Machine Learning Seminar Series

Location: ED 612

Speaker: Agustin D'Alessandro, University of Regina

Title: Deep Sets

Abstract:

Deep sets are neural network architectures to process unordered sets of inputs by being permutation-invariant, i.e. ensuring that their output is invariant to permutations of the input set elements. They achieve this by decomposing the network into two key components: (1) Feature Extraction: each element in the set is processed individually by a shared function (like a small neural network) to learn features from it,  and (2) Aggregation: the features from all the elements are combined using an operation that doesn’t depend on their order, like summing, averaging, or taking the maximum value.