This is a highlight from a talk by Diogo Almeida, "Deep learning: Modular in theory, inflexible in practice." Visit Safari to view the full session from the 2016 Artificial Intelligence Conference in New York.
Given the recent success of deep learning, it is tempting to believe that it can solve any problem placed in its path. Just build more neural networks, throw more data at them, and connect them in a modular fashion to do anything. In truth, this approach would never work without thoughtful systems engineering. In this talk excerpt, Diogo Almeida discusses challenging and related issues like data inefficiency and explainability, and he proposes how we should think about tackling these problems.
Deep Learning sessions at the O’Reilly Artificial Intelligence Conference in New York City, June 26-29, 2017