Showing papers by "Ethem Alpaydin published in 2018"

Continuously Constructive Deep Neural Networks

Abstract: Traditionally, deep learning algorithms update the network weights whereas the network architecture is chosen manually, using a process of trial and error. In this work, we propose two novel approaches that automatically update the network structure while also learning its weights. The novelty of our approach lies in our parameterization where the depth, or additional complexity, is encapsulated continuously in the parameter space through control parameters that add additional complexity. We propose two methods: In tunnel networks, this selection is done at the level of a hidden unit, and in budding perceptrons, this is done at the level of a network layer; updating this control parameter introduces either another hidden unit or another hidden layer. We show the effectiveness of our methods on the synthetic two-spirals data and on two real data sets of MNIST and MIRFLICKR, where we see that our proposed methods, with the same set of hyperparameters, can correctly adjust the network complexity to the task complexity.

Dropout Regularization in Hierarchical Mixture of Experts

Abstract: Dropout is a very effective method in preventing overfitting and has become the go-to regularizer for multi-layer neural networks in recent years. Hierarchical mixture of experts is a hierarchically gated model that defines a soft decision tree where leaves correspond to experts and decision nodes correspond to gating models that softly choose between its children, and as such, the model defines a soft hierarchical partitioning of the input space. In this work, we propose a variant of dropout for hierarchical mixture of experts that is faithful to the tree hierarchy defined by the model, as opposed to having a flat, unitwise independent application of dropout as one has with multi-layer perceptrons. We show that on a synthetic regression data and on MNIST and CIFAR-10 datasets, our proposed dropout mechanism prevents overfitting on trees with many levels improving generalization and providing smoother fits.

Convolutional Soft Decision Trees

Abstract: Soft decision trees, aka hierarchical mixture of experts, are composed of soft multivariate decision nodes and output-predicting leaves. Previously, they have been shown to work successfully in supervised classification and regression tasks, as well as in training unsupervised autoencoders. This work has two contributions: First, we show that dropout and dropconnect on input units, previously proposed for deep multi-layer neural networks, can also be used with soft decision trees for regularization. Second, we propose a convolutional extension of the soft decision tree with local feature detectors in successive layers that are trained together with the other parameters of the soft decision tree. Our experiments on four image data sets, MNIST, Fashion-MNIST, CIFAR-10 and Imagenet32, indicate improvements due to both contributions.

What's in a Game? The Effect of Game Complexity on Deep Reinforcement Learning.

Abstract: Deep Reinforcement Learning (DRL) combines deep neural networks with reinforcement learning. These methods, unlike their predecessors, learn end-to-end by extracting high-dimensional representations from raw sensory data to directly predict the actions. DRL methods were shown to master most of the ATARI games, beating humans in a good number of them, using the same algorithm, network architecture and hyper-parameters. However, why DRL works on some games better than others has not been fully investigated. In this paper, we propose that the complexity of each game is defined by a number of factors (the size of the search space, existence/absence of enemies, existence/absence of intermediate reward, and so on) and we posit that how fast and well a game is learned by DRL depends on these factors. Towards this aim, we use simplified Maze and Pacman environments and we conduct experiments to see the effect of such factors on the convergence of DRL. Our results provide a first step in a better understanding of how DRL works and as such will be informative in the future in determining scenarios where DRL can be applied effectively e.g., outside of games.