When a neural network is split between a small edge device and a server, the edge device has to send its intermediate features over a noisy wireless link. Spiking neural networks (SNNs) are networks whose neurons talk in sparse binary pulses ("spikes") over a few time steps. They are attractive for the edge half of such a split for two reasons. They replace multiplications with additions, which cuts compute energy. Their output is already a stream of bits, so it needs no separate quantization before transmission. The leading spiking design for this setting, SNN-SC, sends every spike with the same protection against channel errors. Yet a spike tensor is very uneven. Some channels and time steps matter far more to the server's prediction than others. The two ways a bit can go wrong are also not alike. A spurious spike that appears where there was silence is a different event from a real spike that is erased.
We propose to treat the transmitted spike tensor the way video coding treats its frames. We would measure how much the downstream task loss rises when each group of bits is corrupted in each direction. A fixed transmission budget would then be divided across groups, with options ranging from "drop entirely" through "send uncoded" to several strengths of error-correcting code. The bet is that, at an equal number of channel uses, this unequal error protection (UEP) gives better accuracy than uniform protection. A second bet is that it also lowers server-side energy. A spiking receiver's dynamic energy scales with the number of spikes it processes (Tavanaei et al., Jang et al.), so channel-created spurious spikes cost both accuracy and energy.
Core Claim
The claim has two parts, and both are testable. The first concerns importance: in a directly trained deep SNN split at a bottleneck layer, flip sensitivity varies by at least an order of magnitude across groups indexed by (feature channel, time step, error direction). The second concerns allocation: a static, offline protection map built from that sensitivity beats two alternatives at matched channel uses. One alternative is uniform protection with noise-aware training. The other is the simpler "drop the unimportant spikes and protect the rest uniformly" scheme. The protection map needs no per-frame side information, because the group layout is fixed by the architecture. That sidesteps the fragile-header problem that dogs hand-built UEP. Error direction is the ingredient specific to spikes. We expect insertions (0→1) to be more harmful than deletions in most groups. The expectation follows from SoftSNN, which found that bit flips raising stored weights make neurons hyperactive and do far more damage than flips lowering them. If this asymmetry holds for transmitted activations, it can be exploited even without codes. Under on-off signaling, moving the receiver's detection threshold trades insertion probability against deletion probability for free. Coding redundancy can then be spent where direction-blind protection is insufficient.
The allocation itself is a small knapsack problem. Let group $j$ have $n_j$ spike bits, receive protection option $r_j$ costing $c(r_j)$ channel uses per bit, and suffer residual insertion and deletion rates $p^{+}_j(r_j)$ and $p^{-}_j(r_j)$ at the operating signal-to-noise ratio. We choose
$$\min_{\{r_j\}} \sum_j \left[ S^{+}_j\, p^{+}_j(r_j) + S^{-}_j\, p^{-}_j(r_j) \right] \quad \text{s.t.} \quad \sum_j n_j\, c(r_j) \le B,$$
Here $S^{\pm}_j$ is the measured increase in validation loss per unit flip rate of each direction in group $j$, and $B$ is the channel-use budget. The "drop" option has $c=0$ and turns every spike in the group into a deletion. This makes the drop-and-protect alternative a special case of the search space. It is therefore a baseline the method must beat, rather than one it can sidestep. The linear-in-flip-rate objective is a first-order approximation, valid at the low residual error rates targeted after coding. We check it by measuring loss at the chosen allocation directly.
Relation to Prior Work
Two kinds of prior work motivate the importance hypothesis. Classical UEP theory shows the payoff when importance is uneven. Borade et al. prove that critical message classes can keep positive error exponents at channel capacity if less critical classes absorb more errors. Video systems realize this gain by concentrating error correction on base layers and reference frames (Wang et al., Yang et al.). Their importance labels come from codec structure; ours must be measured from a learned network. On the spiking side, ReSpawn and SoftSNN show strong asymmetries in how SNNs respond to bit flips: most-significant versus least-significant weight bits, and upward versus downward flips. Both papers protect stored weights inside an accelerator, using bit remapping and clipping circuits. Both study shallow networks trained with spike-timing-dependent plasticity on MNIST. Neither touches transmitted activations, channel codes, or deep directly trained models.
