This is all about applications of deep learning, and give an intuitive explanation of why deep learning works well. G. Hinton and Y. Bengio released several presentations online that present detailed and theoretical part of deep learning. I started my study from there.
In case anyone is curious, the powerpoint presentation Andrew is giving can be found just by Google searching "DeepLearning-Mar2013.pptx". If you want to see some of the slides that Andrew talks to but which do not get into the video, this is helpful.
I remember working on Natural Language Processing as an undergrad in the 80's ... it's come so far ... of course, back then we didn't compete with Neuro-Linguistic Programming for the NLP acronym.
Many thanks for sharing the lecture, it's so clear and concise. Now, I've an overall better understanding of the difference between various machine learning algorithms.
If you are having lag issues ... helper.ipam.ucla.edu/wowzavideo.aspx?vfn=10595.mp4&vfd=gss2012 This video from the site posted by David Sanders (see below) was working better for me.
I feel the same way currently - is that machine learning tends to be just curve fitting and statistics. This is not what I really want to learn when I say I want to learn AI/Machine Learning.
So up to you to invest your spare time into finding new methods, right? I think having a stereoscopic view or a time element, or both will greatly help improve image recognition algorithms. I think humans would also have a lot of trouble recognizing motorcycles if they spend their entire lives living in a world of non-moving 2D images. The reason we can pick apart objects from others is because we have seen them often, at other times, as if moving in 3D space. Humans use a combination of parallax and both eyes to map their 2D view into a 3D understanding of the world. If we want computers to make sense of the world in the same way as we do, the first preprocessing step would be to do the same: try to guess a 3th dimension onto 2D images. Just my thoughts. Might be utter bullshit :)
Wow - great talk! Extremely interesting material. I've been fascinated with AI every since I was a kid. In fact, I think that's what got me into the field of software dev in the first place. Mayhap it's time to truly start playing around with neural nets and learning algorithms.
Should all AI have a standard eventually for a "base" brain? Should an individual AI be on a knowledge island of their own or should they all be interconnected and all be able to share their acquired learning and knowledge?
Well, Gabor-like filters that deep nets tend to discover are the basis of JPEG which is an important part of MPEG compression. But it's only one layer of "features" and there are no learned temporal features that take advantage of the redundancy from frame to frame. The keyframing technique does not count as a learned feature. Geoff Hinton has applied stacked restricted boltzmann machines to video of bouncing balls and found that it can learn temporal features, so yes!
I wish Andrew would move forward from text, audio and images/video as there are more interesting problems and I'd like to hear about possible solutions to these problems via algorithms.
It doesn't actually inspire too much confidence that some of the greatest brains on the planet developing AI that will presumably be responsible for our safety and well being can't figure out how to upload a decent quality RU-vid video ten years after its inception. ;D (jk btw)
The statement at 19:30 "Humans have 20 years of experience looking at unlabeled images" is nonsense. Human experiences are always labelled by their feelings (aka rewards in machine learning).
The early pre-cursor to the tongue "display" (which apparently feels like varying degrees of soda bubbles rather than a 9V battery) was actually a giant chair that replaced pixels with little actuators. Turns out skin isn't so good with definition whereas the tongue is super dense with sensors - more "pixels per inch" if you will.
We're a start up called "#Winning" and we're using convolutional neural networks to predict lottery numbers. We're currently training a recurrent network on coin tosses and dice rolls before moving onto the holy grail (but never holy fail) of predicting next week's lottery numbers
Help!! I am currently working on spatio-temporal feature extraction from videos using deep learning. Unfortunately there hasn't been much work on it. Can anyone provide me with links on deep neural networks for video features???
how can you have so much knowledge and talk about deep learning and advanced AI but not be able to record a 45 minute video? i really don't understand that...
11:00 still i wonder why our brains are so similiar and these brain regions are normally structured according to one specific task, optical cortex for seeing things, and auditory cortex for understanding sound frequency modulations. Why do certain sensory task occur in the same brain regions. That is my question, thanks for the replies if there are any;D really intresting stuff AI!!!!!! 2040 the internet will be alive!;p
+Vrolijke Vent Long wires are expensive. Total connectivity in the human brain is about 15%. It's nonsense to connect A1 with V1. The genes know that and therefore brain region connections are predefined in order to speed up learning.
Voxel Skull it's a chinese surname used by certain dialects, en.wikipedia.org/wiki/Ng_%28surname%29 similar (in chinese character) to those with 'Huang' as surname. You can pronounce it like "earn" but with a 'g' at the end... so well, "earng".
what do want computers to do? Play fetch like man's other best friend? Seems frivolous when you put it that way. Accept if you're one of those who believes the accurate recollection can somehow make us whole again,