Reproduce “Contextual Plackett-Luce: An Efficient Neural Model for Probabilistic Sequence Selection under Ambiguity” on OpenResearch

About this paper

Noam Mizrachi, Nadav Har-Tuv, Shai Shalev-Shwartz

Researchers from Mobileye developed Contextual Plackett–Luce (CPL), an efficient neural model for probabilistic sequence and subset selection that resolves ambiguity in structured prediction tasks by combining autoregressive expressivity with parallel computational efficiency. CPL achieved top performance in multi-modal path prediction (min-ADE 2.35) and representative subset selection (CluF1 0.853) with significantly faster inference times compared to fully autoregressive approaches.

Key insights

• Incorporating history-dependent interactions through learned pairwise terms is crucial for resolving ambiguity and enforcing structural consistency in sequential and subset selection.
• Efficiently precomputing all interaction parameters in parallel, while maintaining a sequential decision process with lightweight incremental updates, can bridge the gap between the expressivity of autoregressive models and the efficiency of parallel methods.
• A training objective that samples partial ground-truth subsets and induces a uniform target distribution over remaining valid candidates is effective for learning from ambiguous, single-sample supervision.

Summary by alphaXiv.