2024. 1. 18. 08:04ㆍ코딩 도구/LG Aimers
LG Aimers: AI전문가과정 4차
Module 6. 『딥러닝(Deep Learning)』
ㅇ 교수 : KAIST 주재걸 교수
ㅇ 학습목표
Neural Networks의 한 종류인 딥러닝(Deep Learning)에 대한 기본 개념과 대표적인 모형들의 학습원리를 배우게 됩니다.
이미지와 언어모델 학습을 위한 딥러닝 모델과 학습원리를 배우게 됩니다.
Part 1. Introduction to Deep Neural Networks
-Artificial Neural Networks
• A technology that imitates neurons existing in the human brain
-Deep Neural Network (DNN)
• DNN improves accuracy of AI technology by stacking neural network layers
"Non-deep" feedforward neural network는
input layer -> hidden layer -> output layer 순으로 구성되어있고,
Deep Neural Network (DNN)는 예를 들면 input layer -> hidden layer -> hidden layer -> hidden layer ->output layer 이런식으로 layer들이 입출력 사이에 있다.
-Reason Why Deep Learning has been Successful
3박자가 잘 갖춰짐.
Big Data, GPU Acceleration, Algorithm Improvements
-Applications of Deep Learning
Computer Vision
(Object Detection, Image Synthesis)
Natural Language Processing
(Machine Translation, Mail Classification)
Time-Series Analysis
(Stock Price Predction, Speech Recognition & Synthesis)
Reinforcement Learning
(AlphaGo, Atari Gane)
-Perceptron and Neural Networks
-What is Perceptron?
Perceptron
• One kind of neural networks
• Frank Rosenblatt devised in 1957
• Linear classifier
• Similar with structure of a neuron
-Multi-Layer Perceptron for XOR Gate
Is it possible to solve a XOR problem using a single layer perceptron?
→ No. Single layer perceptron can only solve linear problem. XOR problem is non-linear
-Multi-Layer Perceptron
But if we use two-layer perceptron, we can solve XOR problem
→ This model is called multi-layer perceptron
-Tensorflow Playground
https://playground.tensorflow.org/
-Forward Propagation
• 𝑎𝑗(i)
: “Activation” of the 𝑖-th unit in the 𝑗-th layer
• 𝑊 (j)
: “Weight Matrix” mapping from the 𝑗-th layer to the (𝑗 + 1)-th layer
-Sigmoid function
= Logistic function
-MNIST Dataset
MNIST
(Modified National Institute of Standards and Technology)
Handwritten digits from 0 to 9
• 55,000 training examples
• 10,000 testing examples
Each image has been preprocessed
• Digits are center-aligned
• Digit size is rescaled to similar size
• Each image has fixed size of 28 × 28
→ Real number matrix from 0.0 to 1.0
-MNIST Classification Model
+ 아래 사진도 참고!
-Softmax Layer (Softmax Classifier)
• Because of sigmoid outputs, Prediction ∈ 0,1 & Target ∈ 0,1
→ Upper limits exist on loss and gradient magnitude with MSE Loss
• In addition, a better output would be a sum-to-one probability vector over multiple possible classes.
-Logistic Regression as a Special Case of Softmax Classifier
'코딩 도구 > LG Aimers' 카테고리의 다른 글
LG Aimers 4기 Convolutional Neural Networks and Image Classification (5) | 2024.01.20 |
---|---|
LG Aimers 4기 Training Neural Networks (3) | 2024.01.19 |
LG Aimers 4기 Phase 1 온라인 교육 후기 (0) | 2024.01.17 |
LG Aimers 4기 인과추론의 다양한 연구 방향 (0) | 2024.01.17 |
LG Aimers 4기 인과추론 수행을 위한 기본 방법론 (0) | 2024.01.16 |