BiND: Bimanual Neural Discriminator-Decoder

• BiND is a two-stage intracortical decoding framework that first discriminates motion types before applying specialized GRU-based decoders.
• It leverages time-index augmentation to improve phase-awareness, achieving up to 4% cross-session prediction improvements over standard GRU models.
• The architecture supports real-time bilateral BCI control with low inference latency and efficient parameter usage, addressing overlapping motor-cortical activity.

BiND, short for Bimanual Neural Discriminator-Decoder, is a two-stage intracortical brain-computer interface decoder for predicting continuous bimanual hand trajectories from neural recordings. It addresses a central difficulty in bimanual decoding: motor-cortical activity for the left and right hands is partially overlapping, while bilateral actions introduce nonlinear interlimb interactions that degrade a single monolithic decoder. BiND responds to this by first discriminating motion type—unimanual left, unimanual right, or bimanual—and then routing each trial window to a specialized GRU-based decoder augmented with a trial-relative time index. On a publicly available 13-session intracortical dataset from a tetraplegic patient, the method attains mean $R^2$ of $0.76 \pm 0.01$ for unimanual and $0.69 \pm 0.03$ for bimanual trajectory prediction, surpassing the next-best GRU baseline by about $2\%$ in both settings and yielding up to $4\%$ gains in cross-session analyses (Robert et al., 19 Aug 2025).

1. Problem setting and decoding objective

BiND is designed for decoding continuous 2D hand velocities from intracortical recordings in order to support more natural two-hand control in BCIs (Robert et al., 19 Aug 2025). The target behaviors are unimanual left-hand cursor movement, unimanual right-hand cursor movement, and simultaneous bimanual cursor movement. The emphasis on bimanual control is not incidental: the underlying decoding problem is harder than unimanual movement because neural activity related to the two hands overlaps in motor cortex, especially during coordinated actions, and because interdependent bilateral dynamics complicate the mapping from population activity to ipsilateral and contralateral kinematics.

The paper frames BiND around two specific difficulties. First, overlapping neural representations make it difficult for a single decoder to disentangle which part of the observed firing pattern should be attributed to left-hand versus right-hand movement. Second, nonlinear interlimb interactions degrade tuning, particularly for the ipsilateral or non-dominant hand during coordinated motion. In the reported experiments, this manifests as a clear bimanual performance drop for the left hand: during bimanual movement, the left hand shows a $10$–$16\%$ reduction in $R^2$ and a $6$–$9\%$ reduction in correlation relative to unimanual movement, consistent with weaker ipsilateral tuning during bimanual coordination (Robert et al., 19 Aug 2025).

BiND’s central modeling decision is therefore task-aware routing. Rather than learning one entangled mapping from all neural windows to all velocity outputs, it first classifies the motion regime and then applies a decoder specialized to that regime. This suggests that the method treats motion context as a latent simplifier of the neural-to-kinematic map: once the regime is identified, the regression problem becomes more structured and easier to fit.

2. Dataset, neural features, and preprocessing pipeline

The model is evaluated on one tetraplegic participant from a publicly available dataset associated with Deo et al. (2024), comprising 13 open-loop BCI sessions (Robert et al., 19 Aug 2025). Neural activity was recorded from two 96-channel Utah arrays implanted in the hand area, or “hand knob,” of the left precentral gyrus, for a total of 192 channels. The decoding target is 2D cursor velocity for the left hand, right hand, or both hands jointly, depending on task condition.

Feature extraction begins with threshold-crossing spike detection at $0.76 \pm 0.01$0 RMS voltage. Spikes are then binned in 20-ms bins and Gaussian-temporally smoothed with kernel width approximately 40 ms, which the paper reports as empirically optimal for reducing noise while preserving rapid changes (Robert et al., 19 Aug 2025). Counts are normalized per session. No dimensionality reduction is applied.

