Hacker News Reader

Comment by ma2rten | Hacker News Reader

I'm more in Yann LeCun's camp who called unsupervised learning the cake, supervised learning the icing and reinforcement learning the cherry on top of the cake.

View on HN

Wow, I had no idea that Yann LeCun was also this based. I've figured that it was relatively rare to hear people advocate for the superiority of unsupervised methods - but I guess it isn't if a titan like him does. It's good to hear because epistemologically I just do not believe that most learning is anything but unsupervised. There are very few good labels for our data relative to how much data we process in an information theory sense.

> It's good to hear because epistemologically I just do not believe that most learning is anything but unsupervised.

I have a feeling that the lines between the supervised and unsupervised categories will get increasingly blurred, with semi-supervised, self-supervised (eg. like self-attention) and adversarial (eg. GANs) approaches mixing together in strange ways.

I used to think that given enough time, more people would also learn that there are no substantial differences between supervised and unsupervised learning. However, I have come to believe that this might not happen any time soon. The supposed difference between supervised and unsupervised can be easily explained (label / no label -- duh?!). It is much harder to explain, why that distinction is ultimately a mirage. Couple this with a million low quality blog posts on SEO steroids by data grand wizards and machine learning architects, I doubt that even a good text book jointly written by Bengio, Hinton, and LeCun would convince the ML hype train otherwise. But there is always hope!

My opinion is that view is very simplistic and unnecessarily offensive to a whole class of researchers. MuZero, developed by David Silver, uses a combination of RL, supervised learning, and unsupervised learning (state representation) coupled with a planning algorithm. It accomplished things far beyond anything unsupervised learning can ever accomplish.

Unsupervised learning is exactly the wrong way to approach chess or other games that MuZero solves. It's also worth noting that traditional alpha-beta pruning + heuristics are basically neck and neck with the very best of neural network based techniques. I'll trust stockfish over a alpha-zero or MuZero for awhile longer if I'm trying to win a computer chess competition ...

Sure, Stockfish just uses millions of years of evolution to build its heuristics and can't be transferred to any other game. The point remains, calling RL a cherry on the cake compared to unsupervised learning when they are completely orthogonal and not mutually exclusive techniques is simplistic and unnecessarily offensive.

So it's bad to be the cake? I assume he means it's the foundation one falls back on when the more specialized categories of methods are not applicable.

You might not like my analogy either. I think of supervised and unsupervised learning as the majority of the genome of ML, while RL is that little Y chromosome sometimes tacked on to address a few high-profile tasks.

Not disputing your main point, but Stockfish now includes a neural network.

Reply on news.ycombinator.com