François CholletVerified account

Just to be clear, “part of” was meant as in “part of” hybrid systems that include both, as I made amply clear here https://medium.com/@GaryMarcus/the-deepest-problem-with-deep-learning-91c5991f5695 … We both agree that deep learning is part of solution; I was focusing on the part where we might disagree.

There was a talk on statistical + symbolic methods at Clojure Conj last week, and I was struck by the speaker's framing of deep learning as solving *perception* problems --- a subset of intelligence.

Isn't that the very definition in an ensemble?

The problem is when people in the field hear "symbolic" AI, they tune out because their mind goes to GOFAI. In truth, its really about well designed software engineering architectures and more innovative programming for AI.

Great read, thanks! How does “extreme generalization” work in cases when there are no obvious reasoning? Like human speech recognition in any conditions, even unseen?

Would it be possible to train Python-like language constructs into an approximating neural network, and compile those together based on human written source code? Pre-trained neural networks would become the differentiable machine code for human written logic. #MachineLearning