Timeline for Training an energy-based model (EBM) using MCMC
Current License: CC BY-SA 4.0
9 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Dec 17, 2022 at 22:37 | comment | added | Dror Speiser | Just in case you haven't seen them, I'll quickly add that Song has a few nice YouTube videos from earlier this year about this research and paper, and there's also Ermon's IAS talk from two years ago. | |
| Dec 17, 2022 at 22:29 | comment | added | Dror Speiser | $p_{data} $ just means the "empirical distribution", it's the finite uniform distribution over the samples we have. This the same as in normal MLE up to wording; you're probably familiar with the more common wording "find $\theta$ that maximizes $p(\text{data}|\theta)$". And yes to first and second questions, where "for all $x$" means all the samples in our data. | |
| Dec 17, 2022 at 10:53 | comment | added | voila | @DrorSpeiser I see, so we first approximate $\triangledown_\theta \log p_\theta(x)$ for all x? Then we take a mean over gradients at all $x$? But in order to take the mean, we would need to know $p_{\rm{data}}$, is the underlying data distribution known before hand? Thanks. | |
| Dec 17, 2022 at 9:47 | comment | added | Dror Speiser | Yes, the idea is to maximize $\mathbb{E}_{x\sim p_{data}} \log(p_\theta(x)) $. So what we're really doing with the gradient ascent is taking the mean over gradients at all $x$ in our data | |
| Dec 17, 2022 at 0:52 | comment | added | voila | @DrorSpeiser I understand that log likelihood is maximized using gradient ascent. In a vanilla Maximum Likelihood estimation, there is a notion of observed data. Do we observe any data here? Thanks. | |
| Dec 15, 2022 at 16:29 | comment | added | voila | @DrorSpeiser I see, that makes sense, thanks! | |
| Dec 15, 2022 at 9:58 | history | edited | YCor | CC BY-SA 4.0 |
formatting, added tag
|
| Dec 15, 2022 at 8:32 | comment | added | Dror Speiser | Given approximations to the gradient of log likelihood you can get local maxima of the log likelihood as a function of $\theta$ - start anywhere, and then do gradient accent. At no point do you actually know what the log likelihood is, but you know you're getting better (stochastically). | |
| Dec 15, 2022 at 5:09 | history | asked | voila | CC BY-SA 4.0 |