>> The course will be available for free online from early 2023.
Anyone aware of an in-depth intro-level text-based explanation of Stable Diffusion that covers the whole pipeline, including training on an extremely
limited dataset?
Sorry folks for not contributing to this thread earlier - didn't realise this popped up on HN while I was sleeping! I'm Jeremy, and I'll be running this course. Lemme know if you've got any questions about it, or anything related to the topic of DL, stable diffusion, etc.
Yes, but more from a "what do they do, and how do we code them" than "how are they derived mathematically". I think sampling methods might be the bit of this which is the hardest to fully understand given our prerequisites (just high school math) and may require a follow up course to cover more deeply.
This is exactly the kind of course I’ve wanted to do for some time now. Even before stable diffusion it felt like other media synthesis applications like StyleGAN were what I wanted to learn, but most machine learning courses focus on more traditional data science topics.
Of course you can start with a more traditional course and then learn something like stable diffusion afterwards, but as a newbie it’s quite hard to figure out where to even start. A full-fledged course that takes you exactly where you want to go is a lot easier and I think it can help learners to stay motivated because they have a clear goal in mind. If I want to learn how to create cool images, I want to spend as little time as possible predicting housing prices in the Bay Area.
> If I want to learn how to create cool images, I want to spend as little time as possible predicting housing prices in the Bay Area.
I think that's somewhat of a dangerous mindset to have. If you want to create cool images you can use pre-trained models and high-level APIs without needing to understand any of the internals.
But if you want to truly understand how these models work, you need to make effort to study the basics. Maybe not predicting housing prices, but learn the foundational math and primitives behind all of the components from the ground up (and the Diffusion models are a complex beast made up of many components). And getting an intuitive understanding of how models behave when you tune certain knobs takes much longer. Many researchers in the field have spent years developing their intuition of what works and what doesn't.
Both of these are fine, but I think I think we should stop encouraging people to be in the middle. Have courses that that promise "Learn Deep Learning / Transformers / Diffusion models in 7 days!" but then go on and teach you how to call blackbox APIs, giving you an illusion of knowledge and understanding where there is none. I don't know if this applies to this specific course, but there are a bunch of those out there, and highly recommend staying away from those. I know it's a hard sell in this modern instant gratification age, but if you actually want to understand something you need to put in some possibly hard work.
>> But if you want to truly understand how these models work, you need to make effort to study the basics.
I was very confused by this in the beginning of my journey. I was trying to learn everything involved with ML/DL, but in the end everything is already implemented with APIs, and your boss doesnt care if you know how to implement a MLP from scratch or if you use Tensorflow.
My (poor) analogy is: you don't need to know how a car works (or how to build one) in every detail to drive it. When I understood it, it was liberating.
I agree, and I think that's the first use case I mentioned above. You can use Deep Learning libraries and pre-made models without a deep understanding of anything and you get some nice results. And that's great.
What's not so great is the huge number of people believing to understand something when they don't, i.e. the illusion of knowledge they're getting from some of these marketing-driven courses and MOOCs. I see that in job applications. Every resume has "Deep Learning, PyTorch, Tensorflow" on it now, but if you ask them why something works (or why a variation may not work) these candidates have no idea. And for some jobs that's totally fine, but for other jobs it's not. And the problem is when you can't tell the difference.
It's kind of like putting "compilers" on your resume because you've managed to run gcc.
It's the opposite of leetcode because it tests understanding, not memorization. For example, you could ask why position embeddings are necessary and what would happen without, reasoning behind certain terms in an objective function and what would happen without them, which part of an architecture the bottleneck for convergence is, intuitively what tuning a certain hyperparameter does, show them a timeseries of gradients for something that doesn't convergence and ask what's wrong with them, etc.
I'm just making these up because the questions we previously asked were domain-specific to our applications, e.g. "why is this specific learning objective hard" or "what would you modify to help generalization in case X"
These questions are very easy to talk about for someone with a strong ML background. They may not always know the answer and often there is no right answer, but they can make reasonable guesses and have a thought process around it. Someone who just took a MOOC likely has no idea how to even approach the question.
> I don't know if this applies to this specific course, but there are a bunch of those out there, and highly recommend staying away from those.
fast.ai do stuff pretty well. FWIW, I did one of their earlier free courses and, as a maths grad, got my fill of maths learning as well as my fill of practical 'doing stuff with ML' stuff. If I didn't have my plate full I'd probably pay the 500 quid or whatever to do this course now rather than wait for the free version.
Sounds like Emad Mostaque of Stability AI / stable diffusion thinks this course probably won't fall into "do this, no understanding needed" trap (I'm not contradicting anything you said here).
AFAIK Emad Mostaque is not (yet) an AI expert at all, he's a rich guy (former hedge fund manager) building a business that provides the funding for AI experts to do their thing. Stable diffusion itself was built by a team of academics [1], Emad is not a coauthor. Not to take away anything from what he's accomplished -- it's quite incredible -- but it doesn't mean he knows how to (learn to) build AI systems or do AI research himself.
