CVPR Daily - Thursday

maybe also say quote-unquote because again, there is still a lack of understanding and grounding in many cases, but there are impressive capabilities which we should, and we are looking at them and exploring them. Can you give an example? What’s something we can’t do today that you expect we’ll be able to do five years or ten years from now? There are probably too many things to name a specific one when it comes to reasoning tasks. When we need to reason about, especially think of any long documents or things which require external knowledge, things which require multiple hops of sort of reasoning steps, I think we're not quite there. Many realistic problems out there require such reasoning. Humans can do this and solve complicated problems. Machines are just not there yet. Maybe one example that kind of inspires me is if you think of a very prominent issue of misinformation online. To actually understand, detect, and prevent the spread of misinformation, it's a very complex process where human fact- checkers are involved. They have to sit and sift through data and analyze data enabling this kind of process. Automating that with machines would be hugely impactful because we could stop the harmful misinformation from spreading. But accessing knowledge, accessing sort of reasoning skills, all of these are still out of reach for now. Maybe somebody who will be able to make a real breakthrough in this area would win the Nobel Prize for that? They don't give Nobel Prizes for that one. No chance. Maybe one day there will be a Nobel Prize for computer science. We'll see. Can you tell me about your team? I'm part of the group together with Professor Trevor Darrell and many bright students and postdocs at Berkeley. Without naming names, perhaps, but there are many. You said many. Just to give an idea, how many? Goodness, gracious! You're asking difficult questions … 21 DAILY CVPR Thursday Ukrainian dress UK AINE CORNER Anna Rohrbach