Rotation-equivariant networks save parameters by reusing one learned filter in every orientation. A filter trained to detect a vertical edge is rotated to detect edges at 0°, 90°, 180° and 270°, so the network never has to learn those copies separately. The cost appears elsewhere: each feature map now carries one sub-map per orientation, which makes activations four to eight times larger. On a GPU this means more memory. On neuromorphic hardware, which spends energy only when a neuron emits a binary pulse (a "spike"), it means something worse. If every orientation copy fires on its own, the spike count can grow with the number of orientations, and the energy advantage that justifies spiking hardware disappears. We propose to make only the best-matching orientation fire. At each location and channel, the orientation copies compete, and only the copy whose internal voltage is highest may spike. This competition commutes with rotation, so the network stays exactly equivariant. The goal is a spiking network that keeps the parameter savings of equivariance while holding its spike budget to roughly that of an ordinary, orientation-blind spiking network.
The setting is event cameras, sensors that report per-pixel brightness changes asynchronously instead of full frames. Rotation matters here because the event stream depends on how the sensor and the object move relative to each other. Current spiking networks for event data, such as the detector in Bio-Vision-Inspired Spiking Neural Networks for Object Detection with Event Cameras, contain no geometric structure. They must learn rotated versions of the data by brute force.
Core Claim
An equivariant spiking network is useful only if its accuracy per spike beats the alternatives. We claim that a structural constraint, a winner-take-all rule over the orientation copies of each feature, achieves a better accuracy-versus-spike-count trade-off than two simpler options. The first is a planar spiking network trained with rotation augmentation. The second is a group-equivariant spiking network whose firing is suppressed by an ordinary spike-rate penalty. We further claim that this advantage is largest when the network is tested on rotations it never saw during training.
The energy accounting is simple. On accumulate-only hardware, dynamic energy is approximately
$$E \approx e_{AC} \sum_{l} f_l \sum_{t=1}^{T} S_l(t).$$
Here $e_{AC}$ is the cost of one addition, $S_l(t)$ is the number of spikes emitted by layer $l$ at time step $t$, $f_l$ is the number of downstream connections each spike drives, and $T$ is the number of simulation time steps. The proportionality to spike count follows Tavanaei et al. and Venkatesha et al..
A group convolution over the rotation group $p4$ (translations plus 90° rotations) multiplies the number of neurons in each layer by $|G|=4$. Its fan-out $f_l$ also grows by $|G|$, because each spike feeds all output orientations. With winner-take-all over the group dimension, each spatial location and channel emits at most one spike per time step. $S_l(t)$ is then bounded by the spike count of a planar network with the same number of base channels, and parameters remain shared. Fan-out still carries the factor $|G|$. The proposal therefore reports spike counts and synaptic operations separately, and it compares against planar networks widened to match either parameters or operations.
Relation to Prior Work
Two papers supply the method. Group Equivariant Convolutional Networks defines the $p4$ and $p4m$ group convolutions ($p4m$ adds mirror reflections) that we reuse unchanged. It also documents the $|G|$-fold growth in feature maps that motivates this work. Efficient Equivariant Network attacks that growth for dense networks by decoupling kernels. It reduces FLOPs but leaves activations real-valued and dense. Our target is instead the count of discrete events. Exploiting Cyclic Symmetry in Convolutional Neural Networks supports the expectation that a planar network needs roughly four times the parameters to match a cyclic-equivariant one. Our parameter-matched and operation-matched baselines test that expectation directly in the spiking setting.
Lateral inhibition as a way to suppress redundant spikes comes from Optimal ANN-SNN Conversion with Group Neurons. That paper has member neurons with staggered thresholds inhibit one another so that together they approximate one continuous activation, which yields near-lossless conversion at $T=2$. Its "group" is an ensemble of neurons, not a symmetry group. Its inhibition runs over threshold levels of a single scalar. Ours runs over the orientations of a symmetry group. That changes both the justification (equivariance must be preserved) and the failure mode (the network may lose information about secondary orientations).
