Microsoft Research334 тыс
Опубликовано 8 июля 2016, 1:00
ML Day 2014 - "Perceptual Annotation": from Biologically Inspired, to Biologically Informed Machine Learning
Many machine learning applications, explicitly or implicitly, attempt to mimic natural human abilities in a machine. Indeed, any setting where human-provided labels are used as ground truth – whether the system aspires to be biologically-inspired or not – is ultimately driven by the human visual and cognitive system and its ability to provide accurate exemplary labels. However, human-provided ground-truth labels are in many ways just the tip of the iceberg of the information that can be extracted from human judgments. I will describe a new approach – called "perceptual annotation" – in which we use an advanced online psychometric testing platform to acquire new kinds of human annotation data, and we incorporate these data directly into the formulation of a machine learning algorithm. A key intuition for this approach is that while it may be infeasible to dramatically increase the amount of data and high-quality labels available for the training of a given system, measuring the latent exemplar-by-exemplar landscape of difficulty and patterns of human errors can provide important information for regularizing the solution of the system at hand. Finally, I will conclude by exploring how this approach can be extended to incorporate an even greater diversity of different kinds of biological data
Many machine learning applications, explicitly or implicitly, attempt to mimic natural human abilities in a machine. Indeed, any setting where human-provided labels are used as ground truth – whether the system aspires to be biologically-inspired or not – is ultimately driven by the human visual and cognitive system and its ability to provide accurate exemplary labels. However, human-provided ground-truth labels are in many ways just the tip of the iceberg of the information that can be extracted from human judgments. I will describe a new approach – called "perceptual annotation" – in which we use an advanced online psychometric testing platform to acquire new kinds of human annotation data, and we incorporate these data directly into the formulation of a machine learning algorithm. A key intuition for this approach is that while it may be infeasible to dramatically increase the amount of data and high-quality labels available for the training of a given system, measuring the latent exemplar-by-exemplar landscape of difficulty and patterns of human errors can provide important information for regularizing the solution of the system at hand. Finally, I will conclude by exploring how this approach can be extended to incorporate an even greater diversity of different kinds of biological data
Свежие видео
Случайные видео