Jimmy Whitaker, a software engineer at Digital Reasoning, will be speaking at the GPU Technology Conference (GTC) in San Jose, Calif., April 4-7. GTC is the largest event of the year for GPU developers, and showcases the most important work in the visual computing industry today – including Artificial Intelligence, Deep Learning, Virtual Reality and Self-Driving Cars.
The presence of public data on the web has allowed for many innovations in how humans are able to search, process, and understand information. The majority of these innovations have relied on processing textual data and have yielded tremendous results. In the past 10 years, however, the amount of digital media, such as digital images, on the public web has seen tremendous growth and with it, the need to be able to process and understand it.
Recently deep learning techniques such as Deep Convolutional Neural Networks (ConvNets) have achieved state-of-the-art results in many computer vision tasks. The data-driven nature of deep learning normally requires a large number of labeled examples in order to achieve high accuracies. Unfortunately, much of the publicly available data on the web is not labeled, thus requiring human labelers for large datasets or unsupervised machine learning techniques.
In his session, titled Bootstrapping Labels for One-Hundred Million Images, Whitaker will describe how to create an iterative labeling process to perform data science on +100 million images using a GPU-powered workflow with Convolutional Neural Networks. This labeling process involves using transfer learning to extract convolutional features from a set of images. The images are then clustered based on the extracted features to create weak classifiers. The clusters are assigned weak labels. We then fine-tune the classifier based on the clusters, and repeat the process. The GPU architecture has made the iterative process feasible for a fraction of the cost of CPU-based architectures.
This labeling process has been used to enable investigative tools to explore unlabeled digital media. Pending material sensitivity, the session may contain concrete examples from human trafficking applications, media aggregation, and/or medical imaging.
Convolutional Neural Network architecture speeds are 30-60x faster than the CPU implementations. The overall speed and performance of our approach is still being explored. Additionally, another GPU based hardware configuration is being planned which should allow for even faster performance.
To find out more about the GPU Technology Conference, presented by NVIDIA, please visit www.gputechconf.com.