NeuCODEX: Edge–Cloud SNN Inference

• NeuCODEX is an edge–cloud co-inference architecture for SNNs that jointly compresses intermediate spike tensors and adapts processing timesteps via a confidence-based early exit.
• It employs a learned encoder–decoder bottleneck and integrates both spatial and temporal redundancy reduction to minimize communication costs and computational delay.
• The system demonstrates significant energy savings and latency improvements on static and neuromorphic datasets using ResNet-18 and VGG-16, with accuracy drops typically under 2%.

NeuCODEX is an edge–cloud co-inference architecture for Spiking Neural Networks (SNNs) that combines spike-driven learned compression at the edge–cloud split point with dynamic early exit over timesteps. It was proposed to address a specific deployment problem: although SNNs are often regarded as energy-efficient, full edge-side inference remains difficult because accuracy is tied to repeated timestep processing, which increases latency and energy, while naive split inference can incur prohibitive feature-transmission cost. NeuCODEX therefore targets both spatial redundancy in intermediate spike tensors and temporal redundancy across timesteps, and was evaluated on static images and neuromorphic event streams using ResNet-18 and VGG-16 backbones in a real edge-to-cloud testbed (Hassan et al., 23 Sep 2025).

1. Concept and deployment rationale

NeuCODEX is organized around the claim that efficient SNN deployment requires simultaneous control of computation, communication, and timestep count. The motivating difficulty is not merely model depth, but the fixed and high timestep overhead of SNN inference: even when per-timestep computation is cheap, repeated processing over timesteps raises latency and edge energy. Offloading part of the network to the cloud can reduce edge computation, but raw split-layer spike tensors can be so large that communication becomes the dominant bottleneck (Hassan et al., 23 Sep 2025).

The architecture therefore jointly optimizes two forms of redundancy. The first is spatial redundancy, addressed by a learned bottleneck that compresses intermediate spike feature maps before transmission. The second is temporal redundancy, addressed by a confidence-based stopping rule that terminates inference as soon as the current output is sufficiently confident. The paper’s central thesis is that prior work generally studied compression only or early exit only, whereas these mechanisms are complementary: compression reduces the cost of each transmitted timestep, and early exit reduces the number of transmitted timesteps (Hassan et al., 23 Sep 2025).

This framing places NeuCODEX within neuromorphic edge–cloud systems rather than generic ANN split computing. The paper explicitly contrasts it with standard ANN split computing, noting that ANN methods typically compress dense intermediate activations, whereas NeuCODEX operates on spike-based intermediate tensors and exploits time-domain adaptivity, which is especially natural for SNNs. It also contrasts the system with prior SNN co-inference methods such as EC-SNN, which transmits raw uncompressed spike features, and with compression-focused approaches such as SpikeBottleNet and SNN-SC, which do not address temporal redundancy through input-adaptive early exit (Hassan et al., 23 Sep 2025).

2. System architecture and inference workflow

NeuCODEX partitions an SNN into an edge-side front end and a cloud-side back end, denoted in the paper as $f_{\text{edge}}$ and $f_{\text{cloud}}$. At the split point it inserts an encoder–decoder bottleneck: the edge encoder compresses the split-layer spike tensor, and the cloud decoder reconstructs an approximate feature tensor for the remaining cloud-side network. The per-timestep workflow is fixed. The edge runs early SNN layers, compresses the split-layer spike activation $X_t$ into $Z_t$, transmits $Z_t$ to the cloud, the cloud decodes it into $\hat{X}_t$, processes the remaining layers, returns logits $y^{(t)}$ to the edge, and the edge decides whether another timestep is necessary (Hassan et al., 23 Sep 2025).

