Olivier Grisel

Engineer at @probabl_ai, scikit-learn developer.

More active at:

[Olivier Grisel @ogrisel]

Here is the recording of the presentation:

https://t.co/J4XSc8wTkY

[Olivier Grisel @ogrisel]

I will introduce theoretical considerations such as how to assess the quality of probabilistic classifiers, and conversely, how the choice of the performance metric impacts the ability of the hyper-parameter tuning procedure to reliably find good models.

[Olivier Grisel @ogrisel]

For the record, here are my slides:

https://t.co/FHHIdhWN8v

The video recording should follow hopefully soon-ish.

[Olivier Grisel @ogrisel]

@fhuszar @andrewgwils I also found:

https://t.co/1VBczEOIif

and: https://t.co/1K0boZaota

[Olivier Grisel @ogrisel]

@DHolzmueller I wonder if schedule-free AdamW could help here.

[Olivier Grisel @ogrisel]

@glouppe Maybe this is just a copy and paste error.

[Olivier Grisel @ogrisel]

@jn2clark How many epochs? Maybe ADOPT works best (as an optimizer) but makes it easier to overfit as a result? Strong regularization might be needed (e.g. dropout or SAM-style updates?)

[Olivier Grisel @ogrisel]

@nworbmot Indeed, I read some of the details in the links of the thread after asking :) I suppose this kind of energy storage can help justify the increase of demand side flexibility from 5% to 10%.

[Olivier Grisel @ogrisel]

@PreetumNakkiran Or "creativity".

[Olivier Grisel @ogrisel]

I am looking forward to giving a keynote presentation at @PyDataParis 2024, September 25-26, Cité des Sciences. This will be an opportunity to reflect on the meaning of "noise" and predictive uncertainty in machine learning applications.

[Olivier Grisel @ogrisel]

@benjaminwalker @srush_nlp But one could argue this is just a UI problem: the LLM could be prompted to output an inner voice "train of thoughts" first followed by a final answer and the UI hides the transient inner voice output to the user. I am sure that some 'chat' platforms already implement this.

[Olivier Grisel @ogrisel]

@karpathy s/state/step/

[Olivier Grisel @ogrisel]

@karpathy > No production-grade *actual* RL on an LLM has so far been convincingly achieved and demonstrated in an open domain, at scale.

The solver network of AlphaProof seems like a good state in that direction:

https://t.co/eiSGRwW49x

although it might be a stretch to call it an LLM.

[Olivier Grisel @ogrisel]

@tomgoldsteincs Do you think this idea could be adapted to generative image / video models (stable diffusion and the like)?

[PyData Paris @PyDataParis]

The schedule of PyData Paris 2024 is LIVE! 🚀

Have you secured your tickets yet? 🎟️

Join us for two inspiring days of open-source data science, insightful talks, sprints, and countless networking opportunities, on September 25-26! 🌐

https://t.co/bOZDXHFmfJ