Analog in-memory computing does a neural network's matrix multiply inside the memory array. Inputs become voltages on the rows, weights become conductances (how easily each memory cell passes current), and the column wires sum the currents for free. Doing this avoids moving weights between memory and processor, which dominates the energy of deep learning (Hardware for machine learning: Challenges and opportunities). Two problems eat most of that advantage. The first is reading the analog sums out: one benchmark of binary networks on resistive memory found that analog-to-digital converters (ADCs) need 8 bits of resolution to keep accuracy, and at that resolution they take up to 89% of chip area (Comprehensive Benchmarking of Binary Neural Networks on NVM Crossbar Architectures). The second is noise. Spiking neural networks (SNNs) are networks whose neurons emit 0/1 pulses over several time steps and are trained with "surrogate gradients," a smooth stand-in for the derivative of the spike threshold. When such networks run on non-ideal crossbars they collapse, because analog error is injected at every time step and piles up in the neuron's internal state (Examining the Robustness of Spiking Neural Networks on Non-ideal Memristive Crossbars).
We propose to make the spike decision happen in the readout circuit itself. Each weight is stored as a pair of two-state resistive cells whose difference encodes −1, 0 or +1. A 1-bit sense amplifier compares the positive and negative column currents of a crossbar tile and outputs a single bit. That bit is the only thing that leaves the analog domain. The network is trained with a surrogate gradient that is not a hand-picked smooth curve. It is the derivative of the comparator's measured probability of firing under device noise. Two things would be new. This is the first fully 1-bit system, with binary spike inputs, ternary differential weights and 1-bit sensing, and no DACs or ADCs. And it tests whether binarizing a common-mode-cancelled current before temporal integration turns the "temporal trap" into a small rate of bit flips.
Core Claim
The proposal rests on one hypothesis. When noise enters an SNN as a continuous analog error added to the membrane potential (the neuron's running internal sum), every time step contributes uncorrelated error and every systematic error, such as IR drop along the wires or conductance drift, adds up coherently. If the analog quantity is instead the difference of two currents drawn along matched wires, and it is thresholded to one bit before anything is integrated, two things change. Systematic errors shared by $G^+$ and $G^-$ largely cancel, and what remains reaches the digital membrane as occasional bit flips with a probability set by the comparator margin.
The comparator's behaviour can be written down. Let $s_i(t)\in\{0,1\}$ be the input spike on row $i$, $G^\pm_{ij}$ the two conductances of weight $w_{ij}$, $\theta_j$ the sense threshold of tile column $j$, and $\sigma_j(t)$ the standard deviation of the differential current noise. The tile output fires with probability
$$p_j(t)=\Phi\!\left(\frac{\sum_i s_i(t)\,(G^+_{ij}-G^-_{ij})-\theta_j}{\sigma_j(t)}\right),$$
where $\Phi$ is the standard normal cumulative distribution. Here $\sigma_j(t)$ is not a constant. It grows with the number of active rows whose cells sit in the low-resistance state, because device variation scales with conductance.
We use $\partial p_j/\partial(\cdot)$ as the backward-pass surrogate. Its width therefore shrinks when a column has few active low-resistance cells and widens when it has many. This rule is fixed by the hardware, not chosen by the researcher. In the forward pass we sample the noisy comparator rather than evaluating an ideal Heaviside step. The digital neuron then adds up the tile bits, with learned per-tile scale factors, into a leaky integrate-and-fire membrane and emits the layer's output spike.
Research Program
Each direction in the file plays a different role here. Signed weight representation supplies the cell. We reuse the 2T2R ternary mapping of Low Power In-Memory Implementation of Ternary Neural Networks with Resistive RAM-Based Synapse, where +1 is (low, high) resistance, −1 is (high, low) and 0 is (high, high). It was shown on a 130nm RRAM chip for static networks. We also reuse its error taxonomy: sign flips, and zeros misread as nonzero. The single-sense readout in Implementation of Ternary Weights With Resistive RAM Using a Single Sense Operation Per Synapse shows the zero state can be resolved reliably and tolerates up to 18.5% zero-to-±1 errors with little accuracy loss. Both papers motivate our choice of the zero state, since it draws almost no current and so reduces IR drop. Neither puts the cell inside a temporal spiking neuron.
