Efficient visual learning with reduced supervision

目前，我们目睹了现实世界系统和应用（例如便携式设备和移动车辆）中对视觉对象检测和识别的惊人需求。视觉识别是下一代智能应用的关键和限制，例如自动驾驶，无约束环境中的机器人和增强现实。识别模型目前都基于一种学习范式，在这种范式中，训练数据（即示例）被提供给机器以自动执行复杂任务。鉴于互联网上生成和共享的信息量越来越大，这种基于数据的机器学习方法尤为及时。深度卷积神经网络等现代机器学习算法可以利用大量数据作为训练样本。然而，必须解决两个主要挑战：i）视觉识别/学习中的当前最先进的方法是基于强大的蛮力方法，其不考虑何时以及如何使用可用的计算。这些技术的高计算成本是使用大量数据进行学习以及在计算受限的平台（例如嵌入式和低功耗设备）上部署这些识别算法的限制因素。ii）标准学习算法不仅需要训练样本（输入），还需要相应的注释或标签（输出）。鉴于可用的“大数据”几乎没有注释，专家需要执行额外的注释步骤（监督）以利用它进行学习。因此，在这些条件下，训练数据的监督注释往往成为大数据学习的瓶颈之一，无论是成本还是可行性。我的研究的主要贡献将是开发计算效率高的视觉学习方法，减少监督。这项研究将集中在数据和算法的属性的深入研究，允许智能和有效的学习减少注释。这将包括：i）有效的视觉学习和推理技术，ii）减少监督的学习，iii）视觉数据探索。这项研究高度相关，适用于各种数据形式的视觉识别任务，包括图像（对象分类、定位、分割）、视频（动作识别和情感分析），和文本（文档分析和语言建模）。我希望我未来的研究可以在减少监督的情况下实现有效和强大的视觉学习。从长远来看，我预计在这个“大数据”时代，有效学习和减少监督将是进一步发展机器学习技术广泛应用的关键因素。

We are currently witnessing an incredible demand for the detection and recognition of visual objects in real-world systems and applications, such as portable devices and moving vehicles. Visual recognition is at once the key and the limitation towards the next generation of intelligent applications, such as autonomous driving, robotics in unconstrained environments and augmented reality. Recognition models are all currently based on a learning paradigm in which training data (i.e. examples) are provided to machines to automatically perform a complex task. This data-based approach to machine learning is particularly timely given the increasingly vast amount of information being generated and shared over the internet.Modern machine learning algorithms such as deep convolutional neural networks can leverage massive amounts of data as training examples. However, two main challenges must be addressed:i) Current state-of-the-arts methods in visual recognition/learning are based on powerful brute-force approaches, which do not reason about when and how to use the available computation. The high computational cost of these techniques is then the limiting factor for learning with massive data and for deploying these recognition algorithms on computation-limited platforms, such as embedded and low power devices.ii) Standard learning algorithms require not only training samples (input), but also the corresponding annotations or labels (output). Given that the available "Big Data" are equipped with little to no annotation, experts are required to perform an additional annotation step (supervision) to leverage it for learning. Thus, in these conditions, supervised annotation of training data often becomes one of the bottlenecks for learning on Big Data, in terms of both cost and feasibility.The primary contribution of my research will be to develop computationally efficient methods for visual learning with reduced supervision. This research will focus on an in-depth study of the properties of data and algorithms that allow intelligent and efficient learning with reduced annotations. This will include: i) efficient techniques for visual learning and inference, ii) learning with reduced supervision, iii) visual data exploration.This research is highly relevant and applicable to visual recognition tasks in various data modalities, including images (object classification, localization, segmentation), videos (action recognition and sentiment analysis), and text (document analysis and language modelling).I expect that my future research can lead to effective and robust visual learning with reduced supervision. Long term, I expect that in this era of "Big Data", learning efficiently and with reduced supervision will be a key factor to further develop machine learning techniques in widespread applications.