The closest system is SNN-SC. It sends binary spikes from an edge SNN to a cloud SNN and degrades gracefully under bit errors, but every spike receives equal treatment. Its entropy-maximization loss, which spreads spike activity evenly, may work against our hypothesis. If it flattens importance across channels, there is little left for UEP to exploit. We treat this as a real risk and test it directly. Its hybrid neuron also sends residual membrane potentials as continuous values. We keep the link purely binary so that the bit-level analysis is clean. Firing Bits Where It Matters uses an SNN to decide where source bits go in machine-oriented video coding, but it assumes an error-free channel. We address the complementary channel-side question.
For method, we reuse direct training with threshold-dependent batch normalization from Zheng et al. to obtain deep spiking ResNets. From Ninkovic et al. we take the idea of class-weighted error objectives; that paper learns UEP codes with dense floating-point autoencoders. We deliberately avoid learning a new neural channel code. The protection options are short off-the-shelf codes with cheap hard-decision decoders, so the decoder overhead is easy to count. A learned spiking codec is a possible later step, not the contribution.
Research Program
The first stage builds the testbed. We train spiking ResNets with threshold-dependent batch normalization on DvsGesture, CIFAR-10, and CIFAR-100, using few time steps (four to six). Each network is split at an early and at a mid-depth bottleneck, and each split is trained in two versions: with and without an SNN-SC-style entropy regularizer. Three channels are modeled: a binary symmetric channel, binary phase-shift keying over additive Gaussian noise, and on-off keying over additive Gaussian noise with an adjustable detection threshold. The last one produces a tunable asymmetric bit channel.
The second stage is the sensitivity study, which on its own produces a publishable measurement. For each group (feature channel × time step), we inject insertions or deletions at controlled rates on held-out data and record the loss increase $S^{\pm}_j$. We report how spread out these values are and how stable they are across random seeds, datasets, and signal-to-noise ratios. We also report how well they correlate with cheap proxies: time index, mean firing rate, and gradient magnitude. If a free proxy such as firing rate ranks groups almost as well as the measurement does, the method gets simpler. That result is also worth reporting.
The third stage is allocation. We solve the knapsack problem greedily for each signal-to-noise ratio on a grid, using a protection menu of {drop, uncoded, rate-2/3, rate-1/2, rate-1/3} short codes, plus the free detection-threshold shift under on-off signaling. The result is a small lookup table indexed by channel quality. Since the table is fixed per signal-to-noise bin, the only side information needed is the bin index, which is a few heavily protected bits per frame. A short noise-aware fine-tuning pass is then run under the chosen allocation, so that every compared method gets the same adaptation step.
The fourth stage is accounting. We report accuracy against channel uses per inference and against signal-to-noise ratio. For energy, we report the server SNN's synaptic accumulate operations per inference and the edge-side and server-side decoder operation counts. All costs are expressed in operation counts, with no assumed joules-per-operation. That keeps the energy claim honest without neuromorphic hardware, while still exposing whether spurious spikes inflate receiver work.
Baselines
The naive reference point is uncoded transmission of the spike tensor, with the network trained on clean features. It shows how bad the problem is and gives the floor for every curve.
The strongest existing method is SNN-SC-style transmission: a spiking encoder and decoder trained end to end through the noisy channel with the entropy regularizer and no explicit unequal protection (SNN-SC). We give it the same backbone, the same split point, and the same channel-use budget. Any budget beyond the raw spike bits is spent on a uniform code of matching rate. Its regularizer coefficient is tuned on the same validation grid used for our fine-tuning, starting from the value of 1.0 reported as best in that paper. The file gives no reported CIFAR-100 accuracy number for SNN-SC. It reports only that SNN-SC outperforms DNN-based semantic communication once the bit error rate exceeds 0.15, along with its compression ratios, so we reimplement it rather than compare against a quoted figure. As clean-channel ceilings we use the STBP-tdBN results from Zheng et al.: 96.87% on DvsGesture and 93.16% on CIFAR-10. Beating the baseline means higher accuracy at equal channel uses across a contiguous range of signal-to-noise ratios. Equivalently, it means equal accuracy with at least 20% fewer channel uses, with gaps larger than the seed-to-seed spread over three seeds.
