Yann LeCun: Deep Learning, Convolutional Neural Networks, and Self-Supervised Learning

Audio:

Video:

Clip on autonomous vehicles:

1 Like

You can support Lex Fridman’s AI podcast at Patreon:

Lex just posted an AMA video where he answers Patreon supporters’ questions (including one of mine!). You can watch the video if you sign up as a patron.

1 Like

My question for Yann LeCun:

Curious if anyone can think of a better way to do it than I suggested here.

Really fun interview. No big insights, but Yann is great at explaining things (even the things I disagree with him on).

What do you disagree with Yann about?

Yann downplays the potential of NNs in the short to medium term. His perspective is that of a mainstream academic in that he’s very wary of public expectations being too high, and perhaps too familiar with the myriad of unknown things that need to be investigated. This leads to academically minded experts directing the public to not expect any material gains for the foreseeable future. For some reason academics don’t mind being wrong as long as they are wrong in the “right” way, and Yann seems, to me, to be like that. He famously suggested in 2014 that reinforcement learning and unsupervised learning wouldn’t amount to much for a long time and he was terribly wrong about that but isn’t embarrassed by that error. But he’s very careful not to make the opposite error.

In my opinion that approach may be good for academia, but it’s bad overall. I believe society underinvests in short and medium term infrastructure and policy changes. Social institutions are already conservative and don’t tend to change until a crisis forces them to and academics are always happy to supply institutional managers with excuses to put off investment. This results in mal-investment on a massive scale. A lot of big problems are having their solutions delayed by academics telling institutions that they shouldn’t expect meaningful change for another 20 or 30 years.

Of course, academics won’t sign up to this blunder. They will publicly say, for instance, global warming is important and we need to start investing in it today. But then they also turn around and say that cost effective technologies for dealing with it won’t arrive for decades. So the academics feel like they are promoting solutions, but in fact their net effect is just the opposite because they downplay what can be accomplished in the near term and thus misrepresent the cost-benefit of near term investment. They feel morally superior while they continue to contribute to the problem. And because they feel morally superior there’s no impetus for them to change their behavior.