Julie Chang says that for the last hundred years humans have been the end users of images and photographs. This has led to technology that creates images for humans. Artificial intelligence, however, needs different information. Researchers in artificial intelligence developed systems to use hundreds of thousands of images to build their databases and decision structures. Neural networks can be envisioned as a brain made of layers of mathematical neurons, each taking information and acting like Instagram filters. In her TEDxBeacon Street talk Why we need to design cameras for robots, Chang explains how machines can benefit from images built specifically for machines, and how that can help humanity.
After a neural network is shown thousands of images of cats the network can take the information inside those images and also recognize additional pictures of cats. One drawback is that neural networks and becoming larger and requiring more time, energy, and memory to perform properly. Chang envisions a world where our cell phones could be image recognition tools, but says that current methods would need to be highly compressed to fit on our phones and lose accuracy. Sending images from our phone to the cloud might take too much computing time or use too much data. She also brings up the scary possibility of a self-driving car losing signal while processing images for collision avoidance.
Traditionally, Chang says, the best cameras have been the ones that input and process information exactly like our brains do. The next logical jump is building cameras that will process information like artificial intelligence does, or build the camera directly into the algorithm.
Julie Chang is an engaging speaker, able to bring her own experiences with light and Instagram filters into the discussion of image processing. As she demonstrates the cameras and neural network tools under development at the Stanford Computational Imaging Lab, it’s easy to see how robots of the future will have better data points to make decisions.