Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Show HN: Boldly go where Gradient Descent has never gone before with DiscoGrad

github.com

Post reply on HN
Print view View original

61–67 of 67 posts

Re: Show HN: Boldly go where Gradient Descent has never gone before with DiscoGrad

#61

Post by usgroup »

frankling_ wrote: post #58
usgroup wrote: post #55

Earlier quoted context omitted.

Does that imply that every possible branch is optimised separately and then convoluted thereafter? If so, does it scale for very branchy programs? Do you have any comparisons to a Gibbs based approach for any of the use case examples?

The convolution is approximated via a form of sampling with additional bookkeeping at each encountered branch. How well that scales for deeply branching programs depends on the probabilities of the branches and the diversity in the output on the resulting paths, the worst case being a program where all branches are equally likely and each path generates an entirely different output (as in a hash function, for example…

Thank you. Your helpfulness and balanced evaluations are refreshing.

Re: Show HN: Boldly go where Gradient Descent has never gone before with DiscoGrad

#64

Post by big-chungus4 »

andyferris wrote: post #50

Pretty cool. It is somewhat like Julia’s autodiff ecosystem - which also handles native, branching code.

Julia's autodiff packages, as well as PyTorch, can differentiate through branching code -the gradients simply flow through whatever branch was used in forward pass. However, derivatives with respect to conditional values, such as `a` in `if i > a`, are mathematically zero. If you plot a graph of how function value depends on a conditional `if i > a`, it is flat with a single drop when `i` becomes bigger than `a`. DiscoGrad, on the other hand, doesn't use true mathematical derivatives, instead it calculates useful, smoothed gradient approximations for those conditionals.

Re: Show HN: Boldly go where Gradient Descent has never gone before with DiscoGrad

#65

Post by big-chungus4 »

zaitanz wrote: post #48

So this confuses me slightly and I am keen to know the advantage of using this. I work on projects that heavily use auto-differentiation for complex models. The models are defined by user input files at run-time, so the state and execution pathway of the model is unknown at compilation time. Would this help? Given that all auto-differentiation is an approximation anyways. I've found existing tooling (CppAD, ADMB, ADO…

In `if i > x`, derivative with respect to x is mathematically 0 at all points. DiscoGrad gives you a useful smooth approximation that is not 0 and lets the function learn those conditional values.

Re: Show HN: Boldly go where Gradient Descent has never gone before with DiscoGrad

#66

Post by dwrodri »

usgroup wrote: post #34
dwrodri wrote: post #19

This is the sort of thing I expected to see when Chris Lattner moved to Google and started working on the Swift for Tensorflow project. I am so grateful that someone is making it happen! I remember being taught how to write Prolog in University, and then being shown how close the relationship was between building something that parses a grammar and building something that generates valid examples of that grammar. Whe…

Not really. The world of Bayesian modelling has much fancier tools: Hamiltonian MC. See MC Stan. There’s also been Gibbs samplers and other techniques which support discrete decisions for donkeys years. You can write down just about anything as a BUGS model for example, but “identifying the model” —- finding the uniquely best parameters, even though it’s a globally optimisation —- is often very difficult. Gradient de…

I wrote a Gibbs Sampler to try and fit a Latent Dirichlet Allocation model on arXiv abstracts many moons ago! I'd probably have to start from primitive stuff if I were to give it another go today.

I agree with everything you've said so far: getting to the point where you can use gradient descent to solve your problem often requires simplifying your model down to the point where you're not sure how well it represents reality.

My lived experience--and perhaps this is just showing my ignorance--I've had a much harder time getting anything Bayesian to scale up to larger datasets and every time I've worked with graphical models it's just such a PITA compared to what we're seeing now where we can slap a Transformer Layer with embeddings and get a decent baseline. The Bitter Lesson has empowered the lazy, proverbially speaking.

Tensorflow has a GPU-accelerated implementation of Black Box Variational Inference, and I've been meaning to revisit that project for some time. No clue about their MC sampler implementations. Then I stumbled across https://www.connectedpapers.com/ and Twitter locked up it's API, so admittedly both of those took a lot of the wind out of my sail.

Currently saving up my money so that I can buy Kevin Murphy's (I think he's on here as murphyk) two new books that released not too long ago https://probml.github.io/pml-book/. The draft PDFs are on the website, but unfortunately I'm one of those people who can't push themselves to actually read a text if it's not something I can hold in my hands.

Re: Show HN: Boldly go where Gradient Descent has never gone before with DiscoGrad

#67

Post by dwrodri »

YeGoblynQueenne wrote: post #57
dwrodri wrote: post #19

This is the sort of thing I expected to see when Chris Lattner moved to Google and started working on the Swift for Tensorflow project. I am so grateful that someone is making it happen! I remember being taught how to write Prolog in University, and then being shown how close the relationship was between building something that parses a grammar and building something that generates valid examples of that grammar. Whe…

>> "If you can build a program which follows a set of rules, and the rules for that language can be differentiated, could you not code a simulation in that differentiable language and then identify the optimal policy using it's gradients?" What's a "policy" here? In optimal control (and reinforcement learning) a policy is a function from a set of states to a set of actions, each action a transition between states. In…

Thank for taking the time to explain such a worked out example. I was indeed picturing something a long the lines of "If you could write a program equivalent to a game where you solve a maze, could you produce a maze-solver program if the game were made in this runtime.
Post reply on HN