TranscriptAILearnerFromYoutube / TwoTranscriptQuotesFromIlyaSutskever.md
awacke1's picture
Upload 4 files
f39338a
|
raw
history blame
No virus
3.32 kB

https://www.youtube.com/watch?v=9EN_HoEk3KY&t=172s

1:42 program the does very very well on your data then you will achieve the best 1:48 generalization possible with a little bit of modification you can turn it into a precise theorem 1:54 and on a very intuitive level it's easy to see what it should be the case if you 2:01 have some data and you're able to find a shorter program which generates this 2:06 data then you've essentially extracted all the all conceivable regularity from 2:11 this data into your program and then you can use these objects to make the best predictions possible like if if you have 2:19 data which is so complex but there is no way to express it as a shorter program 2:25 then it means that your data is totally random there is no way to extract any regularity from it whatsoever now there 2:32 is little known mathematical theory behind this and the proofs of these statements actually not even that hard 2:38 but the one minor slight disappointment is that it's actually not possible at 2:44 least given today's tools and understanding to find the best short program that

https://youtu.be/9EN_HoEk3KY?t=442 5 to talk a little bit about reinforcement learning so reinforcement learning is a framework it's a framework of evaluating 6:53 agents in their ability to achieve goals and complicated stochastic environments 6:58 you've got an agent which is plugged into an environment as shown in the figure right here and for any given 7:06 agent you can simply run it many times and compute its average reward now the 7:13 thing that's interesting about the reinforcement learning framework is that there exist interesting useful 7:20 reinforcement learning algorithms the framework existed for a long time it 7:25 became interesting once we realized that good algorithms exist now these are there are perfect algorithms but they 7:31 are good enough to do interesting things and all you want the mathematical 7:37 problem is one where you need to maximize the expected reward now one 7:44 important way in which the reinforcement learning framework is not quite complete is that it assumes that the reward is 7:50 given by the environment you see this picture the agent sends an action while 7:56 the reward sends it an observation in a both the observation and the reward backwards that's what the environment 8:01 communicates back the way in which this is not the case in the real world is that we figure out 8:11 what the reward is from the observation we reward ourselves we are not told 8:16 environment doesn't say hey here's some negative reward it's our interpretation over census that lets us determine what 8:23 the reward is and there is only one real true reward in life and this is 8:28 existence or nonexistence and everything else is a corollary of that so well what 8:35 should our agent be you already know the answer should be a neural network because whenever you want to do 8:41 something dense it's going to be a neural network and you want the agent to map observations to actions so you let 8:47 it be parametrized with a neural net and you apply learning algorithm so I want to explain to you how reinforcement 8:53 learning works this is model free reinforcement learning the reinforcement learning has actually been used in practice everywhere but it's