Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
102 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

LSQ+: Improving low-bit quantization through learnable offsets and better initialization (2004.09576v1)

Published 20 Apr 2020 in cs.CV, cs.LG, and stat.ML

Abstract: Unlike ReLU, newer activation functions (like Swish, H-swish, Mish) that are frequently employed in popular efficient architectures can also result in negative activation values, with skewed positive and negative ranges. Typical learnable quantization schemes [PACT, LSQ] assume unsigned quantization for activations and quantize all negative activations to zero which leads to significant loss in performance. Naively using signed quantization to accommodate these negative values requires an extra sign bit which is expensive for low-bit (2-, 3-, 4-bit) quantization. To solve this problem, we propose LSQ+, a natural extension of LSQ, wherein we introduce a general asymmetric quantization scheme with trainable scale and offset parameters that can learn to accommodate the negative activations. Gradient-based learnable quantization schemes also commonly suffer from high instability or variance in the final training performance, hence requiring a great deal of hyper-parameter tuning to reach a satisfactory performance. LSQ+ alleviates this problem by using an MSE-based initialization scheme for the quantization parameters. We show that this initialization leads to significantly lower variance in final performance across multiple training runs. Overall, LSQ+ shows state-of-the-art results for EfficientNet and MixNet and also significantly outperforms LSQ for low-bit quantization of neural nets with Swish activations (e.g.: 1.8% gain with W4A4 quantization and upto 5.6% gain with W2A2 quantization of EfficientNet-B0 on ImageNet dataset). To the best of our knowledge, ours is the first work to quantize such architectures to extremely low bit-widths.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (5)
  1. Yash Bhalgat (23 papers)
  2. Jinwon Lee (21 papers)
  3. Markus Nagel (33 papers)
  4. Tijmen Blankevoort (37 papers)
  5. Nojun Kwak (116 papers)
Citations (192)

Summary

An Analysis of LSQ+: Improvised Low-Bit Quantization Through Learnable Offsets and Initialization

The paper "LSQ+: Improving Low-Bit Quantization through Learnable Offsets and Better Initialization" presents a novel approach to address the shortcomings of existing quantization methods in deep neural networks, specifically focusing on low-bit arithmetic. The authors propose a refinement to the Learned Step Size Quantization (LSQ) framework by introducing LSQ+, which incorporates a learnable asymmetric quantization scheme alongside an MSE-based initialization technique for quantization parameters.

Key Contributions

The authors identify that conventional quantization methods, predominantly unsigned quantization, discard negative activations, which leads to significant performance losses in networks employing non-ReLU activations like Swish and Mish. To ameliorate this, LSQ+ employs an asymmetric quantization mechanism that features trainable scale and offset parameters. This enables retention and effective utilization of negative activation values without necessitating additional sign bits.

LSQ+ Highlights:

  • Asymmetric Activation Quantization: The scale and offset parameters are learned dynamically across the training phase, facilitating accommodation of skewed activation ranges that are characteristic of Swish and similar functions.
  • Enhanced Initialization: Recognizing the variance and instability issues associated with quantization-aware training, an MSE-based initialization is incorporated. This yields more stable performance across different runs, especially prominent in lower bit-width scenarios.

Empirical Evaluation

Experiments underscore LSQ+'s efficacy by demonstrating state-of-the-art performance in benchmark architectures such as EfficientNet and MixNet. The methodology results in notable accuracy improvements over baseline LSQ under configurations of W4A4, W3A3, and notably W2A2 quantization. For instance, with EfficientNet-B0, LSQ+ achieved up to a 5.6% enhancement at W2A2 levels compared to the standard LSQ.

When applied to ReLU-based architectures like ResNet18, LSQ+ maintains competitive performance, thus validating its cross-framework applicability. Results show that configurations with learned offsets (Configurations 3 and 4) consistently outperform those without, underscoring the utility of asymmetric quantization in modern networks with diversified activation functions.

Theoretical and Practical Implications

From a theoretical perspective, LSQ+ challenges the adequacy of symmetric and unsigned quantization methods in capturing complete activation functions, suggesting that incorporating learnable offsets could become a standardized practice in emerging architectures. Practically, the ability to achieve high accuracy with extremely low-bits holds significant implications for deployment in edge computing environments where both computation and power resources are severely limited.

Future exploration could involve the integration of LSQ+ with automated neural architecture search frameworks or expanding its application to even more granular quantization levels like 1-bit schemes. Further, the hardware-specific optimizations for LSQ+'s asymmetric approach remain an open avenue for subsequent research and development.

Conclusion

LSQ+ presents a robust solution to the prevalent challenges faced in low-bit quantization of neural networks. By emphasizing the learnability of quantization parameters and leveraging an initialization strategy that mitigates performance variability, it establishes new benchmarks across a variety of models and activation functions. The proposed methods have broad implications for advancing neural network efficiency in resource-constrained settings, marking a substantive step forward in quantization-aware training and deployment.