September 26-27, 2016
New York, NY

Managing the deep learning computer-vision pipeline with DIGITS

Jon Barker (NVIDIA)
11:50am–12:30pm Tuesday, 09/27/2016
Implementing AI
Location: River Pavilion B Level: Intermediate
Average rating: *****
(5.00, 1 rating)

Prerequisite knowledge

  • Firsthand experience with deep learning
  • What you'll learn

  • Gain the basic knowledge needed to begin deep learning trainingĀ in your organization
  • Description

    The process for deploying an effective neural network is iterative. Before an effective neural network is reached, many parameters must be evaluated and their impact on performance assessed. Jon Barker offers an overview of DIGITS, a deep learning GPU-training system designed to provide a real-time interactive user interface targeted toward accelerating the development process.

    DIGITS is a easy-to-use tool that allows researchers and scientists to quickly develop a trained neural network for their dataset. It possesses functionality for a wide range of popular use cases for deep learning and features for parameter testing. The real-time performance features during training enable easy monitoring and decision making. This is especially advantageous because the training can sometimes last days or even weeks.

    The initial release of DIGITS allowed users to quickly create a convolutional neural network for classification. Since then it has been expanded for more use cases and includes detection, segmentation, regression, and auto-encoders. DIGITS simplifies the dataset creation and network tuning processing of deep learning. It has a graphical interface that works with two popular deep learning frameworks, Caffe and Torch, which makes it easy to deploy trained networks. With DIGITS, a hyperparameter search can easily be performed with a given dataset, and performance can be compared. The best performer can then be downloaded and deployed on the target platform with the framework used to create it.

    Networks can be visualized to assess network performance. Some research indicates that artificial neurons can activate more than others and in some cases rarely at all. Each layer of a trained network can be visualized, showing individual neuron activations. This permits developers to view and evaluate neuron activity to different input classes and prune if needed.

    For each training case, the performance as a function of the epoch is displayed and saved. This feature displays underfitting and overfitting if either has occurred. If is common for networks to overfit to the data. Saved networks enable evaluation at different points during the training process. If overfitting is discerned halfway through a single training, the network from before this point can be tested and used for deployment. In the event that underfitting occurs, it is easily discerned in the graphical display along with the training parameters so that it is easy to prevent the configuration from being used again accidentally.

    Photo of Jon Barker

    Jon Barker

    NVIDIA

    Jon Barker is a solution architect with NVIDIA, helping customers and partners develop applications of GPU-accelerated machine learning and data analytics to solve defense and national security problems. Jon is particularly focused on applications of the rapidly developing field of deep learning. Prior to joining NVIDIA, Jon spent almost a decade as a government research scientist within the UK Ministry of Defence and the US Department of Defense R&D communities. While in government service, he led R&D projects in sensor data fusion, big data analytics, and machine learning for multimodal sensor data to support military situational awareness and aid decision making. Jon has a PhD and BS in pure mathematics from the University of Southampton, UK.