Representation Learning with Deep Auto-Encoder Hanock Kwak, Byoung-Tak Zhang Department of Computer Science and Engineering, Seoul National University, hnkwak@bi.snu.ac.kr Department of Computer Science and Engineering, Seoul National University, btzhang@bi.snu.ac.kr Backgrounds Experimental Results Dimensionality reduction facilitates the classification, visualization, communication, and storage of high-dimensional data. Auto-encoder is a nonlinear generalization of PCA that uses an adaptive, multilayer encoder network to transform the high- dimensional data into a low-dimensional code. Gradient descent can be used for fine-tuning the weights neural networks. Reconstruction results Methods Interpolation on each layer Starting with random weights in the two networks, they can be trained together by minimizing the discrepancy between the original data and its reconstruction. The required gradients are obtained by using the chain rule to backpropagate error derivatives first through the decoder network and then through the encoder network. raw h1 h2 h3 raw h1 h2 h3 Loss curves for each layer h1 h2 h3 Reconstruction of noisy inputs Putting Bernoulli’s random noise to test robustness of auto- encoder. Discussion The manifold of digit data is flattened in deep hidden layers which is shown in the interpolation experiment results. Contractive penalty always helps an auto-encoder to perform better, and competes or improves upon the representations learned by ordinary auto-encoders. The higher representation of deep auto-encoder can eliminate minor noises in the input through forwarding operations of multilayer perceptrons. Here, the input is more abstracted on deeper hidden layer. References Vincent, Pascal, et al. "Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion." The Journal of Machine Learning Research 11 (2010): 3371-3408. Y. Bengio, G. Mesnil, Y. Dauphin, S. Rifai, Better mixing via Deep Representations, ICML 2013 Hinton, G. E., & Salakhutdinov, R. R. (2006). Reducing the dimensionality of data with neural networks. Science, 313(5786), 504-507 Rifai, Salah, et al. "Contractive auto-encoders: Explicit invariance during feature extraction." Proceedings of the 28th international conference on machine learning (ICML-11). 2011. Biointelligence Lab, Seoul National University | Seoul 151-744, Korea (http://bi.snu.ac.kr)