The split-layer spike tensor is formalized as

$X_t \in \{0,1\}^{C \times H \times W},$

and the transmitted bottleneck code as

$Z_t \in \mathbb{R}^{C' \times W' \times H'},$

with $C' \ll C$, $f_{\text{cloud}}$0, and $f_{\text{cloud}}$1. The compressed representation is therefore the communicated object, not the original spike tensor. Compression is performed by

$f_{\text{cloud}}$2

where $f_{\text{cloud}}$3 is a 2D convolution, $f_{\text{cloud}}$4 is batch normalization, $f_{\text{cloud}}$5 is a spiking neuron layer implemented as LIF, and $f_{\text{cloud}}$6 are learned encoder weights. Reconstruction at the cloud uses

$f_{\text{cloud}}$7

where $f_{\text{cloud}}$8 is a 2D transposed convolution and $f_{\text{cloud}}$9 are learned decoder weights (Hassan et al., 23 Sep 2025).

NeuCODEX was implemented on spiking ResNet-18 and VGG-16, with LIF neurons inserted after each convolutional and fully connected layer. Candidate partition points are denoted $X_t$0. The paper states that higher-index split points correspond to increasing feature sizes—larger $X_t$1 and larger $X_t$2—and therefore greater transmission costs, although it also notes that this naming is somewhat unconventional (Hassan et al., 23 Sep 2025).

The system supports both static image and neuromorphic event-stream inputs. Static images from CIFAR-10 and Caltech are converted to spikes by rate coding over two timesteps, with higher pixel intensity producing a higher firing rate. CIFAR10-DVS and N-Caltech are already in spike/event form, so no static-to-spike conversion is required. Other than input encoding, the edge–cloud workflow is unchanged across modalities (Hassan et al., 23 Sep 2025).

3. Spike-driven compression and dynamic early exit

The compression module is a learned neural bottleneck rather than a full rate–distortion codec. The paper does not describe an explicit entropy model, arithmetic coding, Huffman coding, masking scheme, quantization module beyond the spike representation itself, or any explicit rate term in the objective. It also does not report a reconstruction loss for the bottleneck. Instead, the encoder–decoder pair is optimized indirectly through downstream classification performance, with the bottleneck explicitly treated as lossy in the sense that $X_t$3 in general (Hassan et al., 23 Sep 2025).

The early-exit mechanism is defined over timesteps, not over depth. At timestep $X_t$4, the cloud outputs logits

$X_t$5

which are converted into probabilities

$X_t$6

The confidence score is then

$X_t$7

and inference stops at the earliest timestep satisfying

$X_t$8

If no timestep reaches the threshold, inference continues to $X_t$9. The final prediction is

$Z_t$0

The paper reports that $Z_t$1 gave a good latency–accuracy tradeoff across datasets, and many reported results use $Z_t$2 (Hassan et al., 23 Sep 2025).

A key design point is that the stopping policy is edge-controlled but cloud-informed. The cloud computes logits, sends them back, and the edge computes confidence and decides whether to continue. The paper does not define a temporal logit-accumulation rule such as averaging or cumulative evidence integration across timesteps; confidence is computed directly from timestep-$Z_t$3 logits (Hassan et al., 23 Sep 2025).

The observed exit behavior is data dependent. With $Z_t$4, the main paper reports that roughly 50–70% of samples exit after the first timestep. The appendix gives more specific cases: for ResNet-18 on CIFAR-10, average transmitted tensors were 1.18 instead of 2; VGG-16 on static datasets had roughly 66–97% first-step exits; on CIFAR10-DVS with VGG-16, 0% exited at the first timestep, which the paper attributes to early spikes being too sparse (Hassan et al., 23 Sep 2025).

The compression gains are correspondingly large when the bottleneck is effective. For ResNet-18 on CIFAR-10 at split point SP7, the paper reports transmission reduction from 16,384 bits to 8 bits, a 2048× reduction. Other reported reductions include 1056× on Caltech at SP5, 154× on CIFAR10-DVS at SP3, and about 458× on N-Caltech at SP3. For VGG-16, reported figures include 1024× on CIFAR-10 at SP8, 128× on CIFAR10-DVS at SP3, 1024× on Caltech at SP8, and 256× on N-Caltech at SP3 (Hassan et al., 23 Sep 2025).

