Microsoft Challenges Google’s Artificial Brain With ‘Project Adam’ | Hacker News Reader

Microsoft Challenges Google’s Artificial Brain With ‘Project Adam’

wired.com

20 pointsdigital555 comments

As an academic researcher in object recognition, I'm starting to wonder if I can still make significant contributions in this area, given that I'm competing with larger teams with far greater resources at Google, Microsoft, etc. For people like me, it is probably time to pivot to new problems in machine perception; the challenge is getting funding for them.

There's a lot of room left! Can you get accuracy at 120 fps? Can you do that in an embedded system? Can you use temporal and contextual queues to build a self-supervised system? Can you handle stereo and depth? Can the system handle a wide variety of image sizes? Can it handle rotation? Can it handle obstruction?

There is so so much more to vision than ImageNet! Go visit your robotics lab and they'll give you a giant laundry list of things they need but don't have.

Sure, I agree with you that there is more to vision than ImageNet (and many-category object recognition), and that's why I indicated I'm looking for other computer vision problems to work on that I think I can solve.

Incidentally, these deep convolutional neural networks do handle rotation very well (see Table 1, Krizhevsky et al): http://arxiv.org/pdf/1301.3530.pdf

It sounds like you're not very interested in you research. If you were, you would be excited that more people are working on it, instead of being scared of competition.

This is a nice result. I'm sure a lot of people had this particular idea but never got to actually do it. Not surprising because this is actually why to human brain scales so well. This was also hinted by AutoEncoders where pieces of a net are built separately bottom-up. I'm guessing a final settling phase would also be useful here.

One unfortunate thing is that different runs will produce different nets. But if training is radically faster and accuracy differences are small than this is one new technique to add to the toolbox.

Reply on news.ycombinator.com