Spikachu: Causal, Energy-Efficient Neural Decoder

• Spikachu is a scalable, causal, energy-efficient spiking neural network decoder designed for real-time brain-computer interfaces using only present and past inputs.
• It integrates multi-scale SNN modules, a Perceiver-style harmonizer, and an observer-based readout to extract latent features and produce continuous behavioral predictions.
• Empirical benchmarks show up to 418× lower energy consumption with competitive decoding accuracy compared to traditional ANN-based decoders.

Spikachu is a scalable, causal, and energy-efficient neural decoding framework based on spiking neural networks (SNNs), designed specifically for brain-computer interfaces (BCIs) in resource-constrained settings. Spikachu processes extracellularly recorded spikes from arbitrarily structured electrode arrays, harmonizes heterogeneous neural signals, and produces continuous behavioral predictions, such as kinematic velocity, operating strictly with present and past inputs for real-time BCI compatibility. Empirical results demonstrate Spikachu's superior energy efficiency—achieving between 2.26× and 418.81× lower energy consumption than established ann-based decoders—while delivering decoding accuracy on par with state-of-the-art offline models (Mentzelopoulos et al., 23 Oct 2025).

1. Architectural Overview

Spikachu's architecture comprises three principal modules: a harmonizer for latent space projection, stacked multi-scale causal spiking modules for feature extraction, and an observer-based readout for behavioral decoding. The input consists of binned spike counts (e.g., 10 ms bins), where each active recording unit per bin is treated as a "token." A learnable embedding function UnitEmb maps each token (unit index $u_i$) to a d-dimensional vector $x_i$, which are concatenated as $X \in \mathbb{R}^{n_u \times d}$.

To reconcile inconsistent array geometries across sessions and subjects, Spikachu utilizes a Perceiver-style cross-attention harmonizer. For a set of learned latent queries $Z_0 \in \mathbb{R}^{n_0 \times d}$:

$Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$

$$Z_1 = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right) V$$

$Z_1$ is flattened and projected into a vector $Z_2 \in \mathbb{R}^{n_v}$—the session- and subject-agnostic representation termed "virtual units," enabling session-invariant downstream processing.

The feature extraction pipeline includes:

• Multi-Scale SNN-I: $Z_2$ is processed in parallel by $P$ two-layer SNN-MLPs, each initialized with a distinct, learnable membrane time constant $x_i$0, yielding concatenated outputs $x_i$1.
• Spiking Self-Attention (SSA): Long-range dependencies across timescales are modelled via spiking-attention over $x_i$2, where spikes serve as hard nonlinearity, circumventing the need for softmax and retaining causal, local computations. A spiking MLP reprojects this to $x_i$3.
• Mixing MLP and Multi-Scale SNN-II: $x_i$4 is flattened and compressed via a spiking MLP, then passed through a second bank ($x_i$5) of multi-timescale SNN-MLPs, producing $x_i$6.
• Observer Readout: A final observer layer accumulates the membrane potential without emitting spikes. The observer state $x_i$7 is linearly projected to $x_i$8, the predicted (vₓ, vᵧ) velocity.

All operations, including attention modules, are strictly causal—no peeking ahead—facilitated in attention modules by lower-triangular temporal masking.

2. Spiking Neuron Model and Learning Dynamics

All spiking layers employ the Leaky Integrate-and-Fire (LIF) neuron model. The continuous-time subthreshold potential $x_i$9 follows:

$X \in \mathbb{R}^{n_u \times d}$0

which is discretized per time step $X \in \mathbb{R}^{n_u \times d}$1 as:

$X \in \mathbb{R}^{n_u \times d}$2

Upon surpassing a threshold $X \in \mathbb{R}^{n_u \times d}$3, the neuron emits a binary spike $X \in \mathbb{R}^{n_u \times d}$4 and the voltage resets to $X \in \mathbb{R}^{n_u \times d}$5. To permit gradient-based training through discontinuous spikes, a smooth surrogate derivative, specifically $X \in \mathbb{R}^{n_u \times d}$6, is adopted for backpropagation.

3. Training Procedure and Causality

Spikachu is trained end-to-end to minimize the squared error between predicted and true velocity per bin:

$X \in \mathbb{R}^{n_u \times d}$7

For center-out (CO) tasks, errors during reaching epochs are upweighted by a factor of 5. Optimization employs the LAMB optimizer with a weight decay of $X \in \mathbb{R}^{n_u \times d}$8, a fixed learning rate of $X \in \mathbb{R}^{n_u \times d}$9 for the first 75% epochs, followed by cosine annealing. Typical training entails 1,000 epochs (batch size 128) for single sessions; multi-session pretraining uses batch size 512 for 400 epochs.

