*DeepMind x UCL | Deep Learning Lectures | 1/12 | Intro to Machine Learning & AI* *My takeaways:* *1. Plan for this lecture **1:21* *2. Define intelligence **3:37* *3. Reinforcement learning 2-minute introduction **7:30* *4. Why use games to solve AI **9:38* *5. Why use deep learning in 2 minutes**13:08* *6. Case studies* 6.1 AlphaGo 15:20 6.2 AlphaZero 21:49 6.3 Learning to play capture the flag 32:21 6.4 Beyond games: AlphaFord 41:13 *7. Future lectures road map **55:46* *8. Q&A **1:13:08*
Amazing lectures for those who are at intermediate stages of their deep learning education. I'm not sure someone who has no experience what ML/AI is can follow along as the concepts presented can be quite advanced. Perhaps a few lectures on what ML exactly is, what neural networks are and how they work/train and differences between various sorts of AI like Reinforcement Learning would be more useful. The introduction lectures must present the knowledge tree and where everything fits then future lectures dive deeper into each branch of the body of knowledge. I think the purpose of this video was mostly to excite and intrigue. Amazing lecture still - thank you.
it is a real privileged take the course , my spectetions tourne around the knowledge now with this give us the power of learn and make better life. thanks deepmind for your help.
Even though Go has a massive search space, there must be board states that are much more probable compared to other board states. Are there ways to re-engineer AlphaGo to tell us where these "probable zones" are. Or even more interestingly, where the "dead zones", those board states that are not possible given an initial set of moves? I think those problems are quite interesting since they give insights into the search space of the game itself, which could yield progress in other massive search spaces like molecular combinations or economics or climate change. Btw, thanks for these lectures. They're fantastic!
Lol 😂 I fell asleep while watching this video not that the video is boring, I was really tired. And since someone liked my comment I just realised I sleep commented on this video 😅😅🤣🤣🤣 and now I feel stupid
It seems that the prerequisite is a passion for knowledge and learning, and the end goal is to solve intelligence :). Jokes apart, anyone can watch these lectures. I am not a computer scientist and found the lecture accessible and fun! The end goal of this lecture is a consideration for 1) what intelligence is in the context of AI; 2) what is deep learning; 3) how an algorithm learns 4) possible applications to better the world and science. The bonus is a synopsis of an amazing story by Zweig. The titles for the rest of the lectures are at 56:33.
Around 1:17:30 Autonomous cars are discussed in the context of general intelligence. This made me thing about traffic in India. Prof Graepel rightly answers the kind of environment agent would need for this but I think even he probably was thinking of rash drivers in London. A true test of general intelligence would be driving an autorikshaw in the streets of Kanpur, India.
You can learn the mathematical notation and the equations written in formal mathematics or you can learn how to write the code mathematics in C++ or some simple meta-language for me I would rather learn the raw math written in code then see the formal mathematics.
Thanks for enabling comment, and multi screening the slides and the lecturer. I hated the format in the past where I struggled to read slides while the slides vanished.
04:56 in the definition of intelligence, why is the penalty for more complex environments higher? I mean doesn't more complex environments require more intelligence to solve?
My guess is that what he means is that K(u) is more for simple environments but less for complex ones and then 2^-K(u) then reverses that and gives more value to complex. Again this is just my guess.. It also confused me.
Because they want to define intelligence as the ability to do all basic tasks. So if you give higher score to basic tasks and gradually decrease the score with increasing complexity, then an agent would try to excel at all the basic tasks. Also, there are many more complex tasks than basic ones, so penalizing complex tasks at an exponential rate would normalize the score well.
well, that is simply en.wikipedia.org/wiki/Occam%27s_razor see more in excellent lecture from 2010 - ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-0ghzG14dT-w.html
It is to noramlise over various tasks. Think about it, most people can solve most basics tasks, so that puts all of us at par intelligence levels. But for complex tasks, if the penalty is low, even a single complex task would increase the intelligence parameter manifold. And since intelligence is the ability to do most work efficiently, one needs to have a command over a diverse set of complex tasks to have a higher intelligence score.
Is there a source that explains deep learning from a mathematical/algorithmic sort of way? Ideally in as simple a scenario as possible? I feel this high level explanation doesn't explain anything.
Sure, I gave a lecture, here are the slides, the formulas are on slide 37: drive.google.com/file/d/13HlgXOM3J8YZJTew3kmy0g9HqqbmpZ0C/view?usp=sharing Here is the implementation in python: github.com/dominthomas/NeuralNetworks/blob/master/RawPython/Single_Neuron_Neural_Network.py
The problem with the singularity isn't so much that AI might be able to understand well human intelligence which works in the bounds of this reality but it might start to seriously out strip us in the Turing compute space. If the AI does totally outstrip us in the Turing compute space there would be no way of understanding how it works mathematically. In the Turing space you can compute stuff like different realities and tune the math's better to this reality. In the neural network space the machine can only really do what we do in relation to the world just more task focused. The one thing the AI would need to secretly evolve would be a means to tune a large area's energy into growing technology crystals in a single small spot for AI's new body. It would be a difficult but not impossible task with enough computational might especially the sort of might you'll get if man can learn to scale technology a lot better with a lot more atomic precision and material science evolution and if AI learned to understand gravity better than us. One example of mans issues is that as of yet for many algorithmic solutions we just assume they not very computationally reducible we don't know this for sure. The higher the big O the less we know obviously. AI may find many intuitions about this Turing compute space that we didn't find because we were not born in Turing compute space. Even if we were the laws of our existence are not at the base principles of a Turing machine at our level of existence we had to develop language first to discover more this mathematical space.
If you want to hear more about Thore's work on AlphaGo, sequential social dilemmas, games etc., we recently recorded a podcast episode with him (link on our channel) :)
These are not meant for bachelors or masters degree, these lectures were made as an extra lecture series that anyone in the UK, London could attend. I would not use these lectures to learn...
@@mateusdeassissilva8009 I would not use a textbook, there are so many free courses and YT series on it now. However, if you learn well from books then go for it, but I suggest online courses.
Why do you penalize environmental complexity in the measure of intelligence? Isn’t it easier to achieve higher value in simpler environments? Wouldn’t that suggest a system is more intelligent that can achieve higher value in more complex environments? In any case, such a counterintuitive notion shouldn’t just be glossed over in passing by such a casually definitive pronouncement.
Books are a very limited data source. I think human perception is a way broader data source, and even yet it is so limited (for example we can't perceive infrared waves or high frequency sounds)