22 February 2018
To help apply and develop their literature research on Deep Learning, Jake Rankin attended the Deep Learning Workshop. This would both provide experience with applying deep learning and an opportunity to discuss specific considerations with leading practitioners for implementing deep learning systems. The Centre for Modelling and Simulation (CFMS) Hosted a Deep Learning Workshop at the Bristol and Bath Science park, consisting of talks from Cray, Nvidia Deep Learning Institute, TSystemes and CFMS. This was then supported by a workshop that used Nvidia’s Deep Learning GPU Training System (DIGITS). The purpose of this event was to help organisations understand the challenges of implementing deep learning and the presentations were focused on business applications of deep learning and clarifying the mechanisms of deep learning.
Chris Hegarty (Cray, AI Enterprise Sector Development Lead) gave a useful presentation on the hardware and software required for the implementation of Artificial Intelligence Systems, highlighting several useful toolkits such as CNTK, TensorFlow and Caffe2 (All of which are available for use), but also highlighted a word of caution with regards to the data-processing required (Exceeding well above 25 Gflops) and the costs for algorithm training (~$25k). Adam Grzywaczewski (Nvidia, Deep Learning Solution Architect), who also led the Deep Learning Institute workshops, explained the workings behind neural networks very concisely, with focus on Convolutional Neural Networks (CNN), which would be the focus of our workshop.
The Deep Learning Institute workshop applied deep learning techniques to image processing. This was broken down into three lessons:
- Image Classification with DIGITS
- Object Detection with DIGITS
- Neural Network Deployment with DIGITS and TensorFlow
The conference was a useful experience that has helped to focus literature research and consider practical constraints for the project. The networking was a useful insight of business opinions on various deep learning methods and helpful feedback was shared.
For those interested in finding out more about Nvidia DLI and future workshops, please visit www.nvidia.com/dli