Journal: iScience

Article Title: System-level FPGA validation of a trainable and robust multiplier-free spiking neural network

doi: 10.1016/j.isci.2026.115985

Figure Lengend Snippet: SNN architecture and TTFS encoding module implemented on FPGA (A) Overall FPGA-based implementation of the SNN system. An on-chip controller orchestrates the execution of inference and training submodules under external commands. Dashed arrows denote control signals, while solid arrows indicate data flow between functional modules. (B) Hardware structure of the TTFS encoder. Pixel inputs are compared with a global countdown counter to generate spike signals, and a range decoder extracts the relative spike timing to produce positional outputs Pos i .

Article Snippet: By implementing a sensitive-noise threshold (SNT) mechanism and a limited remote supervised method (LReSuMe) on a Xilinx ZCU104 FPGA platform, the proposed design achieves high throughput, low power consumption, and preserved robustness to the noise.

Techniques: Control, Functional Assay

Journal: iScience

Article Title: System-level FPGA validation of a trainable and robust multiplier-free spiking neural network

doi: 10.1016/j.isci.2026.115985

Figure Lengend Snippet: FPGA-based implementation of the neuron computation, weight storage, decay, and training modules (A) Neuron computation module implemented on FPGA, illustrating the datapath from TTFS-encoded spike trains to the final output score. The module consists of an adder tree for spike accumulation, a membrane potential updating unit, and an output layer that detects threshold crossings based on SNT and updates the output score accordingly. (B) Organization of 4,000 synaptic weights across 20 on-chip block RAMs (BRAMs). Each BRAM stores weights associated with a single output neuron, enabling parallel access to all corresponding synaptic weights during inference and training by fixing the neuron address. (C) Multiplier-free circuit for realizing a fixed decay coefficient D = 0.75 using a shift-and-add structure. Partial results generated by binary shifts are combined through addition to obtain the scaled output, representing a standard hardware-efficient implementation for constant multiplication in FPGA designs. (D) FPGA-based training module implementing fixed-point weight updates using an LReSuMe-based learning rule. The module includes a shift-and-add learning multiplication unit and a zero replacement unit (ZRU) to mitigate quantization-induced null updates.

Article Snippet: By implementing a sensitive-noise threshold (SNT) mechanism and a limited remote supervised method (LReSuMe) on a Xilinx ZCU104 FPGA platform, the proposed design achieves high throughput, low power consumption, and preserved robustness to the noise.

Techniques: Membrane, Blocking Assay, Generated

Journal: iScience

Article Title: System-level FPGA validation of a trainable and robust multiplier-free spiking neural network

doi: 10.1016/j.isci.2026.115985

Figure Lengend Snippet: Noise modeling, training accuracy comparison, and fixed-point saturation behavior (A) Illustration of impulse noise modeling in the MNIST dataset. From left to right: original image, image corrupted with random impulse noise (random positions and random values), and image corrupted with impulse noise (random positions with pixel values replaced by either 0 or 255). (B) Training accuracy comparison among six different network configurations and quantization settings on the clean MNIST training dataset. Accuracy is reported per 100-image chunk. (C) Evolution of final output scores under different fixed-point formats during training. Results are shown for FPGA Q6.10 (left) and Q6.26 (right) implementations. The x axis denotes the uniformly sampled time step index selected from the first 180 training chunks of the hardware simulation, and the y axis represents the final output score S o j . At each sampled time step, the maximum, mean, and standard deviation of S o j across all output neurons are computed and visualized.

Article Snippet: By implementing a sensitive-noise threshold (SNT) mechanism and a limited remote supervised method (LReSuMe) on a Xilinx ZCU104 FPGA platform, the proposed design achieves high throughput, low power consumption, and preserved robustness to the noise.

Techniques: Comparison, Standard Deviation