Expanding Usability of Convolutional Neural Networks
Time: 4:00 PM, Wednesday, March 28, 2018
Place: 235 Weir Hall
Speaker: Torumoy Ghoshal
Abstract: Deep convolutional neural networks show superior performance when there is significant spatial locality present, finding their best use in image and their derivative datasets. Attempts to use text-based data with convolutional networks have usually relied on some text-representation models. Success in such models is still dependent on the efficacy of the text-vectors. Little research into linguistics suggests that incorporating contextualized information may be a better representation for text based models. However, the vectors may be sparse in nature and may not show enough spatial locality. Using dimensionality reduction techniques a dense representation of sentence vectors can be achieved. While improving representation and reducing dimensionality may improve spatial locality, the architecture of convolutional networks still provides research opportunities for improvements. The final Softmax layer in traditional convolutional architectures may be replaced by a stochastic, differentiable, and back-propagation compatible decision tree model that enables end-to-end training. In this report I summarize techniques and findings from three papers related to the concepts mentioned above.