When working through Michael Nielsen's deep learning tutorial[1], this was exactly the question I had. I used to think deep learning involved 1000s of layers. T

When working through Michael Nielsen's deep learning tutorial[1], this was exactly the question I had. I used to think deep learning involved 1000s of layers. Turns out that training a network a few layers deep can already be a real pain. Currently we seem to be training one-off networks that are very good at a single thing (e.g. recognising hand written digits in 28x28 pixels), but we haven't been able to really connect multiple networks together to get to more generalised applications. I didn't see this model coming, but it show DeepMind is taking an interesting direction here, potentially leading to more generally applicable ML capabilities. [1]: http://neuralnetworksanddeeplearning.com http://neuralnetworksanddeeplearning.com