Whether selecting training data, finding an optimal experimental design, exploring in reinforcement learning, or designing recommender systems, selecting a high-quality but diverse set of items is a core challenge for ML.
Any task that requires selecting multiple, non-similar items leverages the concept of negative dependence. Negatively-dependent measures and submodularity are powerful, theoretically-grounded tools that can aid in this selection.
Determinantal point processes are arguably the most popular negatively-dependent measure, with past applications including recommender systems, neural network pruning, ensemble learning, summarization, and kernel reconstruction. However, the spectrum of negatively-dependent measures is much broader.
This workshop will discuss with the ICML audience the rich mathematical tools associated with negative dependence, delving into the key theoretical concepts that underlie negatively-dependent measures and investigating fundamental applications.
SUBMISSIONS
We invite submissions of papers on any topic related to negative dependence in machine learning, including (but not limited to):
– Submodular optimization
– Determinantal point processes
– Volume sampling
– Recommender systems
– Experimental design
– Variance-reduction methods
– Exploitation/exploration trade-offs (RL, Bayesian Optimization, etc.)
– Batched active learning
– Strongly Rayleigh measures
ORGANIZERS
– Mike Gartrell (Criteo AI Lab)
– Jennifer Gillenwater (Google Research NY)
– Alex Kulesza (Google Research NY)
– Zelda Mariet (MIT)