Spike-Train Level Direct Feedback Alignment: Sidestepping Backpropagation for On-Chip Training of Spiking Neural Nets

Jeong-Jun Lee, Renqian Zhang, Wenrui Zhang, Yu Liu, Peng LiFrontiers in Neuroscience 2020 · DOI

Abstract

Spiking neural networks (SNNs) present a promising computing model for bio-plausible information processing and event-driven, ultra-low power neuromorphic hardware. However, training SNNs to reach the performance levels of conventional deep artificial neural networks (ANNs) is challenging due to the non-differentiable nature of spike activities. This paper introduces Spike-Train Level Direct Feedback Alignment (ST-DFA), a hardware-friendly and biologically plausible alternative to backpropagation (BP). By employing algorithm-hardware co-optimization and efficient online neural signal computation, the proposed ST-DFA achieves competitive classification accuracy (96.27% on MNIST and 84.88% on TI46) while significantly reducing hardware overhead. Compared to state-of-the-art BP implementations, ST-DFA reduces functional resources by 76.7% and backward training latency by 31.6% on a Xilinx ZC706 FPGA.

Introduction

Spiking Neural Networks (SNNs) are often hailed as the "third generation" of neural networks, prized for their ability to mimic the brain's energy efficiency through event-driven processing. Despite their potential, training deep SNNs remains a significant hurdle. The standard training algorithm, Backpropagation (BP), relies on gradients that are difficult to compute for discrete, non-differentiable spikes. Furthermore, BP requires "weight transport"—the use of perfectly symmetric weights in both forward and backward paths—which is biologically implausible and hardware-intensive for on-chip training.

Methodology: From Backpropagation to Direct Feedback Alignment

To overcome the limitations of BP, this work leverages Direct Feedback Alignment (DFA). In traditional BP, error signals are propagated backward through each layer sequentially, requiring the transpose of the forward weights. In contrast, DFA bypasses this "bucket brigade" by sending the error signal from the output layer directly to each hidden layer through fixed, random feedback weights.

Comparison of (A) Backpropagation with symmetric weights and (B) Direct Feedback Alignment with random feedback weights.

This approach eliminates the need for weight symmetry and allows for parallelized error propagation, which is significantly more efficient for hardware implementation.

The ST-DFA Algorithm

The proposed Spike-Train Level Direct Feedback Alignment (ST-DFA) extends DFA to the temporal domain of SNNs. Unlike methods that calculate errors at every millisecond time-step (which leads to a memory explosion), ST-DFA operates on the overall spike train.

Architecture of the proposed Spike-Train Level Direct Feedback Alignment (ST-DFA).

As shown in the figure, the error signal is multiplied by random feedback matrices ($B^k$) to generate the local error ($\delta^k$) for each hidden layer. This mechanism allows the network to learn effectively even though the feedback signals are random, as the hidden layers adapt to make these signals useful for global error reduction.

Hardware Implementation and Co-optimization

To further optimize the design for FPGA deployment, the authors introduced several key techniques:

  • ST-DFA-2: Random feedback weights are restricted to signed powers of two. This allows the hardware to replace complex multipliers with simple bit-shift operations, drastically reducing resource consumption.
  • Online S-PSP Calculation: By accumulating Post-Synaptic Potentials (S-PSP) during the forward pass, the system avoids storing the entire history of spike timings, saving significant memory.
  • Concurrency: The direct nature of the feedback allows the backward pass to occur in parallel across layers, reducing training latency.

Experimental Results

The ST-DFA system was implemented on a Xilinx ZC706 FPGA and evaluated against the state-of-the-art HM2-BP algorithm.

Classification Performance

The network achieved highly competitive results across multiple datasets:

  • MNIST: 96.27% accuracy (using 14x14 reduced resolution).
  • N-MNIST: Demonstrated robust performance on neuromorphic event-based data.
  • TI46 Speech Corpus: 84.88% accuracy on a challenging 16-speaker dataset, proving its efficacy for temporal audio tasks.

Hardware Efficiency

The primary advantage of ST-DFA lies in its efficiency. Compared to the HM2-BP baseline:

  • Resource Reduction: DSP usage was slashed by 76.7%.
  • Latency Improvement: Backward training latency was reduced by 31.6%.
  • Energy Consumption: Training energy ranged from 0.45 mJ to 1.77 mJ per example, making it suitable for edge devices.

Conclusion

ST-DFA represents a significant step toward efficient, on-chip training for SNNs. By sidestepping the complexities of backpropagation and embracing biologically inspired direct feedback, the authors have demonstrated a path toward high-performance neuromorphic hardware that is both power-efficient and fast. Future work will focus on scaling these architectures to deeper networks and more complex real-world applications.

Summary updated

Abstract

Spiking neural networks (SNNs) present a promising computing model and enable bio-plausible information processing and event-driven based ultra-low power neuromorphic hardware. However, training SNNs to reach the same performances of conventional deep artificial neural networks (ANNs), particularly with error backpropagation (BP) algorithms, poses a significant challenge due to inherent complex dynamics and non-differentiable spike activities of spiking neurons. In this paper, we present the first study on realizing competitive spike-train level backpropagation (BP) like algorithms to enable on-chip training of SNNs. We propose a novel spike-train level direct feedback alignment (ST-DFA) algorithm, which is much more bio-plausible and hardware friendly than BP. Algorithm and hardware co-optimization and efficient online neural signal computation are explored for on-chip implementation of ST-DFA. On the Xilinx ZC706 FPGA board, the proposed hardware-efficient ST-DFA shows excellent performance vs. overhead tradeoffs for real-world speech and image classification applications. SNN neural processors with on-chip ST-DFA training show competitive classification accuracy of 96.27% for the MNIST dataset with 4× input resolution reduction and 84.88% for the challenging 16-speaker TI46 speech corpus, respectively. Compared to the hardware implementation of the state-of-the-art BP algorithm HM2-BP, the design of the proposed ST-DFA reduces functional resources by 76.7% and backward training latency by 31.6% while gracefully trading off classification performance.
BibTeX
@inproceedings{lee2020spike,
  author = {Jeong-Jun Lee and Renqian Zhang and Wenrui Zhang and Yu Liu and Peng Li},
  title = {{Spike-Train Level Direct Feedback Alignment: Sidestepping Backpropagation for On-Chip Training of Spiking Neural Nets}},
  booktitle = {Frontiers in Neuroscience},
  year = {2020},
  doi = {10.3389/fnins.2020.00143},
}
Figure