A new story from This Week in Machine Learning & AI

Automatic TRANSCRIPT

More scientific is a science not on engineering field. You won't understand which is the same as neuroscience in some ways what are the principles that enable the same network to perform the task in also do transfer learning from one to the other very fast it one of the areas that seeing some interesting research on the machine learning site is The idea of multi task learning so we can train these networks to do two things at the same time instead of one or and things at the same time instead of one and their You know that somehow has this kind of regularizing. You know that makes them perform better. Is there any kind of inspiration to that or biological parallel to to that? Yeah yeah actually. We discussed that in the paper. Also multitask train and When we see the world we don't just do up to condition we. We know that The distance of the object. We can grasp he'd we have some idea about the The texture a week we can multi scale perception that we can look at individual. You know eyebrow on someone's phase in Fades we can you know? Do you know figure someone sat up. So we do a multi task multi cultural embedding. And if you want the same brain is being all these things. So that's one of the areas. Where with the brains very different than my shade and I think there are people that are single case. Maybe we should like we train on network. All the tasks Mehboob generalize also better be more revised. It will be you know. So that Sorta of a in-interesting direction that we are interested in. Maybe you can. You're absolutely right. That's very important that the problem again becomes. How do you train these? Because if you rely on human labeling all this data that you need to have humor liberal the collar detect and then you sort of again limited by Jaba. You could like Brute force approach at least in toy examples to show that this is the right direction. Is there any one particular direction in Kinda this entire space biological systems to deep learning deep learning to biological systems that you're most excited about the good question? I think that Mostly thing right now in the next few years the most exciting directions the cognitive or behavioral level and the representation level. Because of you know like any maybe a practical thing because we we have good baseline models that relate covid cost function. What their training and the cover presentations that. We can measure right mostly excited about this to higher levels in term so building models that are going get information about best or inspiration in a more or less may be ambitious way from the brain to advance if you go down to the implementation level obliging. He may happen at some point this too much complexity down there in with understand. Why is this complex? It could be implementation it could be biological with where constraints for example energy constraints like you. The brain has chemistry any doesn't have silica and then Need neurotransmitters because of Biology..

Coming up next