For model input construction, each trial is segmented into overlapping 600-ms windows, corresponding to 30 time bins, with 300-ms stride and therefore 50% overlap. Each input window has feature shape $0.76 \pm 0.01$1, to which a trial-relative time index is appended at each time step. The time index $0.76 \pm 0.01$2 indicates the relative position of the window with respect to trial onset, or the “go” cue, and is concatenated to the neural feature at each time bin as

$0.76 \pm 0.01$3

This time-index augmentation is a distinctive aspect of BiND. It does not lengthen the window, but it injects coarse phase information, allowing the decoder to distinguish early-, middle-, and late-trial segments that may otherwise have similar short-term firing patterns. Empirically, the paper reports $0.76 \pm 0.01$4–$0.76 \pm 0.01$5 $0.76 \pm 0.01$6 gains depending on hand and motion type, with average improvement around $0.76 \pm 0.01$7 (Robert et al., 19 Aug 2025).

3. Two-stage architecture and mathematical formulation

BiND consists of a motion-type discriminator followed by specialized decoders (Robert et al., 19 Aug 2025). The discriminator receives the time-indexed neural window and predicts one of three classes: unimanual left ($0.76 \pm 0.01$8), unimanual right ($0.76 \pm 0.01$9), or bimanual ($0.69 \pm 0.03$0). Its architecture is a 128-unit LSTM, dropout $0.69 \pm 0.03$1, a 64-unit dense layer, dropout $0.69 \pm 0.03$2, and a sigmoid output layer yielding class probabilities

$0.69 \pm 0.03$3

Routing is by $0.69 \pm 0.03$4, which selects one of three decoder branches.

The second stage contains three GRU-based decoders. The L-Decoder is trained only on unimanual left trials and outputs left-hand velocities $0.69 \pm 0.03$5. The R-Decoder is trained only on unimanual right trials and outputs right-hand velocities $0.69 \pm 0.03$6. The Bi-Decoder is trained on all motion types and outputs both hands jointly as a 4D vector

$0.69 \pm 0.03$7

Each decoder uses a single 512-unit GRU followed by a linear readout layer.

The recurrent update is standard GRU:

$0.69 \pm 0.03$8

$0.69 \pm 0.03$9

$2\%$0

$2\%$1

The regression loss is mean squared error,

$2\%$2

and the classification loss is cross-entropy,

$2\%$3

The combined objective is

$2\%$4

Performance is reported with coefficient of determination and Pearson correlation, with $2\%$5 preferred as the main metric:

$2\%$6

The architectural rationale is asymmetrical. GRUs are used in the decoders because they achieved similar accuracy to LSTMs with better efficiency, at approximately $2\%$7 of LSTM training time, whereas the discriminator retains an LSTM because it gave slightly better classification (Robert et al., 19 Aug 2025). The reported model sizes are approximately 500k parameters for the discriminator and about 2M for each decoder, for total storage near 10 MB, with inference latency of 100–150 ms on GPU.

4. Training protocol and evaluation design

Training proceeds in two stages. First, the discriminator is trained on motion-type labels using classification loss. Second, the decoders are trained for velocity regression. The branch-specific sample allocation is deliberate: the L- and R-Decoders receive only their respective unimanual trials, whereas the Bi-Decoder receives all samples so that it can learn both shared structure and interaction patterns across uni- and bimanual movement (Robert et al., 19 Aug 2025).

The paper reports dropout $2\%$8 as the only explicit regularization in the discriminator. No additional weight decay is reported, and curriculum learning and early stopping are not explicitly described. Optimizer, learning rate, batch size, epoch count, initialization, and hardware are also not reported (Robert et al., 19 Aug 2025).

Evaluation uses both within-session and cross-session protocols. For earlier sessions, a 10-fold cross-validation procedure is used with identical preprocessing and segmented windows, while maintaining comparable class distributions. For cross-session generalization, models are trained on all sessions preceding a target session, then fine-tuned on 40% of the target session and evaluated on the remaining 60%. An additional cross-session analysis over the last three sessions is used to quantify robustness to session variability (Robert et al., 19 Aug 2025).

This protocol is important because the paper’s claims are not limited to average decoding accuracy. Session variability, domain shift, and the need for limited fine-tuning are treated as core operational constraints. Per-session normalization, Gaussian smoothing, and time-index augmentation are presented as measures that reduce noise, mitigate nonstationarity, and stabilize phase-aware decoding across sessions.

5. Benchmarks, quantitative performance, and ablations

BiND is benchmarked against six baselines: SVR with RBF kernel, XGBoost, FNN, CNN, Transformer, and a single-branch GRU decoder, all trained on the same preprocessed threshold-crossing windows with identical smoothing, normalization, windowing, and train/validation/test splits (Robert et al., 19 Aug 2025). The intent is to isolate architecture and training strategy rather than feature engineering.

