Hacker News new | past | comments | ask | show | jobs | submit login
Facebook’s Quest to Build an Artificial Brain Depends on Yann LeCun (wired.com)
44 points by ajtulloch on Aug 14, 2014 | hide | past | favorite | 9 comments

Lecun's work on LeNet was pioneering, and basically set the template for deep learning. However, it's not as simple as him being "right all along"; incremental advances in neural net architectures, a lot of developments in stochastic gradient methods, and orders-of-magnitude improvements in hardware and data availability have been what made deep learning as powerful as it is today.

The main point is that skeptics in the 2000's were basically right about neural nets being of limited use, but other technologies advanced and broke down the barriers.

Lecun's original convolutional nets in 1998 were run on the then-gigantic dataset of 60,000 images. Consider that a company like Facebook can provide billions of images with some form of tagging, and you see the different world we live in.

I don't think its fair, or sincere, for you to use past tense to describe anyone's current work.

I'm calling his work on LeNet in 1998 pioneering. I'm not trying to disparage his current work.

I'd love to learn more about AI. What are some good resources to start learning?

The pricey but standard recommendation would be buying and working through "Artificial Intelligence - A Modern Approach". It's one of my favourite IT books :)

There's a pretty fun/good course on one of the free MOOCs that pretty much follows the book and uses pacman as an example. It uses Python. It's the course that has a cute robot on the slides :) Edit (found it, this one): https://www.edx.org/course/uc-berkeleyx/uc-berkeleyx-cs188-1...

Edit2: There's also this one but it's not the one I'm thinking of (it's also good though, Norvig is one of the authors of AI-AMA): https://www.udacity.com/course/cs271

Here's a good course on NNs: https://www.coursera.org/course/neuralnets

Most likely you and your kids are gonna be wiped out by intellegent machines just in a few decades.

Probably, but it's unlikely to be from convolutional neural networks.

Maybe not wipe out, but potentially take your job.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact