tinyML Research Symposium: AugViT: Improving Vision Transformer Training by Marrying Attention and..
Zhongzhi YU, PhD Student, EIC Lab at Georgia Institute of Technology
AugViT: Improving Vision Transformer Training by Marrying Attention and Data Augmentation
Despite the impressive accuracy of large-scale vision transformers (ViTs) across various tasks, it remains a challenge for small-scale ViTs (e.g., 1G inference floating points operations (FLOPs) as in LeViT) to significantly outperform state-of-the-art convolution neural networks (CNNs) in terms of the accuracy-efficiency trade-off, limiting their wider application, especially on resource-constrained devices. As analyzed in recent works, selecting an effective data augmentation technique can non-trivially improve the accuracy of small-scale ViTs. However, whether existing mainstream data augmentation techniques dedicated to CNNs are optimal for ViTs is still an open question. To this end, we propose a data augmentation framework called AugViT, which is dedicated to incorporating the key component in ViTs, i.e., self-attention, into data augmentation intensity to enable ViT’s outstanding performance across various devices. Specifically, motivated by ViT’s patch-based processing pipeline, our proposed AugViT integrates (1) a dedicated scheme for mapping the attention map in ViTs to the suggested augmentation intensity for each patch, (2) a simple but effective strategy of selecting the most effective attention map within ViTs to guide the aforementioned attention-aware data argumentation, and (3) a set of patch-level augmentation techniques that matches the patch-aware processing pipeline and enables the varying of augmentation intensities in each patch. Extensive experiments and ablation studies on two datasets and ten representative ViT models validate AugViT’s effectiveness in boosting ViTs’ performance, especially for small-scale ViTs, e.g., improving LeViT-128S’s accuracy from 76.6% to 77.1%, achieving a comparable accuracy to EfficientNet-B0 with 21.8% fewer inference FLOPs overhead on ImageNet dataset).
Видео tinyML Research Symposium: AugViT: Improving Vision Transformer Training by Marrying Attention and.. канала The tinyML Foundation
AugViT: Improving Vision Transformer Training by Marrying Attention and Data Augmentation
Despite the impressive accuracy of large-scale vision transformers (ViTs) across various tasks, it remains a challenge for small-scale ViTs (e.g., 1G inference floating points operations (FLOPs) as in LeViT) to significantly outperform state-of-the-art convolution neural networks (CNNs) in terms of the accuracy-efficiency trade-off, limiting their wider application, especially on resource-constrained devices. As analyzed in recent works, selecting an effective data augmentation technique can non-trivially improve the accuracy of small-scale ViTs. However, whether existing mainstream data augmentation techniques dedicated to CNNs are optimal for ViTs is still an open question. To this end, we propose a data augmentation framework called AugViT, which is dedicated to incorporating the key component in ViTs, i.e., self-attention, into data augmentation intensity to enable ViT’s outstanding performance across various devices. Specifically, motivated by ViT’s patch-based processing pipeline, our proposed AugViT integrates (1) a dedicated scheme for mapping the attention map in ViTs to the suggested augmentation intensity for each patch, (2) a simple but effective strategy of selecting the most effective attention map within ViTs to guide the aforementioned attention-aware data argumentation, and (3) a set of patch-level augmentation techniques that matches the patch-aware processing pipeline and enables the varying of augmentation intensities in each patch. Extensive experiments and ablation studies on two datasets and ten representative ViT models validate AugViT’s effectiveness in boosting ViTs’ performance, especially for small-scale ViTs, e.g., improving LeViT-128S’s accuracy from 76.6% to 77.1%, achieving a comparable accuracy to EfficientNet-B0 with 21.8% fewer inference FLOPs overhead on ImageNet dataset).
Видео tinyML Research Symposium: AugViT: Improving Vision Transformer Training by Marrying Attention and.. канала The tinyML Foundation
Показать
Комментарии отсутствуют
Информация о видео
Другие видео канала
![tinyML Talks Taiwan in Mandarin and English: Discovering tinyML](https://i.ytimg.com/vi/hPtRLygc4LE/default.jpg)
![tinyML Summit 2022: Sensing Applications as a Driver for TinyML Solutions](https://i.ytimg.com/vi/pUAAGVVYgLQ/default.jpg)
![tinyML Neuromorphic Engineering Forum - Sensors Session](https://i.ytimg.com/vi/Mt9X0ALdCpI/default.jpg)
![tinyML Vision Challenge - Himax & Edge Impulse](https://i.ytimg.com/vi/6tmCEzNSIas/default.jpg)
![tinyML Talks Chao Xu: Enabling Neural network at the low power edge: A neural network compiler...](https://i.ytimg.com/vi/Rs1yeTSuZLA/default.jpg)
![SensMACH 2020 Daniel Situnayake: Embedded machine learning in the real world](https://i.ytimg.com/vi/a67hWPT1NLE/default.jpg)
![tinyML Talks: Empowering the Edge: Practical Applications of Embedded Machine Learning on MCUs](https://i.ytimg.com/vi/tkqNS611cRc/default.jpg)
![tinyML Talks: Efficient AI for Wildlife Conservation](https://i.ytimg.com/vi/FfvcZEMn2l0/default.jpg)
![tinyML Research Symposium 2022: Towards Agile Design of Neural Processing Units with Chisel](https://i.ytimg.com/vi/xlP1xdKRrqc/default.jpg)
![tinyML Talks Phoenix: Novel Device and Materials in Emerging Memory for Neuromorphic Computing](https://i.ytimg.com/vi/_apkQF1ZL6A/default.jpg)
![tinyML Talks - Phoenix meetup: Analog TinyML for health management using intelligent wearables](https://i.ytimg.com/vi/bCzg8y6aRi8/default.jpg)
![tinyML Talks India: Single Lead ECG Classification On Wearable and Implantable Devices](https://i.ytimg.com/vi/uHywaYleCtA/default.jpg)
![tinyML Summit 2023:Personal Computing devices use-case and applications enabled by Smart Sensors](https://i.ytimg.com/vi/9hvz6ZB5G8A/default.jpg)
![tinyML Talks: From the lab to the edge: Post-Training Compression](https://i.ytimg.com/vi/Ada9Tq8JAX8/default.jpg)
![tinyML Talks: State of Hardware & Software Ecosystem for Low-Power ML Applications on RISC-V](https://i.ytimg.com/vi/Rcbrc2rnXlk/default.jpg)
![tinyML Talks: Meetup Italy with small-medium industries](https://i.ytimg.com/vi/sAmRSm-tdd4/default.jpg)
![tinyML Hackathon Challenge 2023 - Infineon XENSIV 60GHz Radar Sensor and devkit explanation](https://i.ytimg.com/vi/yL6f61MKzFo/default.jpg)
![tinyML Auto ML Tutorial with Qeexo](https://i.ytimg.com/vi/qo0JTM6gaIE/default.jpg)
![tinyML On Device Learning Forum - Warren Gross: On-Device Learning For Natural Language Processing..](https://i.ytimg.com/vi/ERLFRluwRjA/default.jpg)
![EMEA 2021 tiny Talks: Building Heterogeneous TinyML Pipelines](https://i.ytimg.com/vi/p-Rtnvj4L4I/default.jpg)
![tinyML EMEA 2022- Eran Treister: Wavelet Feature Maps Compression for Image-to-Image CNNs](https://i.ytimg.com/vi/fPmvwecx7TY/default.jpg)