Timezone: »
Efficient on-device learning requires a small memory footprint at training time to fit the tight memory constraint. Existing work solves this problem by reducing the number of trainable parameters. However, this doesn't directly translate to memory saving since the major bottleneck is the activations, not parameters. In this work, we present Tiny-Transfer-Learning (TinyTL) for memory-efficient on-device learning. TinyTL freezes the weights while only learns the memory-efficient bias modules, thus no need to store the intermediate activations. To maintain the adaptation capacity, we introduce a new memory-efficient bias module, the lite residual module, to refine the feature extractor by learning small residual feature maps adding only 3.8% memory overhead. Extensive experiments show that TinyTL significantly saves the memory (up to 6.5x) with little accuracy loss compared to fine-tuning the full network. Compared to fine-tuning the last layer, TinyTL provides significant accuracy improvements (up to 33.8%) with little memory overhead. Furthermore, combined with feature extractor adaptation, TinyTL provides 7.5-12.9x memory saving without sacrificing accuracy compared to fine-tuning the full Inception-V3. Code is released at https://github.com/mit-han-lab/tinyML/tree/master/tinyTL.
Author Information
Han Cai (Massachusetts Institute of Technology)
Chuang Gan (MIT-IBM Watson AI Lab)
Ligeng Zhu (MIT)
Song Han (MIT)
More from the Same Authors
-
2020 Poster: MCUNet: Tiny Deep Learning on IoT Devices »
Ji Lin · Wei-Ming Chen · Yujun Lin · john cohn · Chuang Gan · Song Han -
2020 Spotlight: MCUNet: Tiny Deep Learning on IoT Devices »
Ji Lin · Wei-Ming Chen · Yujun Lin · john cohn · Chuang Gan · Song Han -
2020 Poster: Differentiable Augmentation for Data-Efficient GAN Training »
Shengyu Zhao · Zhijian Liu · Ji Lin · Jun-Yan Zhu · Song Han -
2019 Poster: Park: An Open Platform for Learning-Augmented Computer Systems »
Hongzi Mao · Parimarjan Negi · Akshay Narayan · Hanrui Wang · Jiacheng Yang · Haonan Wang · Ryan Marcus · Ravichandra Addanki · Mehrdad Khani Shirkoohi · Songtao He · Vikram Nathan · Frank Cangialosi · Shaileshh Venkatakrishnan · Wei-Hung Weng · Song Han · Tim Kraska · Dr.Mohammad Alizadeh -
2019 Poster: Deep Leakage from Gradients »
Ligeng Zhu · Zhijian Liu · Song Han -
2019 Poster: Cross-channel Communication Networks »
Jianwei Yang · Zhile Ren · Chuang Gan · Hongyuan Zhu · Devi Parikh -
2019 Poster: Visual Concept-Metaconcept Learning »
Chi Han · Jiayuan Mao · Chuang Gan · Josh Tenenbaum · Jiajun Wu -
2019 Poster: Imitation Learning from Observations by Minimizing Inverse Dynamics Disagreement »
Chao Yang · Xiaojian Ma · Wenbing Huang · Fuchun Sun · Huaping Liu · Junzhou Huang · Chuang Gan -
2019 Spotlight: Imitation Learning from Observations by Minimizing Inverse Dynamics Disagreement »
Chao Yang · Xiaojian Ma · Wenbing Huang · Fuchun Sun · Huaping Liu · Junzhou Huang · Chuang Gan -
2019 Poster: Point-Voxel CNN for Efficient 3D Deep Learning »
Zhijian Liu · Haotian Tang · Yujun Lin · Song Han -
2019 Spotlight: Point-Voxel CNN for Efficient 3D Deep Learning »
Zhijian Liu · Haotian Tang · Yujun Lin · Song Han -
2018 Poster: Weakly Supervised Dense Event Captioning in Videos »
Xin Wang · Wenbing Huang · Chuang Gan · Jingdong Wang · Wenwu Zhu · Junzhou Huang -
2018 Poster: Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding »
Kexin Yi · Jiajun Wu · Chuang Gan · Antonio Torralba · Pushmeet Kohli · Josh Tenenbaum -
2018 Spotlight: Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding »
Kexin Yi · Jiajun Wu · Chuang Gan · Antonio Torralba · Pushmeet Kohli · Josh Tenenbaum
