Browse

Lifelong Learning of Everyday Human Behaviors using Deep Neural Networks: Dual Memory Architecture and Incremental Moment Matching : 깊은 신경망 기반 일상 행동에 대한 평생 학습: 듀얼 메모리 아키텍쳐와 점진적 모멘트 매칭

Cited 0 time in Web of Science Cited 0 time in Scopus
Authors

이상우

Advisor
장병탁
Major
공과대학 전기·컴퓨터공학부
Issue Date
2018-08
Publisher
서울대학교 대학원
Description
학위논문 (박사)-- 서울대학교 대학원 : 공과대학 전기·컴퓨터공학부, 2018. 8. 장병탁.
Abstract
Learning from human behaviors in the real world is imperative for building human-aware intelligent systems.

We attempt to train a personalized context recognizer continuously in a wearable device by rapidly adapting deep neural networks from sensor data streams of user behaviors.

However, training deep neural networks from the data stream is challenging because learning new data through neural networks often results in loss of previously acquired information, referred to as catastrophic forgetting.

This catastrophic forgetting problem has been studied for nearly three decades but has not been solved yet because the mechanism of deep learning has been not understood enough.





We introduce two methods to deal with the catastrophic forgetting problem in deep neural networks. The first method is motivated by the concept of complementary learning systems (CLS) theory - contending that effective learning of the data stream in a lifetime requires complementary systems that comprise the neocortex and hippocampus in the human brain.

We propose a dual memory architecture (DMA), which trains two learning structures: one gradually acquires structured knowledge representations, and the other rapidly learns the specifics of individual experiences.

The ability of online learning is achieved by new techniques, such as weight transfer for the new deep module and hypernetworks for fast adaptation.



The second method is incremental moment matching (IMM) algorithm.

IMM incrementally matches the moment of the posterior distribution of neural networks, which is trained for the previous and the current task, respectively.

To make the search space of posterior parameter smooth, the IMM procedure is complemented by various transfer learning techniques including weight transfer, L2-norm of the old and the new parameter, and a variant of dropout with the old parameter.





To provide an insight into the success of two proposed lifelong learning methods, we introduce an insight by introducing two online learning methods of sum-product network, which is a kind of deep probabilistic graphical model.

We discuss online learning approaches which are valid in probabilistic models and explain how these approaches can be extended to the lifelong learning algorithms of deep neural networks.





We evaluate proposed DMA and IMM on two types of datasets: the various artificial benchmarks devised for evaluating the performance of lifelong learning and the lifelog dataset collected through the Google Glass for 46 days.

The experimental results show that our methods outperform comparative models in various experimental settings and that our trials for overcoming catastrophic forgetting are valuable and promising.
Language
English
URI
https://hdl.handle.net/10371/143174
Files in This Item:
Appears in Collections:
College of Engineering/Engineering Practice School (공과대학/대학원)Dept. of Computer Science and Engineering (컴퓨터공학부)Theses (Ph.D. / Sc.D._컴퓨터공학부)
  • mendeley

Items in S-Space are protected by copyright, with all rights reserved, unless otherwise indicated.

Browse