Implicit Behavioral Decoding from Next-Step Spike Forecasts at Population Scale

Abstract: Closed-loop brain-computer interfaces often require both a forecast of upcoming neural population activity and a readout of the animal's behavioral state. A single Mamba forecaster, trained only on next-step spike counts at Neuropixels scale, can deliver both in one forward pass. A lightweight per-session linear head reading the model's predicted rates decodes behavior better than the same linear classifier reading the raw spike counts, under matched temporal context. We test on the Steinmetz visual-discrimination benchmark, which spans 39 sessions, roughly 27,000 neurons, and 1,994 held-out trials. Across three training seeds, Mamba's predicted rates decode mouse choice at 75.7$\pm$0.2% trial vote, roughly 2.3 times chance level, and stimulus side at 66.1$\pm$0.6%, about twice chance. Compared to a matched 500 ms-context linear decoder on the raw spike counts, Mamba wins at trial vote by 4-6 pp on response and 4-6 pp on stimulus side. A session-start calibration block of about 100-150 trials brings the readout within 1-2 pp of asymptote, and the full pipeline fits inside the 50 ms bin budget on workstation-class GPUs typical of tethered chronic Neuropixels recordings.

• The paper demonstrates that a single Mamba forecaster achieves simultaneous real-time spike prediction and implicit behavioral decoding, outperforming matched-context linear decoders.
• The model utilizes a selective state-space approach with causal recurrence and content-aware gating, validated on large-scale Neuropixels datasets.
• The findings imply reduced computational complexity for closed-loop BCIs through minimal per-session calibration and effective exploitation of cross-neuron temporal dynamics.

Implicit Behavioral Decoding from Next-Step Spike Forecasts: An Authoritative Review

Problem Statement and Motivation

The paper investigates whether a single neural population forecaster can simultaneously provide both real-time spike-rate predictions and implicit behavioral decoding at Neuropixels scale. Instead of employing separate models for neural activity forecasting and behavioral readout—a paradigm that imposes considerable computational and deployment costs—the authors propose utilizing the Mamba selective state-space model, trained as a next-step Poisson spike-count forecaster, to encode underlying population dynamics that support downstream behavioral decoding in closed-loop BCI scenarios.

The motivation lies in addressing two central challenges for BCIs: i) anticipating transitions in neurodynamical states and ii) decoding behavioral intent with high accuracy and low latency, using high-dimensional, temporally resolved electrophysiological data.

Methodology

Forecasting Pipeline

• Input Representation: Neuropixels probe recordings are indexed as history matrices spanning 500 ms windows (10 bins, 50 ms width) from up to 1,998 neurons, forming $X_t \in \mathbb{R}^{M \times H}$.
• Model Architecture: The Mamba forecaster, a selective SSM with causal recurrence and content-aware gating, is trained to minimize next-step Poisson NLL, producing continuous-valued firing rate predictions using a softplus activation.
• Behavioral Decode Head: A per-session multinomial linear classifier is post-hoc fitted to the forecaster's predicted rates $\hat{\lambda}_{t+1}$, decoding response choice (3-class), stimulus side (3-class), and stimulus contrast (16-class) from held-out trial bins.

Baseline and Control Comparisons

Three strong linear baselines, all matched for temporal context and raw spike-count input, are systematically compared:

• Single-bin raw counts.
• Summed 10-bin window of raw counts.
• Flattened 10-bin raw counts with high-dimensional ridge classification.

Additionally, the approach is benchmarked against alternative architectures (Transformer, LRU, NDT2-style bidirectional masked attention) under identical training and evaluation protocols.

Experimental Substrates

Behavioral decoding is evaluated on the Steinmetz Visual Discrimination dataset (39 sessions, $\sim$27,000 neurons) and cross-laboratory generalization is probed using the IBL Repeated Site dataset (66 sessions, $\sim$63,000 neurons).