4. SNN formulation and optimization procedure

NeuCODEX uses Leaky Integrate-and-Fire neurons. The membrane dynamics are given by

$Z_t$5

$Z_t$6

$Z_t$7

where $Z_t$8 is the pre-spike membrane potential at timestep $Z_t$9, $Z_t$0 is the post-reset membrane potential, $Z_t$1 is the input current, $Z_t$2 is the output spike, $Z_t$3 is the membrane time constant, $Z_t$4 is the firing threshold, $Z_t$5 is the reset voltage, and $Z_t$6 is the Heaviside step function. The paper uses SpikingJelly default parameters

$Z_t$7

A spike occurs when $Z_t$8 (Hassan et al., 23 Sep 2025).

The training pipeline is staged. First, the base SNN backbone is trained over $Z_t$9 timesteps using surrogate gradient descent and classification loss. Second, the encoder–decoder bottleneck is inserted at a chosen split point. Third, the modified edge–cloud model is fine-tuned end to end so that the network adapts to lossy bottleneck reconstruction. Dynamic early exit is then added only at inference time; the paper does not describe a special exit-aware training loss (Hassan et al., 23 Sep 2025).

The only training objective explicitly named is classification cross-entropy,

$\hat{X}_t$0

The paper does not provide a fuller combined loss and does not describe optimizer type, learning rate, batch size, training epochs, weight decay, scheduler, or the exact surrogate gradient function. This under-specification is one of the method’s reproducibility limitations (Hassan et al., 23 Sep 2025).

The experimental energy model is also indirect rather than hardware-measured. Edge energy is estimated theoretically using approximately 23 pJ per SynOp from Intel Loihi, with SynOps estimated from layer FLOPs multiplied by firing rate and number of timesteps. The paper does not provide a single formal energy equation, and cloud energy cost is not included (Hassan et al., 23 Sep 2025).

5. Evaluation, benchmarks, and empirical results

NeuCODEX was evaluated on four datasets: CIFAR-10, Caltech, CIFAR10-DVS, and N-Caltech. The real deployment setup used an NVIDIA Jetson Orin Nano as the edge device and an NVIDIA A100-SXM4 GPU server as the cloud, communicating over standard Wi-Fi with measured throughput of approximately 18.9 Mbps. The paper compares four configurations: F–B (fixed timesteps, no bottleneck), D–B (dynamic timesteps, no bottleneck), F+B (fixed timesteps with bottleneck compression), and D+B (dynamic timesteps with bottleneck), with D+B corresponding to NeuCODEX (Hassan et al., 23 Sep 2025).

Setting	F–B accuracy	D+B accuracy
ResNet-18, CIFAR-10	94.15%	92.63%
ResNet-18, CIFAR10-DVS	65.8%	67.4%
ResNet-18, Caltech	97.92%	97.62%
ResNet-18, N-Caltech	94.25%	93.95%
VGG-16, CIFAR-10	91.35%	90.09%
VGG-16, CIFAR10-DVS	63.95%	62.8%
VGG-16, Caltech	97.92%	97.17%
VGG-16, N-Caltech	91.45%	89.23%

These results support the paper’s general characterization that NeuCODEX usually incurs less than about 2% accuracy degradation, with some cases even improving accuracy, especially on CIFAR10-DVS with ResNet-18. The abstract states a “negligible accuracy drop of less than 2%,” while the detailed tables show that the effect is usually, but not uniformly, below that level (Hassan et al., 23 Sep 2025).

Latency results are reported against both edge-only inference and naive uncompressed split inference. For ResNet-18 on N-Caltech at SP1, edge-only latency is 21.79 ms and NeuCODEX latency is 14.85 ms, a 31.8% reduction. On Caltech at SP1, latency drops from 107.19 ms to 68.04 ms, a 36.5% speedup. On CIFAR10-DVS at SP3, latency decreases from 28.05 ms to 25.06 ms, about 11% faster. In a CPU edge scenario on CIFAR-10, latency drops from 14.86 ms to 9.39 ms, a 37% speedup. Against naive uncompressed split inference, the improvements are much larger: for ResNet-18 on N-Caltech at SP1, uncompressed split inference yields 597.29 ms total latency, versus 14.85 ms with NeuCODEX, around a 97% reduction and roughly 40× faster (Hassan et al., 23 Sep 2025).