The main aggregate results are concise. BiND achieves mean $2\%$9 of $4\%$0 for unimanual decoding and $4\%$1 for bimanual decoding, exceeding the next-best GRU baseline by approximately $4\%$2 $4\%$3 in both tasks, with correlation improvements around $4\%$4 (Robert et al., 19 Aug 2025). In the final three-session cross-session analysis, BiND reaches about $4\%$5 for unimanual and about $4\%$6 for bimanual decoding, compared with GRU at about $4\%$7 and $4\%$8, respectively, yielding gains of roughly $4\%$9–$10$0.

Detailed limb-wise examples reinforce the asymmetry of bimanual decoding. For unimanual right-hand motion, $10$1 is approximately $10$2 with 95% confidence interval $10$3 for $10$4 and $10$5. For unimanual left-hand motion, $10$6 is approximately $10$7 with 95% confidence interval $10$8 for $10$9 and approximately $16\%$0 with 95% confidence interval $16\%$1 for $16\%$2 (Robert et al., 19 Aug 2025).

The ablation results explain where the gains originate. Removing the motion-type discriminator and replacing the routed architecture with a single GRU decoder reduces accuracy, supporting the claim that specialization by routing contributes materially to the roughly $16\%$3 advantage over standard GRU. Removing the trial-relative time index also degrades performance. The reported $16\%$4 gains from adding $16\%$5 are:

• Unimanual: $16\%$6 to $16\%$7 per output component.
• Bimanual: $16\%$8 to $16\%$9 per output component.
• Average: approximately $R^2$0.

The gains are stronger for the dominant hand in bimanual movement, with roughly $R^2$1 for the right hand versus $R^2$2 for the left hand, while in unimanual movement the gains are more balanced at about $R^2$3 for the right hand and $R^2$4 for the left hand (Robert et al., 19 Aug 2025).

The discriminator itself is assessed qualitatively. UMAP visualizations show well-separated clusters, and more than 80% of bimanual data points are correctly classified, although precise overall discriminator accuracy and a confusion matrix are not reported (Robert et al., 19 Aug 2025).

6. Robustness, limitations, and practical significance

A central claim of BiND is robustness to session variability. The reported cross-session gains over GRU, together with the improvement under limited target-session fine-tuning, indicate that task-aware routing plus temporal context can mitigate domain shift better than a single-branch recurrent decoder (Robert et al., 19 Aug 2025). The paper attributes this to three interacting mechanisms: per-session normalization and Gaussian smoothing reduce nonstationarity and noise; the time index provides phase-awareness across trials; and the discriminator partitions the problem into simpler submappings.

The method nevertheless has clear limitations. Misclassification at the first stage can route a sample to a suboptimal decoder and thereby degrade velocity prediction. Accuracy also declines under high interlimb coupling, particularly for the non-dominant hand during bimanual tasks. More fundamentally, the empirical basis is limited to a single participant and 13 sessions in an open-loop 2D cursor-control setting, so generalization to broader populations, additional motor behaviors, or closed-loop control remains to be established (Robert et al., 19 Aug 2025).

The paper presents several extensions as natural next steps: multi-participant validation, adaptive online closed-loop deployment, lighter-weight implementations for embedded BCI hardware, incorporation of sensory feedback, and extension beyond 2D velocity decoding to reaching, grasp, forces, and potentially 3D kinematics (Robert et al., 19 Aug 2025). These are prospective directions rather than demonstrated results.

From an operational standpoint, BiND is positioned as compatible with real-time BCI use. Reported GPU inference latency is 100–150 ms, and the architecture’s total storage is around 10 MB (Robert et al., 19 Aug 2025). This suggests that the discriminator-decoder cascade is not merely an offline analysis device but a practical architecture for low-latency bilateral prosthetic control, provided that future work confirms its behavior in adaptive closed-loop settings.

BiND’s broader significance lies in its decomposition of the decoding problem. Rather than treating all motor conditions as instances of a single regression function, it models motion context explicitly and injects trial phase directly into the temporal representation. A plausible implication is that future intracortical decoders for complex multi-effector control may benefit less from ever-larger generic sequence models than from structured conditional architectures that separate context inference from kinematic prediction.