Zhenhao Chen

Machine Learning PhD @MBZUAI

[Zhenhao Chen @DavisChan1998]

@LingjingKong Congratulations!!

[Gabriele Berton @gabriberton]

Libraries and tools that every deep learning project should use: loguru, tqdm, torchmetrics, einops, python 3.11, black. Optional: prettytable. Good for debugging: lovely_tensors. Any other ones I've missed?

Below a few words on each of them:

[Zhenhao Chen @DavisChan1998]

@tolusophy @MarkBaard @xwang_lk According to multiple sources and her tweet record, if I apply Bayesian here, I believe it’s not the first time she show such bias to Chinese. Please have a thorough check before defending. If only a word was misused, how could the community shows such disappointment?

[Zhenhao Chen @DavisChan1998]

I don’t think the overall reaction of the community is oversensitive. First, if the slides are offensive, she should give a proper response and apology to the community. But until now, nothing happens. Furthermore, of course people need to pick words to avoid offending! Free speech doesn’t imply people can use any word as they like. Mutual respect is the principle during communication.

[Zhenhao Chen @DavisChan1998]

Good one!

[Bo Wang @BoWang87]

Most MIT profs who I know are extremely honest and morally upright. I hope this is an exception.

[Zhenhao Chen @DavisChan1998]

The legend Prof. Michael I. Jordan gave an insightful talk titled: "Collaborative Learning, Information Asymmeteries, and Incentives" in MBZUAI this Friday. Hearing his unique perspective on the field's evolution and future directions was truly mind-blowing!

[Zhenhao Chen @DavisChan1998]

@NilsLukas7 @mbzuai Welcome on board, Dr. Nils. See u in Abu Dhabi soon! 🥳

[François Chollet @fchollet]

The main take-away here is that modeling the world is not the same as embedding observational data in an interpolative latent space (a curve).

Mind you, you *can* still use a big curve for physics predictions, especially for systems that are expensive to simulate (e.g. the weather). But to do this, you have to restrict yourself to a subspace where:

1. The future will be like the past (stationary regime).
2. The manifold hypothesis applies -- it should be representable as a curve.
3. You're able to collect a *dense* sample of training data.

There are many domains where you can verify these properties. It's just not feasible for "the visual world" in general. You're going to need a different kind of model for that.

[Zhenhao Chen @DavisChan1998]

@NielsRogge Thank you for your interest and appreciation of our work! We uploaded the data to Hugging face is because the dataset package is very convenient to use and it will save a lot of unnecessary time for AI researchers. Exactly as you said.

[Zhenhao Chen @DavisChan1998]

More examples of BenchLMM: (7/7)

[Zhenhao Chen @DavisChan1998]

We propose BenchLMM to assess the robustness of LMMs against three different styles: artistic image style, imaging sensor style, and application style, where each style has five sub-styles. (1/7)
Paper: https://t.co/5FKRagcRd6
Project website: https://t.co/QFYXHmYmzR

@_akhaliq

[Zhenhao Chen @DavisChan1998]

To this end, we devised a Style Prompt Enhancement (SPE) by introducing a text prompt. (6/7)

[Sundar Pichai @sundarpichai]

Search recorded its highest ever traffic in 25 years during the final of #FIFAWorldCup , it was like the entire world was searching about one thing!

[Zhenhao Chen @DavisChan1998]

🤣

[Yann LeCun @ylecun]

LeNet-5 implemented in Minecraft using Redstone dust.
Very meta.