When we set out to learn a function or some property of it (like its maximum), we hope it is differentiable , because that means we have at our disposal a host of well-studied, and often fast, techniques. But sometimes we are not so lucky - and then there are broadly two options: (a) use a technique that doesn’t rely on differentiability, e.g., Bayesian Optimization, or (b) use an approximation…
I attended NeurIPS’24 virtually, and I was happy to see that they had two tutorials on topics that I care about. One was on evaluating LLMs, and the other one was on decoding-time strategies. This post covers the former. I have been meaning to publish this for a while, but this languished as a draft for a long time while life got in the way. Well. Introduction Tutorial Intro - Irina Sigler Quality…
I totally stole the title from a paper (Attenberg & Provost, 2011) . In theory, Active Learning (AL) is a tremendous idea. You need labeled data, but your kind of labeling comes at a cost, e.g., you need to obtain them from a domain expert. Now, lets say, your goal is to use this labeled data to train a classifier that gets to a held-out accuracy of \(90\%\). If you randomly sampled points to…
Jensen’s inequality finds widespread application in mathematical proofs. I am fond of a particular intuitive explanation of it, which doesn’t seem to be very popular. I will try to present it in brief here. I am not sure when this argument originated, but Google does turn up a paper (Needham, 1993) . Even if this is not the source, it is a good reference. On a related note, the author of the…
This post continues our discussion on BayesOpt. This is part-2 of a two-part series . Now we take a look at the other pillar BayesOpt rests on: acquisition functions. My goal is to provide a flavor by looking at a few of them. I’ll go into depth for a couple; this would help us appreciate the role of GPs in conveniently calculating acquisition values. For the rest I’ll provide an overview.…
The real reason I like Bayesian Optimization: lots of pretty pictures! If I wanted to sell you on the idea of Bayesian Optimization (BayesOpt) , I’d just list some of its applications: Hyperparameter Optimization (HPO) (Turner et al., 2021) . Neural Architecture Search (NAS) (White et al., 2021) . Molecule discovery (Gómez-Bombarelli et al., 2018) . Liquid chromatography (Boelrijk et al., 2023) .…
Generative Models have been all the rage in AI lately, be it image generators like Stable Diffusion or text generators like ChatGPT . These are examples of fairly sophisticated generative systems. But whittled down to basics, they are a means to: (a) concisely represent patterns in data, in a way that … (b) they can generate later what they have “seen”. A bit like an artist who witnesses a scenery…
Moving to a new place can be hectic and tiresome. I am moving my blog, from here , and it’s none of those. 1 /s I tend towards writing technical posts when I tend towards writing at all these days, and blogger doesn’t give me the presentation options I need. So, for now, its GitHub pages, but with my own domain. That way, if I decide to move again, my (almost non-existent) readers won’t be sent…