Machine learning methods that make sense to humans

机器学习方法被广泛应用于预测和决策，人们比以往任何时候都更容易受到这些算法决策的影响。人工智能的使用对我们的社会产生了巨大的影响，这反过来又引发了人们对人工智能的透明度、偏见和公平性的担忧。近年来，针对人工智能可解释性的研究已成为一个备受关注的话题。我对可解释性的研究计划将从两个基本方向进行。首先是开发新的方法，在设计时考虑到可解释性。尽管最近深度学习方法很受欢迎，并在物体识别和机器翻译等领域取得了成功的应用，但许多其他领域——尤其是金融、保险，在某种程度上还有健康领域——一直不愿从深度学习的进步中受益，主要是因为人工智能的透明度、信任、偏见和安全性等问题。在这些领域，大多数数据都是结构化的（与音频和视频等非结构化数据相反）。基于少量特征做出的决定更有可能被人类解释。我将重新讨论特征选择，并开发允许通过参考训练集中的相似样本来解释决策的方法。这模仿了医生通常通过参考过去的类似病例来解释他们的决定。我将进一步扩展这个想法，学习可以提供这种解释的深度嵌入空间。另外，我提出了一种从语义空间生成解释的方法。该计划的第二个方向是寻求发展与模型无关的方法，以提高现有黑盒模型的可解释性。在这里，可解释性可以被视为一种事后处理，以阐明模型是如何做出决定的，以及模型所学到的东西对人类是否有意义。沿着这条线，我想检查涉及人类友好概念的方法。人们通常通过引用对人类来说直观但与预测问题没有直接关系的概念来解释他们的决定。领域专家通常有他们关心的概念，并希望检查机器学习模型是否关注这些概念。国际数据公司估计，到2022年，全球对人工智能的投资将从2018年的240亿美元增长到777亿美元。人工智能的广泛使用意味着对社会的影响越来越大。该项目的结果将有助于建立对人工智能的信任，并减轻有关部署人工智能的一些道德担忧。

Machine learning methods are widely used in predictions and decision making, and people are affected more than ever by these algorithmic decisions. Our society has become significantly influenced by the use of AI and this in turn has raised a lot of concerns about AI's transparency, bias and fairness. Research directed towards the interpretability of AI has become a topic of intense interest in recent years. My research program on interpretability will proceed in two fundamental directions. The first is to develop novel methods that are designed with interpretability in mind. Despite the recent popularity of deep learning methods and their successful application in areas such as object recognition and machine translation, many other areas -especially in finance, insurance and, to some extent health- have been hesitant to benefit from advances in deep learning, mainly because of issues rooted in the transparency, trust, bias and safety of AI. Those are areas where most data are structured (as opposed to unstructured data such as audio and video). Decisions made based on a smaller number of features are more likely to be interpretable by human. I will revisit feature selection and develop methods that allow explaining the decision by referring to similar samples in the training set. This mimics how physicians often explain their decision by referring to similar cases in the past. I will further extend the idea to learning deep embedding spaces that can provide such explanations. Alternatively, I propose an approach for generating explanations from a semantic space. The second direction of the program seeks to develop model-agnostic methods for improving the interpretability of an existing black-box model. Here, interpretability may be seen as a post-hoc processing to shed light on how the model came to its decision, and if, what the model has learned makes sense to human. Along this line, I would like to examine methods that involve human-friendly concepts. Humans often explain their decisions by referring to concepts that are intuitive to humans but are not directly related to the prediction problem. Domain experts often have concepts that they care about and would like to examine if a machine learning model pays attention to them. International Data Corporation estimates that worldwide investment in AI will have grown from US$24 billion in 2018 to US$77.7 billion by 2022. The widespread use of AI implies an increasingly significant impact on society. The program results will help establish trust in AI and mitigate some of the ethical concerns about deploying AI.