The von Neumann bottleneck supplies the constraint. The ADC-less SNN accelerator of Hardware/Software Co-Design With ADC-Less In-Memory Computing Hardware for Spiking Neural Networks already replaces ADCs with 1-bit sense amplifiers on partial sums. It reports a 1.2–2.4% CIFAR-10 drop and 2–7× energy reduction, but it keeps 4-bit multi-level weights, trains in three stages and uses no differential cancellation. We remove the multi-level weights and make the sensing differential. That matters because binary-state RRAM cells are far easier to program reliably than multi-level ones.
Binary networks and spike approximation supply the training method. A Self-Ensemble Inspired Approach for Effective Training of Binary-Weight Spiking Neural Networks shows binary-weight SNNs reach 95.04% on CIFAR-10 in ideal software. It uses Spatial Learning Through Time, which drops gradient flow across time steps, together with shortcuts and distillation. We adopt its training recipe as the starting point. SURGE: Surrogate Gradient Adaptation in Binary Neural Networks shows that how gradients are shaped through a binarizer matters, with +3.9% over IR-Net on ImageNet. That supports the idea that the surrogate's shape is worth getting right. SURGE's shape is learned from data, whereas ours comes from device physics.
ReverB-SNN acts as a constraint. Its real-valued spikes would need input DACs on every row, so we keep inputs strictly binary.
We will build a simulator that combines MemTorch-style device non-idealities with a differential tile model. The MemTorch-style part covers finite states, device-to-device variation, stuck-at faults and tile geometry, following MemTorch. The tile model covers wire resistance, comparator offset and thermal noise. We will fit $\sigma_j$ as a function of active low-resistance count from published device statistics and use NeuroSim parameters for area and energy. Training then has four stages: ideal ternary-weight SNN pretraining; insertion of the 1-bit tile comparators; fine-tuning with the noise-calibrated surrogate under sampled noise; and a final noise-aware BN calibration step.
Baselines
The naive reference point is a binary-weight SNN trained in ideal software with the recipe of A Self-Ensemble Inspired Approach for Effective Training of Binary-Weight Spiking Neural Networks. It is deployed unchanged onto the same simulated differential crossbar with 1-bit tile sensing. This measures how badly the temporal trap hits 1-bit weights without any hardware-aware training. The file leaves that question open: we found no reported result on Lacuna for binary-weight SNNs on non-ideal crossbars.
The strongest existing hardware method is the ADC-less SNN of Hardware/Software Co-Design With ADC-Less In-Memory Computing Hardware for Spiking Neural Networks, with its 4-bit weights and three-stage training. It is the strongest competitor because it already removes the ADC, which is the dominant cost. Its reported CIFAR-10 accuracy drop under ADC-less operation is 1.2–2.4%. We will reimplement it in the same simulator with the same tile size, noise model, architecture and training budget, measured in epochs and in hyperparameter-search trials. Beating it means matching its accuracy within about one point while using binary-state cells and differential sensing. It also means showing equal or better accuracy retention as device variation grows, and lower NeuroSim area and energy.
For the temporal trap itself, the reference is Examining the Robustness of Spiking Neural Networks on Non-ideal Memristive Crossbars. It reports a surrogate-gradient VGG9 SNN on CIFAR-10 falling from 86.70% to 14.94% on non-ideal crossbars, against an ANN falling from 91.90% to 86.29%. Its noise-aware BN adaptation is the strongest cheap fix, and we will apply it to every system we compare. We will reproduce its VGG9 setting with matched non-ideality parameters. A 1-bit differential system that stays within a few points of its own software accuracy at the same non-ideality level would count as escaping the trap.
