Computer Vision News - July 2021

in neuroscience, with the goal of better understanding the structure of animal behavior and how it’s controlled by the brain. Within the lab, we work with computer vision tools for doing automated pose estimation and behavior classification, mostly in mice. We also work with groups that are recording the activity of neurons in different parts of the brain. We develop methods to relate what those neurons are doing to the animal’s behavior to try to understand how the brain is encoding behavior and, in the big picture, how circuits of multiple interacting parts of the brain giving rise to behavioral choices in animals. Everything you said so far raises so manyquestions!Sinceyouarestudying how the cognitive process works, do you think our brain is functioningwell? Also, since you are studying animals, do you think the brain is functioning as it should? Would you suggest any improvements? [ laugh s] That’s a good question! During my postdoc, we were looking at very evolutionarily conserved circuits in the brain, or governing survival behaviors, like defending your territory, escaping from predators, feeding, those kinds of behaviors. These are parts of the brain that have been around for a very long time. They have been genetically wired to keep animals alive - to survive and reproduce. These areas have been under tremendous evolutionary pressure. It’s incredible how much you can accomplish just by wiring these things up genetically without any sort of supervised training of the circuits. Animals are born and sometimes run within a couple hours of being born; they can find food, evade predators. All of that is encoded in their genes and to some extent there in the brain from birth. Obviously, there is learning and experience on top of that, butthe level of things that the system can achieve without any supervised learning or reinforcement learning is really fascinating to me. That is something brains do very well! Regarding our learning process, I have noticed that if you show me several fire extinguishers, I will quickly be able to recognize one. Why does a machine need to see the same image thousands of times before recognizing it? How does the brain do this more effectively? I guess you’re talking about few shot learning, the ability to recognize an object the first time you see it. That’s something that people are working on in object recognition: transfer learning or meta learning or few shot learning. This isn’t my area of research so I probably won’t do a great job describing this, but if you learn to recognize enough objects in the world, like cats and boats and airplanes and trees, then you’ve built a repertoire of features general enough that you can train a new part of Ann Kennedy 39 Best of CVPR 2021