Lebanon, Schumer, Julie discussed on This Week in Machine Learning & AI

Automatic TRANSCRIPT

Let's say for radiology readings although others say what you care about maybe human Lebanon Ability is not possible even in the brain. And we've kind of fooling ourselves. You know by saying you know we understand this area. Does these in these areas. Does that Schumer temperate up. But I think that's uninteresting conversation debate that we will see more of it or into ability and I think we if there is such thing is interpreted ability and we understand it in the brain. Then he's going to be easier for engineers to like now. We know that thing exists. Let's try and figure out how we can put it in the bread and other presentational level where we'll call from yours and this is what would most of the the that perspective paper focuses on these representations level if we need to figure out interpret ability and the brain before we're able to get to interpret ability or explain ability in machine learning models. I think we're probably in trouble. Yeah that is true by the interesting though because the tools that people in mass shooting developing to gain interpret it exactly the two that we're not trying to implementing the brand so that's where there's a very strong link again between the two fields. Are there examples that come to mind of successes that we've seen already in Pulling over understanding that we've gained on the machine learning deep learning side to the neuro science side. Yeah so we got the paper in needs last year. Julie was the for stole for Judy's he to gum recordings in these kids from the mouths visualises them and and build. We built a model actor these inception Lube model and then we generated a similarity matrix physically. We showed let's say foul Naturally images the model and we computed the similarity matrix between May Johnny May J. in the models debris models mural spades gate and then we took a neural network and we tried instead of teaching it to say cats and dogs. We also tried to make representations to look more similar are in terms of these presentational seamlessly metrics to the.

Coming up next