I-Corps: Artificial Intelligence and Deep Learning System for Product Search

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

The broader impact/commercial potential of this I-Corps project are in various applications in ecommerce, security, robotics, etc. This project's technology enables computer systems and their users to better contextually understand products and locate them. For example, in e-commerce, this technology enables people to make informed decisions about product selection. This may enhance the e-tail experience and help to eliminate wasted time as people can easily and precisely find things they desire. The aggregated data on product searches can be used by suppliers to better understand and plan for demand. Also, in robotics, as another example, computer vision is an essential element of the technology and the multi-object detection system built into software will enable robots to be more effective as they will "understand" products or objects in the context of our world. In security, this technology will enable better detection of theft and threats. Overall, the technology has many applications - many that can make lives more convenient and safer. This I-Corps project is based on machine learning technology. Billions of data points that include images and textual information about products are trained into artificial neural networks. These neural networks understand products with context with both images and natural language. Further, the neural networks are able to self-learn with new information from the internet. This technology was developed after researching various machine learning methods for large scale objection detection and search. The result is that neural networks provide the most scalable and efficient technology in terms of object detection. The uniqueness of this system is that parts of the learning in the neural network can be changed without affecting the rest of the network and the network can self-learn from the internet about products and other choice-centric decisions.