Abstract

This presentation will start with a computation overview of deep learning models, as well as a quick tutorial of Theano. Compared with other deep learning toolkits (e.g., Caffe and Torch7), Theano is arguably more tailored towards research use or easier to learn for python/numpy users. In the second half of this talk, I will present a case study with Theano for the problem of synonym detection. We implement a neural network model with Theano, which significantly improves the state-of-the-art (IJCAI 15) on two million-scale datasets: F1= 0.83 vs. 0.71 on a medical synonym dataset, and F1 =  0.38 vs. 0.19  on a general WordNet based synonym dataset.