Augmenting supervised neural networks with unsupervised objectives for large-scale image classification

arXiv:1606.06582

왠지 중요한것 같아서라기보다 CVPR2017 best paper가 인용한 논문이 한국사람의 논문이라서 봄

학습할 때 gradient vanishing, explosion등 문제를 해결하는 방법의 일종인듯.. 괜히읽었다. Augmented autoencoder의 variation 세가지를 제안함.

코드는 개인 홈페이지에 공개

Abstract

we investigate joint supervised and unsupervised learning in a large-scale setting by augmenting existing neural networks with decoding pathways for reconstruction.

1. Introduction

Unsupervised로 pre-training 한참 하다가 요즘엔 아무도 안한다. There are many (supervised-learning) technics which were not limited by the modeling assumptions of unsupervised methods.

Supervised랑 unsupervised를 엮으려는 시도들이 조금 있었으나 대량의 데이터에 관한 것이 아니었다. Zhao et al.이 Zeiler et al.의 “unpooling” operator를 사용해 (stacked) “what-where” autoencoder(SWWAE)를 제안했고 CIFAR with extended unlabeled data에서 괜찮은 결과를 얻었으나, 그보다 더 큰 데이터에서는 그렇지 못했다.

Classification network의 일부를 떼어내고 그것을 그대로 뒤집(은 후 augmentation을 더해 얻)은 것을 (reconstructive) decoding pathway로 쓰겠다. 이 autoencoder를 finetuning하면, 앞단(원래 classification에 쓰이던 부분)의 성능도 향상된다.

We will provide insight on the importance of the pooling switches and the layer-wise reconstruction loss.

2. Related work

• sparse coding
• dictionary learning
• autoencoder

3.Methods

3.1. Unsupervised loss for intermediate representations

원래 보통의 loss는 $$\frac{1}{N}\sum^N_{i=1} C(x_i, y_i), C(x,y) = l(a_{L+1}, y)$$ where

\(a_L\) :\(L\)th layer’s output,

\(y\) : ground truth,

\(l\) : cross-entropy loss,

\(C\) : classification loss.

이지만, net 중간중간에 loss를 만들도록 적당히 뭔가를 넣으면(3.2에 나온다) $$\frac{1}{N}\sum^N_{i=1} (C(x_i, y_i)+ \lambda U(x_i))$$ 이렇게 할 수 있다.그냥 뻔한 이야기

3.2. Network augmentation with autoencoders

아래가 저자들이 제안한 것.

아래는 ladder net

4. Experiments

• ImageNet ILSVRC 2012 dataset
• mainly based on the 16-layer VGGNet
  • partially used AlexNet

4.1. Training procedure

Steps

1. 초기화 : encoding pathway는 pretrained, decoding은 Gaussian
2. encoding을 고정시킨 후 SAE-layerwise net을 학습시킴
3. SAE-first/all은 SAE-layerwise를 pretrained net으로 해서 학습시킴
4. encoding/decoding 모두 reduce learning rate으로 finetuning

SGD momentum 0.9 꽤나 크게 줬네? (We found the learning rate annealing not critical for SAE-layerwise pretraining. Proper base learning rates could make it sufficiently converged within 1 epoch.)

4.2. Image reconstruction via decoding pathways

일단 reconstruction이 엄청 잘된다고 한다.

참고: unpooling switch

4.3. Image classification with augmented architectures

VGG-16으로 finetuning했는데 결과가 좋다는 내용(Compared to the VGGNet baseline, the SWWAE-all model reduced the top-1 errors by 1.66% and 1.18% for the single-crop and convolution schemes, respectively. It also reduced the top-5 errors by 1.01% and 0.81%, which are 10% and 9% relative to the baseline errors.)

약간 overfit되는 경향(lower training error)도 관찰된다고 하나 결과가 더 좋은것으로 보아 학습 전반에 긍정적 효과가 있는듯.

Conclusion

1. the pooling switch connections between the encoding and decoding pathways were helpful, but not critical for improving the performance of the classification network in large-scale settings
2. the decoding pathways mainly helped the supervised objective reach a better optimum
3. the layer-wise reconstruction loss could effectively regularize the solution to the joint objective.(= incorporating layer-wise reconstruction loss)