The latency gains are not universal. The paper explicitly notes that for VGG-16 on CIFAR-10, NeuCODEX is slightly slower than edge-only inference because bottleneck overhead outweighs the benefit of offloading in that setting. This suggests that the practical benefit depends strongly on split point, dataset resolution, backbone speed, wireless throughput, and the cost of encoding/decoding relative to saved computation (Hassan et al., 23 Sep 2025).

Reported edge-energy reductions are consistently large. For ResNet-18, the paper reports reductions from 5.45 mJ to 0.43 mJ on CIFAR-10, from 49.83 mJ to 4.90 mJ on CIFAR10-DVS, from 152.76 mJ to 9.93 mJ on Caltech, and from 49.90 mJ to 2.75 mJ on N-Caltech. For VGG-16, reported reductions are from 1.95 mJ to 0.02 mJ on CIFAR-10, from 13.58 mJ to 0.06 mJ on CIFAR10-DVS, from 43.79 mJ to 0.27 mJ on Caltech, and from 16.80 mJ to 0.04 mJ on N-Caltech. These results underpin the abstract’s claim of over 90% edge energy savings (Hassan et al., 23 Sep 2025).

The paper also reports an ablation on larger timestep budgets for ResNet-18 on CIFAR-10 at SP3 with CPU edge execution. Increasing $\hat{X}_t$1 from 4 to 8 improved accuracy from 85.1% to 88.3% while decreasing average latency from 43.5 ms to 28.6 ms. This suggests that a larger maximum timestep budget can improve both accuracy and latency when dynamic early exit is active, because hard samples are allowed more computation while easy samples still terminate early (Hassan et al., 23 Sep 2025).

6. Significance, limitations, and broader context

NeuCODEX’s principal significance lies in making SNN split inference explicitly neuromorphic rather than merely ANN-inspired. Its architectural novelty is not only that it compresses intermediate representations, but that it combines spike-driven compression with timestep-adaptive stopping. In the paper’s formulation, effective deployment of SNNs requires exploiting both spike sparsity in space and redundancy in time. This gives the work a distinct place relative to ANN split computing and prior SNN co-inference methods (Hassan et al., 23 Sep 2025).

The work also exposes several limitations. First, the main experiments use very small timestep budgets, with $\hat{X}_t$2 in the standard setup, so the observed early-exit behavior is demonstrated in a shallow temporal regime. Second, the compression module is simple: it is a learned bottleneck, but it does not include explicit entropy coding or rate–distortion optimization. Third, split-point selection is empirical rather than optimized. Fourth, energy is estimated rather than directly measured on neuromorphic hardware. Fifth, cloud energy cost is ignored. Sixth, some reproduction-critical details—optimizer, learning rate, batch size, scheduler, exact surrogate function—are not stated. Finally, latency benefits are not guaranteed and can reverse on small settings where encoder–decoder overhead dominates (Hassan et al., 23 Sep 2025).

A common misconception is that NeuCODEX is primarily a general-purpose SNN compression method. The paper’s design is narrower and more specific: it is an edge–cloud co-inference system. Compression is important, but it is meaningful only in conjunction with model partitioning and confidence-driven timestep control. Another potential misunderstanding is that the system performs early exit over network depth. It does not; the split point is fixed, and early exit occurs only over timesteps (Hassan et al., 23 Sep 2025).

The paper suggests several future directions, including spiking Vision Transformers, object detection, segmentation, super-resolution, and hardware-level implementation on neuromorphic chips. A plausible implication is that NeuCODEX should be read less as a finalized deployment stack than as a systems template: SNN co-inference can be made practical when communication reduction and temporal adaptivity are treated as a joint optimization problem rather than as isolated engineering add-ons (Hassan et al., 23 Sep 2025).