecast Behavioral Decoding

Methodology from: Implicit Behavioral Decoding from Next-Step Spike Forecasts at Population Scale (Minnick et al., arXiv:2605.12999, May 2026).

Core Insight

A single autoregressive forecaster (Mamba) trained only on next-step spike counts at Neuropixels scale implicitly learns behavioral representations in its hidden states — without any behavioral labels during training. A lightweight per-session linear probe on hidden states recovers behavioral variables with accuracy comparable to dedicated supervised decoders.

Key Findings

1. Unified Forecasting + Decoding: One model delivers both neural population forecasts and behavioral readouts in a single forward pass, eliminating the need for separate forecasting and decoding models.
2. Emergent Behavioral Encoding: Next-step prediction inherently captures task-relevant latent structure — the hidden states organize by behavioral variables even though the model was never trained to predict behavior.
3. Compute Efficiency: Reduces compute requirements for closed-loop BCIs by replacing two separate models (forecaster + decoder) with one.
4. No Behavioral Labels Needed: The forecaster is self-supervised — only spike counts are required for training. Behavioral decoding emerges via post-hoc linear probes.
5. Neuropixels-Scale: Tested at scale with hundreds of simultaneously recorded neurons.

Architecture

Spike Counts (t-1, t-2, ...)  →  Mamba Forecaster  →  Hidden States (t)
                                                              │
                                                    ┌─────────┴─────────┐
                                                    ↓                   ↓
                                            Next-step Spike       Behavioral
                                              Forecast           Decoding
                                              (supervised)      (linear probe)

• Input: Next-step spike counts from Neuropixels recordings
• Model: Mamba (state-space model) — efficient autoregressive sequence modeling
• Output 1: Predicted spike counts at t+1 (training objective)
• Output 2: Hidden states that encode behavioral variables (emergent property)
• Decoder: Per-session linear probe on hidden states → behavioral variables

Methodology Workflow

Step 1: Train Self-Supervised Forecaster

# Train Mamba on next-step spike prediction only
forecaster = MambaSpikeForecaster(
    n_neurons=spike_matrix.shape[1],
    d_model=256,
    n_layers=4
)

# Loss: predict spike counts at t+1 from t, t-1, ...
loss = mse_loss(forecaster(spike_train[:-1]), spike_train[1:])

Step 2: Extract Hidden States

# Forward pass through trained forecaster
hidden_states = forecaster.encode(spike_train)
# hidden_states: (T, batch, d_model)

Step 3: Linear Probe for Behavioral Decoding

# Per-session linear probe (lightweight, session-specific)
probe = Ridge(alpha=1.0)
probe.fit(hidden_states[training_idx], behavior[training_idx])
predicted_behavior = probe.predict(hidden_states[test_idx])

Step 4: Evaluation

• Compare decoding accuracy against dedicated supervised baselines
• Verify accuracy parity: probe performance ≈ supervised decoder performance
• Analyze what behavioral variables are encoded in different hidden state dimensions

Applications

• Closed-loop BCIs: Unified forecasting + decoding reduces latency and compute
• Neural representation analysis: Study what task-relevant features emerge from self-supervised spike prediction
• Multi-area recordings: Apply to any neural population data (cortex, hippocampus, etc.)
• Zero-shot behavioral decoding: Decode behavior in new sessions without retraining the forecaster

Evaluation Metrics

Metric	Description
Decoding R²	Variance explained of behavioral variable
Forecast MSE	Spike prediction accuracy
Latency	Single forward pass vs. dual-model pipeline
Compute	Parameters and FLOPs comparison

Pitfalls

• Per-session probes: Linear probes are session-specific — may need recalibration across sessions
• Linearity assumption: Behavioral encoding might not be fully linear; more complex probes could improve accuracy
• Temporal alignment: Ensure precise temporal alignment between spike counts and behavioral measurements
• Neuropixels specific: Architecture assumes high-channel-count recordings; may need adaptation for lower-density arrays

Related Skills

• spikeprophecy-benchmark — Benchmark suite for evaluating spike forecasters
• mamba-spike-forecaster-bci — Mamba forecaster for closed-loop BCI
• implicit-behavioral-decoding-spike-forecasts — Companion paper on evaluation protocols

Paper Reference

Minnick, J.R., Gonzalez-Ferrer, J., Hussain, K., et al. (2026). Implicit Behavioral Decoding from Next-Step Spike Forecasts at Population Scale. arXiv:2605.12999.