Continuous Rotation Group Equivariant Network Inspired by Neural Population Coding (CRGEN) is the closest conceptual neighbour. It borrows orientation-tuned population codes from neuroscience to get continuous rotation equivariance with very few parameters. It is entirely non-spiking, and it reports feature blurring in deep stacks. Winner-take-all over orientations is the discrete, spiking counterpart of a sharpened population code, and it may counteract that blurring. Neuromorphic Visual Scene Understanding with Resonator Networks does achieve rotation-aware inference on Loihi, with a 171x energy-delay improvement. It does so with handcrafted vector-symbolic bindings on synthetic letter scenes, reaching 84% on full rigid transforms. We instead learn convolutional features end to end on real event data with standard leaky integrate-and-fire neurons, which accumulate input into a decaying voltage and spike on crossing a threshold.
The spiking-network scaffold comes from Going Deeper With Directly-Trained Larger Spiking Neural Networks (tdBN, a batch normalization that scales inputs relative to the firing threshold) and Deep Residual Learning in Spiking Neural Networks (SEW ResNet, a residual block that adds spikes after the neuron). The spiking point-cloud model in 3DSMT shows that energy-efficient spiking networks remain non-equivariant even in inherently geometric domains.
Research Program
We will build three components.
The first is a spiking $p4$ convolution trained directly with surrogate gradients, which replace the zero derivative of the spike function with a smooth approximation during backpropagation.
The second is a group-equivariant tdBN that computes its statistics jointly over the four orientation copies of each channel. Standard tdBN normalizes each channel independently, which breaks equivariance. Pooling the statistics over orientations restores it. The SEW residual addition is elementwise, so it is already equivariant.
The third component is orientation winner-take-all. Each group of four orientation neurons sharing a location and base channel integrates its input as usual. At each time step, if several neurons cross threshold, only the one with the highest membrane potential fires. The others receive either a reset or a subtractive inhibition; we compare both variants. Taking the maximum over the orientation dimension commutes with cyclically shifting that dimension, so the layer remains exactly $p4$-equivariant. Exact ties, which are rare with real-valued potentials, are allowed to fire jointly so that equivariance is preserved. We will verify equivariance numerically by rotating inputs 90° and checking that the output spike trains are permuted exactly. We also test a softer variant, $k$-winner-take-all with $k=2$, because hard competition may discard useful information about secondary orientations. Where that information matters is the main scientific risk.
Inputs are event streams binned into $T$ time steps. A 90° rotation of a square sensor grid is exact and preserves event polarity, so equivariance holds at the level of raw events. Our analysis tracks firing rate per layer, synaptic operations, accuracy, and the spread of the winning orientation. That spread tells us whether competition collapses onto a single orientation everywhere, which would make the orientation channels useless.
Baselines
The naive reference is a planar spiking network trained without rotation augmentation. We use SEW ResNet at matching depth, with the same $T$, event binning, optimizer, and epochs. It shows how badly geometry-blind models fail on rotated test sets.
The strongest existing method is the same SEW ResNet trained with random 90° rotation augmentation. Deep Residual Learning in Spiking Neural Networks reports 97.92% accuracy on DVS-Gesture. For CIFAR10-DVS, the best spiking result in the file is 73.63% from Shrinking Your TimeStep at an average of 5 time steps. We found no reported result on Lacuna for any spiking network on rotated test splits of these datasets, so all rotated-test numbers will be produced by us under identical conditions. We run this baseline in two sizes: matched in parameters to our model, and widened to match our synaptic operation count. Each size receives the same hyperparameter search budget as our model.
The ablation that isolates the key ingredient is the identical spiking $p4$ network with group-equivariant tdBN but without winner-take-all. We run it both unregularized and with an L1 penalty on firing rate, sweeping the penalty weight to trace a full accuracy-versus-spikes curve.
Beating these baselines means the following. At an equal total spike count, our model must achieve higher accuracy on rotated test data than both augmented planar networks and the rate-penalized equivariant network. Its accuracy on upright test data must be within one point of the augmented planar baseline. Its spike count must stay within a small constant factor of the planar network rather than approaching the up-to-fourfold growth expected when every orientation fires independently.
