Summary of Transaxx: Efficient Transformers with Approximate Computing, by Dimitrios Danopoulos et al.
TransAxx: Efficient Transformers with Approximate Computing
by Dimitrios Danopoulos, Georgios Zervakis, Dimitrios Soudris, Jörg Henkel
First submitted to arxiv on: 12 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Hardware Architecture (cs.AR)
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 A recent introduction of Vision Transformer (ViT) models has shown promising results as a competitive alternative to Convolutional Neural Networks (CNNs). However, their high computational requirements limit their practical applicability on low-power devices. To address this issue, the state-of-the-art employs approximate multipliers for DNN accelerators, but no prior research has explored their use on ViT models. This work proposes TransAxx, a framework that enables fast support for approximate arithmetic to evaluate the impact of approximate computing on DNNs like ViT models. The authors analyze the sensitivity of transformer models on ImageNet dataset to approximate multiplications and perform approximate-aware finetuning to regain accuracy. Additionally, they propose a methodology to generate approximate accelerators for ViT models using Monte Carlo Tree Search (MCTS) algorithm. The evaluation demonstrates the efficacy of their approach in achieving significant trade-offs between accuracy and power. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Vision Transformer models are super smart AI systems that can rival Convolutional Neural Networks. But, they need a lot of computer power to work well, which makes them not very practical for things like smartphones or smart home devices. To solve this problem, researchers have been trying to make DNN accelerators more efficient by using approximate multipliers. This paper proposes a new way to do this specifically for Vision Transformer models. They tested their approach on the ImageNet dataset and found that it can work well even with some inaccuracies. They also came up with a new method to generate special hardware for these AI systems that uses a combination of clever algorithms and human expertise. |
Keywords
* Artificial intelligence * Transformer * Vision transformer * Vit
