Leveraging Catastrophic Forgetting to Transfer Sub-Optimal to the Better-Sub-Optimal

Forgetting: A Nature of the Neural Learners Catastrophic forgetting is the main deficit of neural networks occurring in the primal properties of learning process. Neural networks cannot memorize in a long term if the things to memorize are absence in the training data coming newly. By the way, there are some needs to forget the past things in some area. For example, when we learned the '+' opera..

Thinking Is Feeling. Activeness Is an Illusion.

SELF-ORGANIZED INTELLIGENCE One thought just came across my brain. Viewing through a global view, the universe is just a sequence of perturbations of energy through time. An ultimate cause, which we'd like to call it 'Big Bang', triggered all of the interactions between adjacent particles which lead to complicated energy perturbation of the whole universe. Ivestigating on the great perturbations..

역전파 알고리즘의 생물학적 타당성

현재 인공신경망의 주류 학습 알고리즘은 오류 역전파 (backprop) 방식에 의존하고 있다. 해당 알고리즘의 성능이 탁월하다는 것은 다양한 분야에서 이미 입증되었다. 인공신경망은 뇌의 뉴런들의 연결 구조에서 영감을 받아 만든 구조이다. 즉 bio-inspired하다. 그런데 이 bio-inspired한 구조를 학습하는 주류 방법인 backprop 방법은 bio-inspired가 아니다. 하지만 bio-inspired가 아니어도 biologically plausible(생물학적으로 설명 가능)할 수도 있다. 과연 이 backprop 방법이 biologically plausible할까? 즉 실제 우리 뇌에서 실현이 가능할까? 결론은 간접적으로 타당하다. 한번 살펴보도록 하자. 본 글에서는 feedback ..

Deep Reinforcement Learning With Respect to Neural Networks

RL WITH NEURAL NETWORKS The best way to train neural networks is to iteratively give them a batch of samples randomly sampled (i.i.d. in ideal) from a sufficient amount of input-output data pairs. As a nonlinear global function approximator, neural networks rapidly finds a way to mapping certain input with matched output during training. Once this intput-output matching is accomplished, the gene..

Reinforcement Learning: Does 'Model-free' Really Means 'No model'?

MODEL-BASED VERSUS MODEL-FREE RL In reinforcement learning, by looking at the big picture, there typically exist two types of algorithm: model-free, or model-based. Their common goal is to make the agent find the best policy to gain highest cumulative rewards in the given environments. Stacking interactions with environment, the agent can learn how to act, approximating the expected future rewar..