The static binary network with an 8-bit ADC, from Comprehensive Benchmarking of Binary Neural Networks on NVM Crossbar Architectures, is the area/energy reference. Input-encoding defenses such as GBO (Gradient-based Bit Encoding Optimization for Noise-Robust Binary Memristive Crossbar) recover 31.46% to 71.53% under severe noise. They serve as a non-spiking robustness reference.
Ablations isolate each ingredient. The first replaces the calibrated surrogate with a fixed fast-sigmoid surrogate while keeping identical noise injection. The second replaces differential sensing with a single-ended array plus a digital offset. The third thresholds the analog current after integration instead of before. The fourth restricts weights to binary values, dropping the zero state.
Why not just…?
Why not just inject noise during training and keep a standard surrogate? This is the real threat to the surrogate part of the claim. Noise injection alone may give most of the robustness, and the first ablation is designed to settle it. If the calibrated surrogate does not beat a tuned fixed-width surrogate under identical noise, we drop that claim. The proposal would then rest on the architectural question of where to binarize and on the fully 1-bit system. We expect a gap mainly in high-variance columns, where a fixed width gives miscalibrated gradients, so we will report results split by column activity.
Why not just convert a trained ANN to an SNN? Converted SNNs are more robust on crossbars (Examining the Robustness of Spiking Neural Networks on Non-ideal Memristive Crossbars). But they need long spike trains, while direct training of binary-weight SNNs works at $T=2$ (A Self-Ensemble Inspired Approach for Effective Training of Binary-Weight Spiking Neural Networks). We will include a converted baseline and compare accuracy against $T$ and against energy, not accuracy alone.
Why not just apply noise-aware BN calibration after deployment? It only partially recovers accuracy in the cited study, and it does nothing about the ADC. We apply it to every system, so any gain we report comes on top of it.
Why not just use a digital XNOR-popcount accelerator? Weight movement, not the logic, dominates binary inference energy (Hyperdrive). Analog only wins if the ADC is gone. The NeuroSim comparison against a digital binary-network baseline at matched accuracy decides this. If the 1-bit analog system does not come out ahead on energy per inference, we will report that.
Proposed Experiments
The main benchmark is CIFAR-10 with VGG9, chosen to match the temporal-trap study directly. We add a ResNet-style network to compare with the 95.04% software number for binary-weight SNNs, and DVS128 Gesture to test genuinely temporal input. We will sweep device variation, wire resistance, tile size from 64 to 256 rows, and $T$ from 1 to 8. We will report accuracy, NeuroSim area, energy per inference and latency.
The decisive check is a diagnostic of how error grows over time. For each time step we will log the gap between the noisy and ideal membrane potentials in three settings: analog-integrate-then-threshold, single-ended binarize-then-integrate, and differential binarize-then-integrate. The hypothesis predicts that the gap grows steadily with $T$ in the first setting and stays near the comparator flip rate in the third. If the third setting shows the same growth as the first, the central hypothesis is false.
A secondary check tests whether the zero state reduces IR drop enough to matter. We will compare ternary and binary weights at large tile sizes.
References
- Hardware for machine learning: Challenges and opportunities
- Comprehensive Benchmarking of Binary Neural Networks on NVM Crossbar Architectures
- Examining the Robustness of Spiking Neural Networks on Non-ideal Memristive Crossbars
- Low Power In-Memory Implementation of Ternary Neural Networks with Resistive RAM-Based Synapse
- Implementation of Ternary Weights With Resistive RAM Using a Single Sense Operation Per Synapse
- Hardware/Software co-design with ADC-Less In-memory Computing Hardware for Spiking Neural Networks
- A Self-Ensemble Inspired Approach for Effective Training of Binary-Weight Spiking Neural Networks
- SURGE: Surrogate Gradient Adaptation in Binary Neural Networks
- ReverB-SNN: Reversing Bit of the Weight and Activation for Spiking Neural Networks
- MemTorch: An Open-source Simulation Framework for Memristive Deep Learning Systems
- Gradient-based Bit Encoding Optimization for Noise-Robust Binary Memristive Crossbar
- Hyperdrive: A Multi-Chip Systolically Scalable Binary-Weight CNN Inference Engine