NeurIPS 2019 Workshop on Sets & Partitions
Towards deep amortized clustering
Juho Lee, Yoonho Lee and Yee Whye Teh
We tackle amortized clustering, the problem of learning a neural network that can cluster a new dataset with only a few forward passes. We propose a novel learning framework for amortized clustering and demonstrate that our model is capable of identifying a variable number of clusters, even far outside of the range seen during training. Although our main focus is on clustering using a mixture of Gaussians, we additionally provide extensions that enable amortized clustering of complex data generated from distributions that do not have a simple parametric form.