#2. December, 2022

AI hype, supervoxel segmentation, publication, and some recommendations

Hello, and happy holidays !

After Galactica last month, the AI hype-cycle shifted to ChatGPT this month, which I’m not particularly thrilled about. I wrote about it on my French-speaking blog because the news article that annoyed me the most about it was in “Le Soir”. The gist of it is that ChatGPT is a very nice toy, but that reporting on it as if it was a revolutionary new tool that’s going to rival Google is completely missing the point of how it works, and what it can or cannot do. At some point I want to write about the more political aspects of OpenAI and the damages done by “effective altruism”, “longtermism” and associated philosophies on “AI” research, but that’s going to remain future works for now. For a general look at why “large language models” are overhyped and risky, the best resource is probably the famous “Stochastic Parrots” paper by Emily Bender et al.

On my postdoctoral front I spent a lot of time in December on planning, reporting, in meetings and improving the codebase, so I had fewer interesting things to write about on the research blog. The recap is therefore going to be quite short, with two published posts:

• In [Devlog] Supervoxel CT liver segmentation, I play around with “supervoxels”, a classic image analysis technique, to try to segment a mouse liver in a CT scan. I hadn’t done some “proper” image analysis in a while, so it was fun to experiment a bit with more “hands-on” methods than what we typically do in deep learning pipelines. The results are not too bad either!

• In [Publication] Review of digital pathology segmentation challenges, I summarize a paper from my PhD that just came out in the journal Computerized Medical Imaging and Graphics. It was a big piece from the last years of the PhD, so I’m glad that’s finally out.

Some recommandations from across the web:

I have been following two YouTube channels for a while that have really helped me improve my “developer” skills:

• mCoding is a channel which mostly includes short videos with very good explanations and demonstrations of some subtle and/or less known aspects of how Python works. There are not always immediate practical applications for the tricks that he shows, but it often makes me think a little bit beyond the code that I write, at the code that the computer actually executes.

• ArjanCodes is a channel that focuses more on higher-level software design: how to structure the code in ways that make it more testable and easier to maintain in the long term.

That’s it for this month’s recap, and for the year 2022.

Wishing everyone a good start to 2023,

Adrien.