This work proposes a novel approach to capture the uncertainty of the forecast using conditional Normalizing Flows, which allows us to model complex multimodal distributions and to compare various models using negative log-likelihood (NLL).

This paper introduces personalized deep learning models for subjective tasks in natural language processing (NLP), such as detecting hate speech, emotional text, and humor. These models consider individual opinions and beliefs, significantly outperforming generalized approaches across 14 NLP tasks. The study finds that personalized Human Bias (HuBi) methods greatly enhance reasoning quality, with multi-task solutions proving more effective than single-task ones, while human and word embeddings offer supplementary insights.

The proposed HyperShot - the fusion of kernels and hypernetwork paradigm aims to switch the classification module parameters depending on the task’s embedding, and relies on relations between embeddings of the support examples instead of direct feature values provided by the backbone models.