Deep Learning Conference

In this workshop we will build a RNN-based natural language model that can answer reading comprehension questions via chaining facts, deduction, induction...  Imagine a machine learning model that can read a few factoids and then accurately respond to questions about this information.

We will use Keras to reproduce and improve on a recent deep learning paper on question answering from Facebook AI Research. Keras is a minimalist, highly modular neural network library for Python, that uses Theano under the hood for optimized tensor manipulation on GPU and CPU. It was developed with a focus on enabling fast experimentation.

We will evaluate different neural networks architectures, train a model on GPU and create a toy Q&A application.

In our talk we will give an overview of Deep Learning methods used for time series analysis. We will give a short introduction to Time-Delay NN, Recurrent NN, and Long Short-Term Memory networks.

We will also discuss the applications of these methods for Automatic Speech Recognition and Natural Language Processing.

Deep learning algorithms have achieved state-of-the-art results in many domains such as natural language processing, image recognition, and speech recognition. While traditional machine learning algorithms rely heavily on handcrafted input features, deep learning automatically finds useful feature representations for these complex inputs (given enough data). This workshop will be a state-of-the-art implementation for a common machine learning task: object detection.

OpenDeep is a modular deep learning Python framework written on top of Theano. It enables complex networks to be built like Legos - connecting basic layer classes into larger models. You can easily debug networks through visualization and monitoring tools, as well as run everything on the GPU. OpenDeep's motivating factor is to be extensible enough to perform novel research and also simple enough to quickly set up deep learning algorithms at scale.

For the object detection task, we will implement the recent state-of-the-art paper "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks" by Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. This paper exemplifies the modularity of deep learning by combining aspects of separate convolutional neural network (CNN) models.

In this workshop we will highlight neon, an open source python based deep learning framework that has been built from the ground up for speed and ease of use. We will start with a general overview of how to use neon, build Recurrent Neural Networks that can be used to generate and analyze text, and build Convolutional Autoencoders that can be used to generate images and to localize objects. I will also demo the integration of neon with the Nervana Cloud (currently in private beta) for multi-GPU training of deep networks.