Deep Learning with Little Labelled Data

点击翻译按钮获取中文摘要

Machine Learning (ML) is a field of Artificial Intelligence (AI) that aims at providing computers with learning abilities, by inferring models from observations and experience. Many of the recent advances of AI have stemmed from ML, in particular the subfield of Deep Learning (DL). DL aims at learning hierarchical structures, in order to extract features from unstructured raw data (e.g., images, sound, text), allowing significant improvements over the state-of-the-art for many well-studied tasks (e.g., object recognition, automatic translation). However, these techniques are data hungry, requiring huge annotated datasets to be able to efficiently learn how to accomplish some specific tasks. Although such datasets are available for some specific well-established tasks, collecting and annotating sufficiently large datasets for a novel application can be expensive, cost- and time-wise. We propose a research program investigating various approaches to improve the efficiency of ML and DL for solving problems where only small annotated datasets available. We are looking to build upon other related or satellite datasets having the same kind of data, but whose annotations, if they exist, are not adapted to the task. From it, we would produce a working model for our current task, by learning (or fine-tuning) on the small datasets with the task-related annotations. The program is organized around four objectives: 1) To investigate how to better learn models by exploiting knowledge extracted from different tasks, which can be reused or adapted to the current context; 2) To improve meta-learning methods that can learn new concepts from only few samples; 3) To develop new approaches to extract better general representations through unsupervised learning; 4) To apply the novel methods to real-world applications, in order to both assess the practicality of the proposed methods and achieve meaningful contributions in the application domain themselves. Methods investigated within this research program have the potential of making ML and DL usable in a variety of contexts where there are no available big datasets well adapted to the specific task at hand. These ML-based systems will assist humans to adapt quickly to a given task by exploiting similar historical tasks; they will have the capacity for a rapid customization to each user. We are aiming at learning better representations of some modalities (e.g., image, text, speech), usable for a variety of purposes. We will look for discovering mechanisms for AI to adapt to new learning tasks, being able to ramp up performance with only a few examples. Finally, we will apply our techniques to domains such as black-box optimization, super-resolution microscopy, and visual place recognition.