I am not a coauthor but do know a decent amount about AI systems from a maths and computer science degree from Oxford, couple decades coding and being lead architect on https://hai.stanford.edu/watch-caiac amongst other stuff :)
Originally took a break from being a hedge fund manager to build AI lit review systems to investigate ASD etiology for my son along with neurotransmitter pathway analysis to repurpose medication (with medical oversight) to help ameliorate his more severe symptoms.
Had 100% satisfaction from programmers with some math knowledge trying fast.ai and members of team active there, really nice take off point into a massive sector.
It digs nicely into the principles and is not a surface level course. The stable diffusion one will need some good work to get through.
But yeah my job now is to get billions of dollars into open source AI to make the world happier, happy to do my best and let the smart and diligent folk buidl.
Thanks for the response! Sorry for underestimating your background based on what little I had read/heard about you. I appreciate and respect what you're doing.
He is happy now on an n=1 case. Potentiating GABA really helped with him, with other folk it may be balancing glutamate. Our medical system isn't really set up for that which is why I designed and launched CAIAC, it assumes ergodicity in the population while everyone is kinda individual, particularly for these multi-systemic conditions.
Will be aggressively investing in this area and making the output available openly next year after our education launch.
It’s a bit dated now, but the DeepLearning.ai GANs specialization covers topics through StyleGAN. If you have no experience with ML at all I would probably start with the first two courses of their Deep Learning specialization and then jump into the GANs specialization.
I would also highly recommend FastAI’s Deep Learning for Coders (and their new course that came out this year). You’ll start immediately with some cool applications (basic image recognition and NLP) and then drill down from there to learn how they work in detail.
It’s set up such that you can learn as much as you want (basics with no depth: first chapter; basic understanding of how a neural network is trained with SGD: first four chapters; understanding of decision trees, LSTMs, and CNNs: first half; detailed understanding of how to build everything from scratch: whole book).
> in depth course that started right from the foundations—implementing and GPU-optimising matrix multiplications and initialisations—and covered from scratch implementations of all the key applications of the fastai library.
I haven't taken the course, but that sounds like a horrible place to start a course on understanding deep learning. GPU matrix operations are literally an implementation detail.
I think the proper way to teach deep learning "from scratch" would be:
1.) show simple example of regression using high level library
2.) implement same regression by writing a simple neutral network from scratch (explain going through, multiplying weights, adding biases, applying activation function, calculating loss, back propagation).
3. use NN on more complicated problem with more parameters and a larger training set, so that user sees they've hit a wall in performance, and now implementation needs to be optimized
4. at this point, say okay, our implementation of looping and multiplying can be done much faster with matrix multiplication on GPU, and even faster parallelized across GPUs on a network. If you're interested in that, here is an optional fork in the course that gets into specifics. Anything after this point will assume that implementation of NN calls will be using these techniques under the hood
95% should have skipped step 4 and only revisited if they become interested in this for a specific reason. To start with it is crazy. It's like starting a course about flying by explaining how certain composites allowed us to transition from propellers to jets, and let's dive into how those composites are made.
> Three years ago we pioneered Deep Learning from the Foundations, an in depth course that started right from the foundations—implementing and GPU-optimising matrix multiplications and initialisations
Deep Learning from the Foundations IS part 2. They are not talking about how part 1 starts. They are talking about how part 2 starts. This mirrors how their book is written, with the content typically in Deep Learning for Foundations being the second half of the book. I've taken the courses every year, it's always been like that.
You should consider taking a look at part 1 of the course if you want to understand the starting curriculum: https://course.fast.ai/
'Deep learning from the Foundations' was the previous version of this new course, but both built on their respective part 1s. So part 1 is 'practical deep learning for coders', very good intro to DL for getting things done that starts with big wins using pretrained models and then digs into them enough that you know how to make them better. Part 2 is for those wanting to go all the way to implementing the ideas 'from scratch', reading papers and diving into advanced topics.
The course basically goes through the exact steps you described. The main "GPU optimisation" is to basically say "okay, now that we know how to implement matrix multiplication, let's use the optimised pytorch implementation instead".
When I teach people, the most effective thing I find is to give them small usable projects at the end of each little milestone.
That helps them have something they can use right away, which helps with a few things:
- at each milestone, they have a distinct goal that’s reachable
- they can understand concrete use cases immediately
- they get a little dopamine hit if satisfaction of having completed something.
Whenever you’re doing a course that doesn’t structure itself that way, it’s good to try and break down the components and set yourself little tasks as you go.
This course looks interesting. My only concern is that I don't have real experience with NLP. Anybody can recommend resources to get to speed on this pre-requisite? My NLP knowledge is very basic.
Anyone aware of an in-depth intro-level text-based explanation of Stable Diffusion that covers the whole pipeline, including training on an extremely limited dataset?
Here’s an example, but open to suggestions too:
https://huggingface.co/blog/stable_diffusion