Implicit Neural Networks: What they are and how to use them.
- Daniel McKenzie (UCLA)
Feedforward neural networks explicitly describe a series of computations to be done given an input data point b. Implicit networks, on the other hand, describe a condition which must be met, e.g. a fixed point equation depending on b. Recent work has shown implicit networks can match the performance of feedforward networks on common machine learning tasks such as image recognition, while training using a fraction of the memory. In this talk we discuss recent advances in training implicit neural networks, as well as novel applications to problems in which the target output can naturally be thought of as a fixed point (e.g. game theory).