Strict causality is maintained by ensuring (a) each time bin’s tokens are only processed once (no future context), and (b) all self- or cross-attention masks are strictly lower-triangular in time.

4. Energy Accounting and Operational Efficiency

Spikachu quantifies energy cost by enumerating floating-point multiply-accumulate (MAC) and accumulate (AC) operations per forward pass. Standard hardware costs are $Z_0 \in \mathbb{R}^{n_0 \times d}$0 pJ, $Z_0 \in \mathbb{R}^{n_0 \times d}$1 pJ. For ANN layers:

$Z_0 \in \mathbb{R}^{n_0 \times d}$2

$Z_0 \in \mathbb{R}^{n_0 \times d}$3

In SNN layers, only the proportion $Z_0 \in \mathbb{R}^{n_0 \times d}$4 of active (spiking) neurons incurs subsequent operations (spike-induced MACs are replaced by ACs):

$Z_0 \in \mathbb{R}^{n_0 \times d}$5

SSA layers omit softmax, further reducing energy, with all accumulation and multiplication operations performed only with ±1 increments.

Empirical measurements on single-session models yield a mean energy consumption of $Z_0 \in \mathbb{R}^{n_0 \times d}$6J/inference for Spikachu, versus $Z_0 \in \mathbb{R}^{n_0 \times d}$7J for GRUs and $Z_0 \in \mathbb{R}^{n_0 \times d}$8J for noncausal transformer POYO, corresponding to 2.26× and 418.8× lower requirements, respectively.

5. Performance Benchmarks and Scaling Laws

Performance evaluation encompasses 99 single-session models across three monkeys (78 CO and 21 RT sessions), with Spikachu attaining mean $Z_0 \in \mathbb{R}^{n_0 \times d}$9 for CO and $Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$0 for RT tasks. By comparison, GRU achieves $Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$1 at twice the energy, and POYO (noncausal) reaches $Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$2 at over 400× the energy consumption.

Pretraining Spikachu-mp on all sessions (≈43 hours, 110M spikes), followed by session-specific finetuning, increases mean $Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$3 to $Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$4 for CO/RT with energy reduced to ~5.04 $Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$5J. Transfer to held-out sessions from a new subject yields $Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$6 at ≈$Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$7J/inference and converges 3–4× faster. Scaling laws indicate gradual improvement in $Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$8 and energy efficiency as the number of sessions added for pretraining increases from 20 to 99, benefitting both seen and novel sessions, including new subjects.

On the Neural Latents Benchmark tasks (MC-RTT, MC-Maze), scratch-trained Spikachu matches prior top models ($Q = W_Q Z_0,\quad K=W_K X,\quad V=W_V X$9) at $$Z_1 = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right) V$$0J, with finetuned models converging 2–2.6× faster and saving an additional ~3% energy.

Summary table of benchmark results:

Model	CO $$Z_1 = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right) V$$1 / RT $$Z_1 = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right) V$$2	Energy ($$Z_1 = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right) V$$3J/inference)
Spikachu	0.84 / 0.68	5.14
GRU	0.83 / 0.67	11.65
POYO	0.89 / 0.68	2151

6. Component Analysis and Ablation Results

Ablation studies attribute the largest drops in performance to removal of the multi-scale SNN blocks, confirming the centrality of integrating across multiple intrinsic time constants $$Z_1 = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right) V$$4. Exclusion of the harmonizer or SSA reduces multi-session generalization; however, SSA is more critical for generalization than for single-session performance. The event-driven, stateful character of LIF neurons is essential: replacing all spiking activations with ReLU, while keeping network connectivity constant, degrades $$Z_1 = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right) V$$5 to approximately $$Z_1 = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right) V$$6, even in the presence of explicit temporal context. This empirical result demonstrates that Spikachu's efficacy arises from the LIF dynamics rather than architecture alone.

7. Technological Implications and Significance

Spikachu is the first end-to-end spiking-network decoder to operate strictly causally on present and past inputs, harmonize diverse multi-session and multi-subject neural data, combine multi-timescale spiking modules with spiking attention mechanisms for latent feature extraction, and achieve competitive accuracy with orders-of-magnitude lower energy cost. This profile makes Spikachu a viable solution for battery-constrained, implantable BCI systems requiring efficient, scalable, and real-time-compatible neural decoding (Mentzelopoulos et al., 23 Oct 2025).