MINT: Multiplier-less INTeger Quantization for Energy Efficient Spiking Neural Networks

Ruokai Yin, Yuhang Li, Abhishek Moitra, Priyadarshini Panda

TL;DR

MINT (Multiplier-less INTeger quantization) is a novel scheme that slashes the energy and memory demands of Spiking Neural Networks (SNNs) by quantizing both weights and membrane potentials to ultra-low precision (as low as 2-bit). By sharing a single scaling factor across these components, MINT eliminates the need for power-hungry 32-bit multipliers during inference, enabling high-performance AI on resource-constrained edge devices.

What Problem Does This Paper Address?

Spiking Neural Networks (SNNs) are often touted as the "green" alternative to traditional Artificial Neural Networks (ANNs) because they process information using sparse, binary spikes. However, two major hurdles prevent them from being truly efficient on edge hardware:

  1. The Membrane Potential Bottleneck: While many researchers focus on compressing model weights, they often ignore the "membrane potential"—the internal state each neuron must store to track temporal data. As weight precision drops, these potentials start to dominate the memory footprint, accounting for up to 98% of total memory in mini-batch processing.
  2. The Scaling Factor Tax: Standard "Uniform Quantization" (UQ) requires multiplying quantized values by a high-precision (32-bit) scaling factor to maintain accuracy. In hardware accelerators like SpinalFlow, these scaling units can consume nearly half (46.3%) of the total system area, effectively negating the benefits of low-precision arithmetic.

Methods

The authors introduce MINT, which rethinks the fundamental Leaky-Integrate-and-Fire (LIF) neuron equations to remove hardware complexity:

  • Unified Low Precision: MINT quantizes both weights ($w$) and membrane potentials ($u$) to the same extremely low bit-width (e.g., 2-bit or 4-bit).
  • Scaling Factor Sharing: By forcing the scaling factors for weights and potentials to be identical ($\alpha_w = \alpha_u = \alpha$), the authors mathematically transform the LIF update equation. During inference, the scaling factor $\alpha$ cancels out of the "update" stage and is absorbed into a pre-computed integer firing threshold ($\theta$).
  • Multiplier-less Inference: Because the scaling factor is handled offline, the hardware only needs to perform integer additions and bit-shifts (for leakage). The complex 32-bit multipliers usually required for scaling are completely removed from the datapath.
  • Learnable Quantization: To prevent accuracy loss at 2-bit precision, the scaling factor $\alpha$ is treated as a learnable parameter during training, allowing the network to adapt its dynamic range to the quantization constraints.

Datasets

The researchers evaluated MINT across a variety of architectures (VGG-9, VGG-16, ResNet-19) using:

  • Static Image Datasets: CIFAR-10 and TinyImageNet.
  • Neuromorphic (Event-based) Dataset: CIFAR-10 DVS, which captures temporal spikes directly from a neuromorphic camera.

Key Insights and Ideas

  • Membrane Potential is the Real Memory Hog: The paper highlights that in real-world deployment scenarios using mini-batches, the memory cost of storing neuron states is far more critical than the cost of storing weights.
  • Multipliers are Overkill: Most SNN accelerators waste area on 32-bit multipliers just to apply a single scaling factor per layer. MINT proves that with clever mathematical rearrangement, you can achieve near-identical accuracy using only integer accumulators and comparators.
  • Regularization via Quantization: Interestingly, in some cases (like VGG-16 on TinyImageNet), the 8-bit MINT model actually outperformed the full-precision baseline, suggesting that ultra-low precision can act as a form of regularization that prevents overfitting.

Key Results

  • High Accuracy at Low Bits: A 2-bit MINT VGG-16 model achieved 90.6% accuracy on CIFAR-10, representing only a 0.6% drop from the full-precision (32-bit) version.
  • Massive Memory Savings: MINT achieved a 93.8% reduction in total memory footprint compared to full-precision models.
  • Superior Energy Efficiency: Compared to "vanilla" uniform quantization, MINT reduced computation energy by 90% because it removed the need for 32-bit scaling multiplications.
  • Hardware Area Reduction: When implemented on a systolic array accelerator, MINT reduced the Processing Element (PE) area by roughly 85% compared to standard quantization methods.

Limitations and Future Work

  • Uniform Bit-width: The current work uses the same bit-width for all layers. Future iterations could explore "mixed-precision" MINT, where sensitive layers keep higher precision while others drop to 1-bit to further optimize the energy-accuracy tradeoff.
  • Threshold Storage: While MINT removes multipliers, it requires storing a unique integer threshold $\theta$ for each layer. While this overhead is small (less than 6 bits per layer), it is a factor to consider in extremely scaled systems.
  • Scaling to Larger Datasets: While results on TinyImageNet are promising, testing MINT on industry-standard benchmarks like ImageNet-1K would further validate its robustness for large-scale applications.

Summary updated

Abstract

We propose Multiplier-less INTeger (MINT) quantization, a uniform quantization scheme that efficiently compresses weights and membrane potentials in spiking neural networks (SNNs). Unlike previous SNN quantization methods, MINT quantizes memory-intensive membrane potentials to an extremely low precision (2-bit), significantly reducing the memory footprint. MINT also shares the quantization scaling factor between weights and membrane potentials, eliminating the need for multipliers required in conventional uniform quantization. Experimental results show that our method matches the accuracy of full-precision models and other state-of-the-art SNN quantization techniques while surpassing them in memory footprint reduction and hardware cost efficiency at deployment. For example, 2-bit MINT VGG-16 achieves 90.6% accuracy on CIFAR-10, with roughly 93.8% reduction in memory footprint from the full-precision model and 90% reduction in computation energy compared to vanilla uniform quantization at deployment. 1 1 Code is available at https://github.com/Intelligent-Computing-Lab-Yale/MINT-Quantization
BibTeX
@inproceedings{yin2024mint,
  author = {Ruokai Yin and Yuhang Li and Abhishek Moitra and Priyadarshini Panda},
  title = {{MINT: Multiplier-less INTeger Quantization for Energy Efficient Spiking Neural Networks}},
  booktitle = {ASPDAC},
  year = {2024},
  doi = {10.1109/asp-dac58780.2024.10473825},
}
Figure