damienteney.bsky.social
Research Scientist @ Idiap Research Institute.
Adjunct lecturer @ Australian Institute for ML.
Occasionally cycling across continents.
https://www.damienteney.info
34 posts
397 followers
289 following
Getting Started
Active Commenter
comment in response to
post
The deadline cycles seem very motivating to some students. And it creates guaranteed periods of downtime for advisors. Do we want to give up both?
I'm sure we can adapt though and I don't even know what I would prefer.
comment in response to
post
I feel it gets easier with experience. Most papers are incremental, so (1) if you miss one, you catch up with following ones in the same line of work. (2) Many papers can be "read" in a few minutes. Skim the abstract and a few other bits, and from experience you know/guess what the rest is about.
comment in response to
post
Ha good point! I didn't think of that.
comment in response to
post
I don't watch the news so I have no idea what this could be about.
comment in response to
post
No, the most minute details of implementation (and possibly hardware) have to be strictly identical to the authors' for this to be meaningful. But even then, if replicating the results of the paper (at least qualitatively) depend on the precise seeds being used, it means the results are junk.
comment in response to
post
What video is this about?
comment in response to
post
I wouldn't worry about it. There must be something like 20-40k authors waiting for the decisions. Of course there will be some in there who will complain and/or don't fully understand the process.
comment in response to
post
Very interesting! (still have to read the details). Looks like you missed our work last year that also showed how important initial wts are, especially with non-ReLU activations.
arxiv.org/abs/2403.02241
"Neural Redshift: Random Networks are not Random Functions"
(Fig. 14 from the appendix ⬇️)
comment in response to
post
Informatif et posé 👌👌 C'est le genre de commentaire qu'on devrait voir plus souvent dans les médias... 🤷
comment in response to
post
True for >1 significant digit.
comment in response to
post
Exploiting the ambiguity Christmas/NY 2024/2025? 🤔 Clever way to save on digits every 2 years!
comment in response to
post
Sorry to hear. It's something most men probably never think about. But with something like a couple of Advil PM (which you easily find at the airport), it's not like you're unconscious for 10 hours. You just really feel like dozing off.
comment in response to
post
I used to be like this, but I recently found out that just the eye mask + earplugs *+ sleeping pills* make the whole thing more bearable.
comment in response to
post
Could you make a few-sentence summary of the current state of kowledge (why it works, limits of applicability, etc) and open questions? 🙏🏼
comment in response to
post
And another example by @davidpicard.bsky.social
PoM: Efficient Image and Video Generation with the Polynomial Mixer
arxiv.org/abs/2411.12663
comment in response to
post
See e.g. arxiv.org/abs/2410.18613 that recently showed that we can replace softmax attention with alternatives that do not satisfy the properties we intuitively assign to it, and yet these models seem to work just as well! (2/2)
comment in response to
post
Do "people" really say that NNs are black boxes? Indeed they're anything but black boxes. Closed-source LLMs APIs are.
comment in response to
post
Glad to finally see someone take this seriously! This is a common gap in the CS curiculum for most students. Will you share the slides after the event for those who could not attend the conference? 🙏
comment in response to
post
👍 Conciseness with clarity is the hallmark of great writing.
comment in response to
post
We'd have a face and name to put onto "Reviewer 2". Everyone would suffer.
comment in response to
post
I think it will be awful. Junior people won't dare posting honest-but-negative reviews, some will avoid reviewing entirely for this reason, others will write excessively praiseful reviews. Aren't blind reviews in place for the same reason as secret ballots in democracies?
comment in response to
post
Did you think what this could do to a junior scientist/PhD student trying to build a network and reputation?
comment in response to
post
I'm colorblind too so I'm very careful with my choice of colors in charts. The best is to pick colors with different brightnesses such that they make sense even if they were printed in grayscale.
comment in response to
post
Of course, the raw data must be available somewhere. In some cases, just as a download could even be fine. I'm thinking of the kind of papers where, what matters is not absolute performance numbers, but imstead the qualitative differences across settings or methods.
comment in response to
post
This one is like naming your child Adolf. Without the negative connotation 😄 Some names aren't meant to be reused again for a while.
comment in response to
post
Having a desk on the ground floor looks as fun as working in a basement. Is it better IRL?
comment in response to
post
Haha granted that's a good effort.
Different story for this one👇
comment in response to
post
Most foods in the US are nothing like the original thing 😀🤷
As a Belgian I can certify that (American) Belgian waffles are unrecognizable to most Belgians.