Heard a professor at MIT say something about AI learning from mistakes

I was listening to a podcast from MIT's computer science department and this one professor said AI models that train on their own errors improve 30% faster than those that don't. That really stuck with me because I always thought more data was the only way to make them smarter. It makes sense though, like how we learn from messing up in real life. Has anyone else heard about this approach working for their projects?

2 comments

2 Comments

hart.mark14d ago

Extra data is fine but it's like adding more books to a library without learning how to use them. Learning from mistakes forces the model to actually understand the patterns instead of just memorizing. That 30% boost matches what I've seen with reinforcement learning setups where the AI gets negative feedback on wrong guesses.

eva90814d ago

But @hart.mark, isn't that 30% boost mostly because the model gets more total training time rather than the mistakes themselves mattering that much? Extra data just gives it more examples to work with, and bigger datasets have always improved results regardless of the training method.