Jae-hun Shim

Bio: Jae-hun Shim is an academic researcher. The author has contributed to research in topics: Deep learning & Computational complexity theory. The author has co-authored 2 publications.

Neural Architecture Search for Light-weight Multi-touch Classification

Abstract: Multi-touch algorithm has proven its effectiveness in various touch applications. Recently, using convolutional neural network were shown to be effective in accurately classifying multi-touch inputs. However, multi-touch algorithm requires very low computational complexity and size due to the resource limitations of target hardwares. Neural Architecture Search (NAS) is currently being spotlighted as an effective solution to designing optimal light-weight networks. Especially, Once-for-all NAS shows remarkable performance in searching for optimal networks on various hardware platforms. In this paper, we propose an efficient OFA NAS based method for designing optimal CNN based multi-touch classifier with a new shrunk search space. The model searched by our proposed method shows outstanding performance despite its computational simplicity. Compared to MobileNetV2, our model has 7 times less MACs with only 1.25% accuracy drop.

Core-set Sampling for Efficient Neural Architecture Search.

Abstract: Neural architecture search (NAS), an important branch of automatic machine learning, has become an effective approach to automate the design of deep learning models. However, the major issue in NAS is how to reduce the large search time imposed by the heavy computational burden. While most recent approaches focus on pruning redundant sets or developing new search methodologies, this paper attempts to formulate the problem based on the data curation manner. Our key strategy is to search the architecture using summarized data distribution, i.e., core-set. Typically, many NAS algorithms separate searching and training stages, and the proposed core-set methodology is only used in search stage, thus their performance degradation can be minimized. In our experiments, we were able to save overall computational time from 30.8 hours to 3.5 hours, 8.8x reduction, on a single RTX 3090 GPU without sacrificing accuracy.