Increasingly Packing Multiple Facial-Informatics Modules in A Unified Deep-Learning Model via Lifelong Learning

By Steven C. Y. Hung, Jia-Hong Lee, Timmy S. T. Wan, Chien-Hung Chen, Yi-Ming Chan, Chu-Song Chen. 2019

47 citations recorded by the source corpus.

Simultaneously running multiple modules is a key requirement for a smart multimedia system for facial applications including face recognition, facial expression understanding, and gender identification. To effectively integrate them, a continual learning approach to learn new tasks without forgetting is introduced. Unlike previous methods growing monotonically in size, our approach maintains the compactness in continual learning. The proposed packing-and-expanding method is effective and easy to implement, which can iteratively shrink and enlarge the model to integrate new functions. Our integrated multitask model can achieve similar accuracy with only 39.9% of the original size.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Meta & Continual Learning, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Prashant Bhat
2. Bahram Zonooz
3. Elahe Arani
4. Qiang Liu
5. Bo Liu
6. Qian Liu
7. Peter Stone
8. Lemeng Wu

In Computer Vision and Pattern Recognition

1. Bahram Zonooz
2. Elahe Arani
3. Prashant Bhat
4. Qian Liu
5. Bo Liu
6. Lemeng Wu
7. Peter Stone
8. Chien-Hung Chen

In Meta & Continual Learning

1. Prashant Bhat
2. Bahram Zonooz
3. Elahe Arani
4. Peter Stone
5. Bo Liu
6. Qian Liu
7. Qiang Liu
8. Lemeng Wu

In Signal Processing

1. Chien-Hung Chen
2. Peter Stone
3. Chu-Song Chen
4. Steven C. Y. Hung
5. Yi-Ming Chan
6. Bahram Zonooz
7. Qiang Liu
8. Elahe Arani

Citations in the corpus

• Task-Aware Information Routing from Common Representation Space in Lifelong Learning (2023, methodology): “Comparison with evolving architectures: Similar to progressive networks (e.g. PNN, CPG, PAE), dynamic sparse networks (e.g. CLIP, NISPA, PackNet) reduce task interference by learning a nonoverlapping task-specific sparse architecture within a fixed capacity model.”
• Compacting, Picking and Growing for Unforgetting Continual Learning (2019, methodology): “Finally, as shown in the Expand (Exp.) field in Table 1, PAE grows the model and uses 2 times of weights for the 20 tasks.”
• Continual Learning and Private Unlearning (2022, background): “1) Dynamic architecture approaches study how to carefully and gradually expand the learning model to incorporate the learning of new knowledge (Rusu et al., 2016; Yoon et al., 2017; Mallya et al., 2018; Rosenfeld & Tsotsos, 2018; Mallya & Lazebnik, 2018; Hung et al., 2019b;a; Wu et al., 2020a).”
• Firefly Neural Architecture Descent: a General Approach for Growing Neural Networks (2021, methodology): “One branch of approaches in CL consider dynamically growing networks to avoid catastrophic forgetting (Rusu et al., 2016; Li & Hoiem, 2017; Yoon et al., 2017; Li et al., 2019; Hung et al., 2019a).”