This video is a great improvement from the previous videos. As a beginner this is the only video that i understood 100%(at least amonst the ones that i have watched). keep up the great job!
You are awesome! You are leading a new young generation of AI developers in to the future! Keep up the good work. Do you have any recommendations for good books on AI, Deeplearning?
TITAN V ....Please do a review & benchmark of Titan V from a Deep Learning perspective!!! How does is compare to 4x1080ti for the same price? Do the 640 tensor cores really scale up to a 9 times performance increase in real life? Is 12 GB RAM enough? What are your thoughts?
*Hey Siraj!* Really appreciate the videos dude. I have a request for a video. Could you maybe code up a crawler that finds the homepage links (channel website, facebook ....) of a youtube channel, scrapes them and analyze them and determine which country the channel belongs to/ primarily caters to? *Thanks!*
Hey Sriraj, is there a way to apply these concepts to a real business problem? If you have some info on that, please post links. Great video, thank you
Honestly do it like predator, use two points calculate the vector, set boundaries for bounce, because an inteligent AI do trend analys. Not just plot intermediate values. That AI will learn alot of faster because it understand the room and angles. Two points is enough for each player. Now you do not need to collect meaningless sampledata. When and if i build AI it will try learn concepts. Yeah i understand it is just an example ;), but would it not be more interesting if the AI could learn the concept of the room, the concept of moving and the concept of speed and the concept of angle. Because that data is translateable into knowledge/information that can be used again, and then the AI actually learned something not just sampled rawdata without comprehension. "One could even do an elaborate pong where the room change and have angeled obstacles in the course of ball between players. And if the obstacles move and have paths, well it is starting to get hard calculate for the AI :)", there is no awareness or insight in reach of that automata LoL
You're contradicting his statement of not making assumptions about Pong at 2:56. As soon as you say "you can do ... which will make it learn faster" that means you narrowed down and can no longer generalize. Always keep that in mind.
AlphaGo and zero do/did exploit that humans narrowed down the searchspace using assumption "evidently they were wrong", but Go is very hard to conceptualise for a human there is no optimal approach, it is game theory "react on the other player", so there the dataset in not fully conceptualised and explored. So there is a difference between problems that can be conceptualised easily and problems that can't like chess and go or politics and economy "without gigantic memory for a human", and an explored tree for an AI. I do not know if they publish online how many games Alpah GoZero played so far but it must be 100 of billions. Actually they should publish online so we can see how many games it plays a minute."It is playing right now and forever until someone push the stop button"
If i have limited data for my model, what can i do to improve it? Generate new images using GAN? OR use tranfer learning (used style transfer) +Siraj Raval
I don't understand the difference between updating the weights for each move once we get a reward, without knowing which move actually got us a positive reward, and doing one move, "wait" and update the weights of the move once we get a reward. I the second case are we only performing one move really? If not, as I guess, we probably want to update the weights for all the moves which in my understanding is the same as the first setting.. Can anyone clarify?
Siraj, please do something about architecture. Now i know something about machine learning and so on. But what is the best way to connect them? For Example: Newsarticles, i bring it to Elasticsearch/Solr maybe MongoDB. With python i do an word2vec with spaCy for example, i dicided to bring it in a graphical database Neo4j. But how the whole system interact with each other in the best way? What is the best architecture?
Axel Kaliff he have done it...try watching the math of intelligence playlist there he have a video on CNN which are popularly used for image classifier
I don't entirely agree with your statements around the minute mark about compute power and large datasets being drivers of recent AI progress. AlphaGo zero used far less compute than the first AlphaGo(4 TPUs vs 176 GPUs). It also used way less data than the original AlphaGo(only data generated from gameplay vs expert play + data generated from gameplay). But I guess it depends how you define recent
Yep. It depends on you define recent. Andrew Ng, cites data sets and computing power as being the reason for the thawing of the most recent AI winter, and it seems everyone else repeats the same thing. But I think you are 100% right. The future of AI is going to have very little to do with even bigger data sets and massive GPU sets.
The script is really similar to Karpathy's blog (karpathy.github.io/2016/05/31/rl/), so I think there should be more credit given to him (maybe an in-video citation, rather than just being in the description?).
Hi Siraj, thanks for your introduction about the pong and DQN. I have re-implemented the algorithm but it seems not working so well for 12 hours training. I wonder whether there is some problems, and could you please help to review my code? Thank you so much. Request for code review: github.com/JieMEI1994/HelloWorld