@alex107568 Indeed, even when developing a prototype, writing unit tests can help you save time. Ofc it depends on how granular you want to go with your tests, but I usually find it helpful. For products or even for MVPs, tests are ofc mandatory.
I'm surprised that unit testing is not the norm in ML. Just yesterday I found two bugs in some new code I was writing, and I would have never found them by just "visualizing results" or "looking at the training curves", etc. Then again, a case could be made that, "if a bug 1/n
@mudalmis AI startups that went out of stealth since last year" list. I can even restrict the target even more, to startups that went out of stealth this year (but then it becomes even harder to discern between promising and hypey ones). 2/2
@mudalmis Perfect is the enemy of good 🙂 even an incomplete list could be good enough (also, note that I'm restricting the search to startups that went out of stealth no earlier than 2021). The Web is full of "10/20 best" lists, surely it would be possible to cook up a "20 best 1/2
Excited to share DiffDock, new non-Euclidean diffusion for molecular docking! In PDBBind, standard benchmark, DiffDock outperforms by a huge margin (38% vs 23%) the previous state-of-the-art methods that were based on expensive search!
A thread! 👇
We've found a new way to prompt language models that improves their ability to answer complex questions
Our Self-ask prompt first has the model ask and answer simpler subquestions. This structure makes it easy to integrate Google Search into an LM. Watch our demo with GPT-3 🧵⬇️
Excited to announce our new work showing the robustness of conformal prediction to label noise!
If you calibrate on noisy labels, you typically get valid, *conservative* coverage on clean labels at test-time. So if you use conformal you're already safe against noisy labels 🛡️ twitter.com/BatEinbinder/s…
Get faster, more flexible inference on GPUs using our newly open-sourced AITemplate, a revolutionary new inference engine that delivers up to 12X performance improvements on NVIDIA GPUs & 4X on AMD GPUs compared to eager-mode within Pytorch.
Learn more: bit.ly/3rl8F3b
Mais quel courage phénoménal ! Ces filles sont jeunes, très jeunes. Les lycéennes et les collégiennes rejoignent également le mouvement, pour la liberté, contre l’obscurantisme. C’est une révolution féministe. #Iran#MahsaAmini
@Aaroth@crispy_jung It could be nice to compare the two methods in such a setting! I always found group-conformal split CP to require calibration sets so big to be useless in practice. I’ll see if I managed to set something up. Compliments for the paper, really cool!
1/4 Today we released two new link prediction models developed over the past 12m. We solve many problems that prevent GNNs succeeding at link prediction by adding a message passing mechanism that uses sketches of subgraphs.
ArXiv arxiv.org/abs/2209.15594: Analysis of SGD. Sharpness (largest eigenvalue of the Hessian) steadily increases during training until instability cutoff 2/η then it hovers around 2/η. Training loss still decreases. Reason: self-stabilization via cubic term in Taylor expansion.