The third baseline is drop-and-protect. Low-sensitivity groups are removed entirely and the rest receive one uniform code, with the drop set and code rate chosen by the same knapsack restricted to two options. This isolates whether graded protection adds anything beyond pruning.
The fourth baseline is time-index UEP. Protection decreases with time step, a hand rule borrowed from the intuition that early spikes matter most. It tests whether measured sensitivity beats an obvious structural heuristic.
The key ablation removes error direction. The allocation is computed from direction-blind sensitivity, the average of insertions and deletions, with no threshold shift. This isolates the spike-specific ingredient. A second ablation repeats everything without the entropy regularizer, to test whether flattening spike activity erases the available gain. No reported result was found on Lacuna for learned UEP autoencoders on these vision tasks, so Ninkovic et al. is discussed but not run as a baseline.
Why not just…?
Why not just use one strong uniform code on everything? Theory says uniform protection is wasteful when importance is uneven (Borade et al.), but theory does not tell us whether a spike tensor is uneven enough for this to matter at short blocklengths. The SNN-SC baseline at matched channel uses is exactly this alternative. If the sensitivity study finds a narrow spread, uniform coding wins and we report that.
Why not just train the network to tolerate noise? SNN-SC already shows that training through the channel gives graceful degradation. Noise-aware training is real competition. We remove its advantage by giving every method, ours included, the same fine-tuning pass. Training alone cannot change the residual error rate of the bits the server actually receives. It can only teach the server to cope with whatever errors arrive. The comparison at matched budget shows whether redistributing the error rates adds accuracy on top of robustness learned in training.
Why not just drop unimportant spikes and protect the rest uniformly? This is the most serious threat, because it captures much of the UEP intuition with a single knob. That is why "drop" sits inside our protection menu and drop-and-protect is a separate baseline. If graded protection beats it by less than seed noise, the honest conclusion is that spike streams call for source pruning, not unequal channel coding. The note would then report a pruning result instead.
Why not just use a free proxy like firing rate or time step as the importance score? We measure exactly how well these proxies correlate with measured sensitivity, and we run time-index UEP as a baseline. If a proxy matches, the method becomes cheaper. The direction-aware component would still be untested by any proxy.
Decisive Checks
Four results settle the proposal. The first is the sensitivity spread: if measured sensitivities across groups span less than a factor of about three, the UEP premise fails for this setting, and we stop before building allocators. The second is the comparison of graded protection against drop-and-protect at matched channel uses on DvsGesture and CIFAR-100. The third is the direction ablation under on-off signaling. If direction-aware allocation and the threshold shift give no gain over direction-blind allocation, the spike-specific part of the claim is false. Separately, we check whether receiver accumulate counts rise with uncorrected insertions as predicted. The fourth is the entropy-regularizer interaction: does SNN-SC's spike balancing remove the importance spread that UEP exploits? A positive outcome on the second and third checks would support a specific conclusion: protection for spiking links should be allocated by measured, direction-specific sensitivity, using a fixed table with no per-frame side information.
References
- SNN-SC: A Spiking Semantic Communication Framework for Collaborative Intelligence
- Deep learning in spiking neural networks
- Spiking Neural Networks - Part I: Detecting Spatial Patterns
- SoftSNN: Low-Cost Fault Tolerance for Spiking Neural Network Accelerators under Soft Errors
- Unequal error protection: an information-theoretic perspective
- Network coding for reliable video distribution in multi-hop device-to-device communications
- Advanced Video Technologies and Applications for H.264/AVC and Beyond
- ReSpawn: Energy-Efficient Fault-Tolerance for Spiking Neural Networks considering Unreliable Memories
- Firing Bits Where It Matters: Spiking-Guided Just Recognizable Distortion Modeling for Machine-Centric Video Coding
- Going Deeper With Directly-Trained Larger Spiking Neural Networks
- Autoencoder-Based Unequal Error Protection Codes