Key Numerical Results and Claims

• Behavioral Decoding Performance: Mamba's predicted rates enable post-hoc decoding of mouse response at $75.7 \pm 0.2\%$ (trial vote), exceeding chance by 2.3x and outperforming matched-context linear decoders by +4–6 percentage points (pp). Stimulus side decoding attains $66.1 \pm 0.6\%$, also +4–6 pp above baseline. The performance generalizes to Transformer and LRU variants within $\sim$1–3 pp.
• Spike Forecast Fidelity: Population-level rate prediction achieves Pearson $r=0.783$, cosine $0.648$, while per-neuron predictions are modest ($r=0.18$) due to inherent Poisson noise dominance at 50 ms bins.
• Calibration Efficiency: Session calibration with 100–150 trials suffices to bring the linear behavioral decode head within 1–2 pp of asymptotic accuracy, supporting deployment feasibility.
• Computational Feasibility: End-to-end forecasting and decoding runs comfortably within a 50 ms bin budget per batch on workstation-class GPUs (e.g., RTX 5000), meeting closed-loop BCI real-time constraints.

Mechanistic Insights

• Cross-Neuron Coupling: The population shuffle test reveals a $\hat{\lambda}_{t+1}$0 mean drop in per-neuron $\hat{\lambda}_{t+1}$1 when cross-neuron temporal correlations are destroyed, indicating that cross-neuron coupling is a load-bearing property in both forecast accuracy and behavioral informativeness.
• Population-Aware Temporal Integration: By integrating temporal context and exploiting cross-neuron structure, Mamba's predicted rates encode richer behavioral information than context-only summing or short-window decoders.

Limitations

• Behavioral decoding remains a post-hoc process; the forecaster operates in a single-step-ahead regime (not multi-horizon rollout).
• Per-neuron prediction accuracy is bounded by Poisson noise floor, limiting single-neuron fidelity.
• Readout calibration is session-specific, with no demonstrated cross-session transfer.
• Cross-laboratory extension is partial—matched-context performance gains persist primarily for spatially distributed targets (stimulus side), while response and contrast accuracy gains are absorbed under certain task structures (e.g., IBL’s binary forced-choice paradigm).
• No comparison against official NDT2 or CEBRA pipelines is presented.

Practical and Theoretical Implications

Practical Implications

• Deployment in Closed-Loop BCIs: The ability to simultaneously decode behavior and forecast neural activity using a single model substantially reduces compute burden and system complexity, critical for tethered chronic Neuropixels recordings and future implanted neuroprosthetics.
• Calibration Protocol Design: Minimal per-session calibration requirements (<150 trials) facilitate efficient operationalization in diverse experimental contexts.

Theoretical Implications

• Compression of Behavioral Signal: The empirical demonstration that predicted rates afford superior behavioral decodability, even absent behavior labels during forecaster training, underscores the emergent property of population-aware temporal integration under autoregressive training objectives.
• Role of Cross-Neuron Dynamics: The shuffle and DTW analyses concretely indicate that cross-neuron temporal structure—rather than per-neuron autocorrelation—underpins both forecasting fidelity and behavioral decode signal.

Future Directions

• Extending to multi-horizon forecasting and decoding scenarios.
• Exploring session-generalized readouts with adaptive input projections, low-rank embedding adapters, or session-specific normalization.
• Benchmarking against additional architectures (e.g., official NDT2, CEBRA).
• Closed-loop experimental validation with live animals.
• Addressing task-structure-dependent erosion of decode gains in multi-laboratory datasets.

Conclusion

The paper establishes that a single Mamba forecaster trained solely for next-step population spike prediction enables implicit behavioral decoding via its predicted rate outputs, achieving superior accuracy compared to strong matched-context linear baselines in major BCI-relevant targets (response, stimulus side). The approach unifies population-rate forecasting and behavioral readout within real-time and practitioner-centric deployment constraints, leveraging population-aware temporal dynamics as a behavioral signal compressor. The findings delineate a robust mechanism relying on cross-neuron temporal coupling and provide concrete recommendations for future BCI model architecture and pipeline design (2605.12999).