[Graham]

Has anyone else been noticing all of the news about AI recently? Machine learning is something that fascinates me, personally.

Here are two videos I've seen recently that intrigued me:

Googles deep mind division creates AlphaGo, it beats the worlds top Go player not though brute force, but learning
https://youtu.be/TnUYcTuZJpM

A program written by SethBling that plays super Mario world. He gives a great demonstration of how it learns.
https://youtu.be/qv6UVOQ0F44

[Dr Lancer-X]

I've only really been getting into machine learning over the past 2 years or so but I have found it to be a really awesome research field. To me the most interesting thing about deep learning is the giant strides that have been made in image recognition using these techniques - there is absolutely no comparison between modern convolutional neural nets and the content-based image recognition techniques of the past. In other areas the benefits of neural nets are, I think, a bit more dubious. One of the reasons I think they are so attractive is that they free the data scientist from careful feature engineering - but it's not really clear that they are necessarily better than more traditional machine learning approaches. Most of the improvements we've seen have been as a result of getting access to better and larger data sets and more compute, and this has aligned with a resurgence in interest in neural nets as a result of how well CNNs are doing.

[Graham]

Dr Lancer-X, on 07 December 2016 - 01:51 PM, said:

I've only really been getting into machine learning over the past 2 years or so but I have found it to be a really awesome research field. To me the most interesting thing about deep learning is the giant strides that have been made in image recognition using these techniques - there is absolutely no comparison between modern convolutional neural nets and the content-based image recognition techniques of the past. In other areas the benefits of neural nets are, I think, a bit more dubious. One of the reasons I think they are so attractive is that they free the data scientist from careful feature engineering - but it's not really clear that they are necessarily better than more traditional machine learning approaches. Most of the improvements we've seen have been as a result of getting access to better and larger data sets and more compute, and this has aligned with a resurgence in interest in neural nets as a result of how well CNNs are doing.

That's very interesting and I wasn't aware of that aspect of it at all. I was, however, curious about how AlphaGo was "reading" the board. Very cool stuff here.

[Graham]

Here is a fascinating video explaining CNN's for anyone else who is clueless about them CNN's

[Dr Lancer-X]

Graham, on 08 December 2016 - 11:29 AM, said:

That's very interesting and I wasn't aware of that aspect of it at all. I was, however, curious about how AlphaGo was "reading" the board. Very cool stuff here.

To me the most fascinating part is how the training is performed. They did train AlphaGo on a corpus of many recorded Go games by champion Go players - so, supervised learning. However, the bulk of the training was through reinforcement learning, which was performed by having two different AlphaGo systems. They would play against each other and system 1's results were backpropagated through system 2's neural net and system 2's results were backpropagated through system 1's neural net. So it just got better by playing against itself many many times. The initial supervised learning wasn't even necessary, but reportedly it helped greatly reduce the amount of time training took.

[Graham]

Dr Lancer-X, on 09 December 2016 - 06:12 PM, said:

To me the most fascinating part is how the training is performed. They did train AlphaGo on a corpus of many recorded Go games by champion Go players - so, supervised learning. However, the bulk of the training was through reinforcement learning, which was performed by having two different AlphaGo systems. They would play against each other and system 1's results were backpropagated through system 2's neural net and system 2's results were backpropagated through system 1's neural net. So it just got better by playing against itself many many times. The initial supervised learning wasn't even necessary, but reportedly it helped greatly reduce the amount of time training took.

Yes I heard something about that in one of the videos I watched. Learning like a child would. I think it played something like 1000 games against itself?
SethBlings software played Super Mario World many times, saving the results of the small adjustments it made when it made progress over the previous build. But it took far fewer iterations to become proficient at that game :-)

[Graham]

This is the MarI/O script, modified to play Super Mario Kart. Watching the different species try different things and learn from them is so cool.
https://youtu.be/S9Y_I9vY8Qw

This post has been edited by Graham: 10 December 2016 - 04:28 PM