Deep learning is a popular machine learning technique and has been applied to many real-world problems, ranging from computer vision to natural language processing. In most cases deep learning outperformed previous work. However, training a deep neural network is very time-consuming, especially on big data. A popular solution is to distribute and parallel the training process across multiple machines. Indeed, the race is on to parallelize deep learning! Industry and academic research teams around the world are trying to make deep neural networks train as fast as possible on farms of GPU capable servers. We are working with our IBM partners to help develop advanced scheduling and messaging techniques for distributed deep learning. In addition, we will develop two real-world applications of distributed deep learning to demonstrate the efficiency and effectiveness of distributed deep learning. In one application, we address the video surveillance problem of tracking a moving target over a network of video cameras with partial or no overlaps in their coverage. We will use a deep learning approach to identify multiple pedestrians in each video frame, and a particle filter to track moving pedestrians. In the second application, we address the problem of fraud/intrusion detection. We will use graph-based detection that considers relationships between objects or individuals. Graph-based approaches are powerful because they do not operate on objects or individuals in isolation, but also consider their network information. We will emphasize on graph-based fraud detection methods that have a number of applications and potentially large impacts.

Industry Partner(s):IBM Canada Ltd.

Academic Institution:York University

Academic Researcher: Aijun An

Co-PI Name: Amir Asif

Focus Areas: Digital Media

Platforms: Cloud, GPU