Why not just…?
Why not just train a planar spiking network with rotation augmentation? In dense networks, CRGEN reports that augmented or planar models fall below 50% accuracy with only 200 training samples, and below 20% on inverted angles when trained upright. Equivariant models in the same study stay above 90%. The file contains no spiking evidence either way, so our augmented, widened baselines are a real test rather than a straw man. If an augmented planar spiking network matched our accuracy per spike, we would conclude that equivariance adds nothing on this hardware budget and report that result.
Why not just add a spike-rate penalty to an equivariant spiking network? This is the most serious threat. A penalty might silence redundant orientation copies as well as structural competition does, with no architectural change. Our ablation traces the full Pareto curve under this penalty. The claim survives only if winner-take-all dominates that curve, especially at low spike budgets. If the curves coincide, the contribution shrinks to group-equivariant tdBN plus an energy analysis, and we would present it that way.
Why not just pool over orientations early, or run an equivariant dense network on an edge accelerator? Pooling over orientations after the first layer removes the fourfold growth, but it throws away relative-orientation structure in deeper layers. We will test this "invariant after layer 1" variant as an extra baseline. The edge-accelerator option still pays for full multiply-accumulates. Spiking networks in the file report 41.7x lower estimated energy than their dense counterparts (Direct training high-performance spiking neural networks). Collapsing events into frames drops DVS-Gesture accuracy to 71.01% (Is Neuromorphic MNIST Neuromorphic?). We therefore compare against a dense $p4$ network only through operation counts, and we do not claim a win beyond that.
Proposed Experiments
The main benchmarks are CIFAR10-DVS and N-Caltech101, each with standard and 90°-rotated test splits. We add a train-upright, test-rotated protocol and a low-data protocol using 10% of the training set.
We treat DVS-Gesture as a deliberate stress test, because some gestures are defined by direction and rotation invariance may hurt there. The model can still read direction from the equivariant features before the final pooling, and we will report whether it does. Rotated N-MNIST serves only as a sanity check, since Is Neuromorphic MNIST Neuromorphic? shows its temporal content is weak.
Metrics are top-1 accuracy, total spikes, synaptic operations, and estimated energy. For the energy estimate we report raw operation counts and apply one fixed cost ratio between additions and multiply-accumulates to all models, so the conclusions do not depend on that constant. We also report parameter count and the numerically measured equivariance error.
Three results would be decisive. First, spike count as a function of the number of orientations must stay flat under winner-take-all while it grows without competition. Second, the Pareto curve of winner-take-all must dominate that of the rate-penalized ablation on rotated tests. Third, the accuracy gap under train-upright, test-rotated must be large, while upright accuracy stays within one point of augmented SEW ResNet.
A negative outcome on the second check would redirect the work toward regularized equivariant spiking networks. A negative outcome on the third would indicate that rotation structure matters little for current event benchmarks, which would itself be worth reporting.
References
- Bio-Vision-Inspired Spiking Neural Networks for Object Detection with Event Cameras
- Deep learning in spiking neural networks
- Federated Learning with Spiking Neural Networks
- Group Equivariant Convolutional Networks
- Efficient Equivariant Network
- Exploiting Cyclic Symmetry in Convolutional Neural Networks
- Optimal ANN-SNN Conversion with Group Neurons
- Continuous Rotation Group Equivariant Network Inspired by Neural Population Coding
- Neuromorphic visual scene understanding with resonator networks
- Going Deeper With Directly-Trained Larger Spiking Neural Networks
- Deep Residual Learning in Spiking Neural Networks
- 3DSMT: A Hybrid Spiking Mamba-Transformer for Point Cloud Analysis
- Shrinking Your TimeStep: Towards Low-Latency Neuromorphic Object Recognition with Spiking Neural Networks
- Direct training high-performance spiking neural networks for object recognition and detection
- Is Neuromorphic MNIST Neuromorphic? Analyzing the Discriminative Power of Neuromorphic Datasets in the Time Domain