목록Deep Learning (11)

Let's Run Jinyeah

Transfer Learning and Domain Adaptation

“Domain” and “Task” Domain relates to the feature space of a specific dataset and the marginal probability distribution of features Task relates to the label space of a dataset and an objective predictive function Transfer Learning goal is to transfer the knowledge learned from the Task(a) on Domain A to the Task(b) on Domain B common to update the last several layers of the pre-trained network ..
Deep Learning/Theory 2022. 8. 10. 20:43
Entropy and Cross-Entropy

Entropy는 일반적으로 불확실성을 나타내는 지표이다. 딥러닝에서는 이를 정보량으로 볼 수 있다. Rationale Information (𝑿=𝒙i) = -㏒ 𝑷(𝑿=𝒙i) The degree of information delivered by an event xi is low if 𝑷(𝑿=𝒙i) is close to 1 is high if 𝑷(𝑿=𝒙i) is close to 0 즉, 확률이 클수록 정보량이 적다 logarithm: -logx = x가 1에 가까울수록 작다 Expectation E [𝑿] = ∑ 𝒙·𝑷(𝑿=𝒙) Summation of (특정 outcome x 그 outcome이 나올 확률) Entropy H [𝑿] = E [Information(𝑿=𝒙)] = - ∑ 𝑷(𝑿=𝒙) · ㏒𝑷(..
Deep Learning/Theory 2022. 7. 31. 00:05
Normalization

What is the normalization formula used for? Normalization is useful in statistics for creating a common scale to compare data sets with very different values. Deep Learning view? 학습의 안정화: Gradient vanising/exploding 문제를 해결할 수 있음 학습시간의 단축: learning rate를 크게 할 수 있음 성능 개선: local optimum에서 빨리 빠져나올 수 있음 Min-Max Normalization Method normalization formula to [0,1] xnormalized = (x-xmin) / (xmax-xmin) i..
Deep Learning/Theory 2022. 6. 18. 17:31
Bayes Decision Theory

Prior Probability Probability derived by deductive reasoning calculated by existing information regarding a situation, and thus may vary depending the given situation ex) 특정 fish image가 돔(S1)에 속할 확률 Class-conditional Probability / Likelihood probability density function for X(feature), given that the corresponding class is Si also called Likelihood of Si with respect to X ex) 돔(S1)일 때, 몸 너비(X)가 ..
Deep Learning/Theory 2022. 5. 15. 17:00
Objective function/Loss function

To improve the performance of a Deep Learning model the goal is to the minimize or maximize the objective function. For regression, classification, and regression problems, the objective function is minimzing the difference between predictions and ground truths. Therefore, the objective function is also called loss functions. Regression Loss Functions Squared Error Loss Absolute Error Loss Huber..
Deep Learning/Theory 2022. 5. 10. 13:14
Integrating Learning and Planning

알파고에 쓰인 시뮬레이션 기반 Model-Based RL 방법에 대해 알아보겠다. Outline 1. Introduction What is Model? Model-Free RL vs Model-Based RL Advantages of Model-Based RL 2. Model-Based RL Model Learning Planning with a Model 3. Integrated Architectures Dyna 4. Simulation-Based Search Simple Monte-Carlo Search Monte-Carlo Tree Search(MCTS) Temporal-Difference Search Introduction What is Model? MDP including Transition p..
Deep Learning/Reinforcement Learning 2021. 8. 31. 18:48
Policy Gradient

이전 포스트에서 각 상태, 액션에 실제 가치에 근사값을 구하는 방법에 대해 알아보았습니다. Control(Policy Iteration) with Value approximation에서는 가치함수를 기반으로 정책을 그리디하게 정하기 때문에 결정론적인 정책을 가집니다. 이번 포스트에서는 정책함수를 기반으로 액션을 선택하여 stochastic 정책을 생성하는 Contorl with policy gradient 방법에 대해 알아보겠습니다. Outline 1. Introduction to Policy-Based Reinforcement Learning Value-Based vs Policy-Based vs Actor-Critic Advantages of Policy-Based RL Policy Search 2...
Deep Learning/Reinforcement Learning 2021. 8. 31. 01:56
Value Function Approximation

이전 포스트(Model-Free Control)에서 가치함수에 근거하여 액션을 선택하는 Sarsa와 Q러닝 기법에 대해 알아보았습니다. 하지만 상태 공간과 액션 공간이 매우 커서 밸류를 일일이 lookup table에 담지 못하는 상황에서 모든 상태, 액션에 대한 밸류를 어떻게 계산해야할까요? 이번 포스트에서는 뉴럴넷과 강화학습을 접목하여 이에 대한 해결책을 찾아보겠습니다. Outline 1. Incremental Methods Stochastic Gradient Descent Control with Value Function Approximation Incremental Contorl Algorithm 2. Batch Methods Stochastic Gradient Descent with Experi..
Deep Learning/Reinforcement Learning 2021. 8. 30. 23:06
Prev 1 2 Next