Summary of Pixel Embedding: Fully Quantized Convolutional Neural Network with Differentiable Lookup Table, by Hiroyuki Tokunaga et al.
Pixel Embedding: Fully Quantized Convolutional Neural Network with Differentiable Lookup Table
by Hiroyuki Tokunaga, Joel Nicholls, Daria Vazhenina, Atsunori Kanemura
First submitted to arxiv on: 23 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper proposes a novel approach called “pixel embedding” to fully quantize deep neural networks for hardware-friendly and energy-efficient deployment. By replacing each float-valued input pixel with a vector of quantized values using a lookup table, the authors aim to address the issue of representing originally high-bit input data with low-bit values. The proposed method utilizes backpropagation to train the lookup table or low-bit representation of pixels. Experimental results on ImageNet and CIFAR-100 demonstrate that pixel embedding reduces the top-5 error gap caused by quantizing floating points at the first layer to only 1% for ImageNet, and slightly over 1% for CIFAR-100. The method also achieves a speedup of over 1.7 times compared to floating point precision at the first layer during inference. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper is about finding a way to make deep neural networks work on devices with limited energy or computing power. Currently, these networks use a lot of energy and can be slow because they process lots of numbers (floats). The authors suggest a new method called “pixel embedding” that replaces each float value with a short code (like a dictionary) that tells the network what to do. This makes the network more efficient and faster on devices. They tested this method on two types of images, ImageNet and CIFAR-100, and it worked well, reducing errors by only 1%. It also made the network run 1.7 times faster. |
Keywords
» Artificial intelligence » Backpropagation » Embedding » Inference » Precision
