neural networks - seoul national...

Post on 12-May-2020

2 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Neural Networks이경호

Movement research lab

Introduction

• Neural network basics

• Convolutional Neural Networks

• Recurrent Neural Networks

Supervised Learning

Linear Classification

ax + by + c > 0

Multiple classifier

Perceptron

Feed-forward Neural Network

Feed-forward Neural Network

Feed-forward Neural Network

𝑊0 𝑊1𝑊2

( n = 3, h = 4, m = 1 )

Loss function

Regression Regularization

Backpropagation

f(x,y,z) = (x + y)*z

Backpropagation

Optimization

• Gradient Descent

• Adam optimizer

• Mini-batch

Applications

• Motion generation• Predict next frame

• Deep RL• Represents policy, value function

Image Classification

• Multi-dimensional structure( width x height x color )

• High-dimensional data ( 128 x 128 x 3 = 49152 )

- Car- Airplane- Ship- Bike

?

Convolutional Neural Network

• Local Connectivity

• Parameter Sharing

• Pooling layer ( Subsampling )

Convolutional Neural Network

Kernel convolution

6x6 Input

3x3 Filter

17sum

Conv2d

Input n X n Filter m x m

Conv2d Output n X n

Conv2d

Input n * n Filter m * m

Conv2d Output n * n

Input n * n * k Filter m * m * k

Conv2d Output n * n

Max pooling

Strided convolutions

Strided convolutions

Convolutional Neural Network

Convolutional Neural Network

(256x256x3)

Filter Size = 5x5

Convolutional Neural Network

(256x256x3)

Filter Size = 5x5

L1 = (5x5x3) x 4

(256x256x4)

Convolutional Neural Network

(256x256x3)

Filter Size = 5x5

F1 = (5x5x3) x 4

(256x256x4)

F2 = (5x5x4) x 8

(128x128x4) (128x128x8)

Convolutional Neural Network

(256x256x3)

Filter Size = 5x5

F1 = (5x5x3) x 4= 300

(256x256x4)

F2 = (5x5x4) x 8= 800

(128x128x4) (128x128x8) (64x64x8)

(32687 x h)

CNN in Character animation

• Conv1d in time dimension

Filter3x50

Input11 x 50

CNN in Character animation

• Conv1d in time dimension

Filter3x50

Input11 x 50

CNN in Character animation

• Environment model

CNN in Character animation

Recurrent Neural Network

He likes apples.

그는 사과를 좋아한다.

Feed-forward Neural Network

𝑊

𝑥

𝑦

x1

x2

x3

y1

y2

𝑤𝑖𝑗

Recurrent Neural Network

Recurrent Neural Network

Truncated back propagation through time (BPTT)

* Vanishing gradients problem

Long Short-Term Memory(LSTM)

Types of RNN

Types of RNN

Motion recognition Motion prediction Motion generation

top related