Markus Peschl

Deep Learning at Qualcomm AI Research (@QCOMResearch) in Amsterdam 🚲

[Markus Peschl @mlpeschl]

@Plinz @ritwikpavan There are plenty apps doing this but there is a fundamental hardware and user limitation. Getting high quality images of tiny, hard to access spots likely requires better lighting conditions, higher resolution cameras etc.

[Pietro Mazzaglia @pietromazzaglia]

Teach VLAs reasoning and instruction following at training time → generate better actions without thinking at test time

🇧🇷☀️🌴Hybrid Training for VLAs is accepted to @iclr_conf

Read the full paper at: https://t.co/jIOhvVG8HP

w/ @CcansuSancaktar @mlpeschl D. Dijkman

[Markus Peschl @mlpeschl]

@nocontextmemes GPT 4 is being a good sport but gets it right.

[arXiv.org @arxiv]

We are launching a new arXivLabs collaboration with @HuggingFace to make demos related to papers in cs, stats, and eess directly accessible from arXiv!

[Markus Peschl @mlpeschl]

@Twitter is coming up with very innovative ways to increase revenue: Putting ads on ads

[Markus Peschl @mlpeschl]

@mlittmancs I'm sure people have made this joke countless times but that torch is lit man.

[Markus Peschl @mlpeschl]

@timnitGebru I don't see what effective altruism has to do with this. To be honest, I am pretty certain that most EAs would argue that these 44bn USD could have been spent much better at other places, be it for a 'longtermist' or 'shorttermist' cause.

[Markus Peschl @mlpeschl]

@hardmaru I love how it manages to draw square pixels so consistently in this inherently higher resolution

[Markus Peschl @mlpeschl]

@ilyasut What if we wake up from the illusion we like to call 'reality' once we fall asleep?

[Markus Peschl @mlpeschl]

How can we learn reward functions from experts in the presence of conflicting inputs?

Proud to present MORAL: Multi-Objective Reinforced Active Learning, accepted @aamas2022! In collaboration with @4ester @faoliehoek @lucianosiebert.

Check it out at https://t.co/hdS6iWtmbe

[Markus Peschl @mlpeschl]

@fchollet Makes sense. But I suppose the manifold hypothesis persists regardless of the priors we use? Then, end to end DL will never truly get us to the 'system 2' type of capabilities. I guess the uncertainty is in whether we can find good priors to get enough ood generalization?

[Markus Peschl @mlpeschl]

@fchollet I saw you were discussing this with Yoshua Bengio at the AGI conference, to which he replied that the missing piece is getting rid of the independence assumption in the latent space by assuming some additional 'modularity' prior. Do you have any comments on this?

[Markus Peschl @mlpeschl]

@thecapeador @kevinsblake Hook it up to matplotlib and take all my github stars

[Markus Peschl @mlpeschl]

@neu_rips Reward is [BIG] enough (?)

[Markus Peschl @mlpeschl]

@TZahavy This 'Pie' topic sounds interesting, maybe I should start following it...

[Markus Peschl @mlpeschl]

@tudelta @ClaudiaWerker 'Vegetarian restaurants impose a certain eating style on us.'

I find this argument to be very shortsighted. Is an Italian restaurant imposing Italian culture on us just because it only serves pizza? Of course not. The same argument can be made about any type of shop/restaurant.

[Markus Peschl @mlpeschl]

@mark_riedl @_KarenHao As with RL, I do like the safe RL framing of the problem, although I think that 'safe RL' as a field currently is a bit too focused on the problem of safe exploration with 'safe' only being defined in terms of simple toy problems. Talking about what 'safety' means might be good

[Markus Peschl @mlpeschl]

@mark_riedl @_KarenHao Thanks for clarifying, I think I agree with you here. It seems to me that AI ethics is just a very overloaded term nowadays, and the fact that AI 'alignment' carries a generic name but points to a rather niche community does not help either.

[Markus Peschl @mlpeschl]

@mark_riedl @_KarenHao For instance, consider aligning narrowly superhuman models https://t.co/90wGrTiEJW. Getting an AI to 'want' to do what humans want is, in my opinion, very central to the AI ethics debate and, regardless of an agent's complexity, will be necessary for AI alignment.

[Markus Peschl @mlpeschl]

@mark_riedl @_KarenHao Even though this might be the underlying narrative for the AI alignment community, I do not agree that AI ethics is as distant from AI alignment as you suggest here. In fact, I would argue that many questions posed by AI alignment are very relevant for the AI ethics discussion...