We Need To Go with DEEPER Inception Network

  • Slides: 18
Download presentation
We Need To Go with DEEPER Inception Network 1조 최영환 박형준 김종헌 성광현 한준희

We Need To Go with DEEPER Inception Network 1조 최영환 박형준 김종헌 성광현 한준희

CONTENTS 01 02 03 04 Background Basic Building Block Network In Network Versions of

CONTENTS 01 02 03 04 Background Basic Building Block Network In Network Versions of the DNN의 성능 향상과 한계 Inception module의 1 X 1 convolution layer의 Inception network 기본 개념과 구성 요소 필요성

01. Background Deep Neural Network 성능 향상시키기: Increasing Depth & Width

01. Background Deep Neural Network 성능 향상시키기: Increasing Depth & Width

01. Background Convolution Neural Network But. . 1. 더 많은 parameter → overfitting 2.

01. Background Convolution Neural Network But. . 1. 더 많은 parameter → overfitting 2. 연산량의 증가

02. Basic Building Block Sparse & Dense 크기가 다른 Filter와 Pooling 을 여러 개

02. Basic Building Block Sparse & Dense 크기가 다른 Filter와 Pooling 을 여러 개 적용해서 그 결과를 결합

02. Basic Building Block Zoom in & Zoom out

02. Basic Building Block Zoom in & Zoom out

02. Basic Building Block 늘어나는 output 채널 수

02. Basic Building Block 늘어나는 output 채널 수

02. Basic Building Block Problem: Computational Cost 단순 5 x 5 필터만 보아도 필요한

02. Basic Building Block Problem: Computational Cost 단순 5 x 5 필터만 보아도 필요한 곱셈은 28 x 32 x 5 x 192 = 약 1억 2000 만개

03. Network In Network Min Lin, 2013 Network In Network (NIN)

03. Network In Network Min Lin, 2013 Network In Network (NIN)

03. Network In Network Min Lin, 2013 Network In Network (NIN) Feature 추출 능력이

03. Network In Network Min Lin, 2013 Network In Network (NIN) Feature 추출 능력이 더 우수해짐

04. Versions of the Inception network Inception v 1(Goog. Le. Net), v 2, v

04. Versions of the Inception network Inception v 1(Goog. Le. Net), v 2, v 3, v 4…

04. Versions of the Inception network Inception v 1(Goog. Le. Net), v 2, v

04. Versions of the Inception network Inception v 1(Goog. Le. Net), v 2, v 3, v 4…

04. Versions of the Inception network Inception v 1(Goog. Le. Net), v 2, v

04. Versions of the Inception network Inception v 1(Goog. Le. Net), v 2, v 3, v 4…