Ufuk Can Biçici

Ufuk Can Bicici, Tuna Han Salih Meral, Lale Akarun

Conditional computing processes an input using only part of the neural network's computational units. Learning to execute parts of a deep convolutional network by routing individual samples has several advantages: Reducing the computational burden is an obvious advantage. Furthermore, if similar classes are routed to the same path, that part of the network learns to discriminate between finer differences and better classification accuracies can be attained with fewer parameters. Recently, several papers have exploited this idea to take a particular child of a node in a tree-shaped network or to skip parts of a network. In this work, we follow a Trellis-based approach for generating specific execution paths in a deep convolutional neural network. We have designed routing mechanisms that use differentiable information gain-based cost functions to determine which subset of features in a convolutional layer will be executed. We call our method Conditional Information Gain Trellis (CIGT). We show that our conditional execution mechanism achieves comparable or better model performance compared to unconditional baselines, using only a fraction of the computational resources.

Ufuk Can Biçici, Cem Keskin, Lale Akarun

Deep neural network models owe their representational power to the high number of learnable parameters. It is often infeasible to run these largely parametrized deep models in limited resource environments, like mobile phones. Network models employing conditional computing are able to reduce computational requirements while achieving high representational power, with their ability to model hierarchies. We propose Conditional Information Gain Networks, which allow the feed forward deep neural networks to execute conditionally, skipping parts of the model based on the sample and the decision mechanisms inserted in the architecture. These decision mechanisms are trained using cost functions based on differentiable Information Gain, inspired by the training procedures of decision trees. These information gain based decision mechanisms are differentiable and can be trained end-to-end using a unified framework with a general cost function, covering both classification and decision losses. We test the effectiveness of the proposed method on MNIST and recently introduced Fashion MNIST datasets and show that our information gain based conditional execution approach can achieve better or comparable classification results using significantly fewer parameters, compared to standard convolutional neural network baselines.