Beyond Deep Associative learning

数学家和科学家一直在追求人工智能（AI）超过60年。人工智能中的一个重要学科是机器学习，通过接触大量数据来训练系统以学习新情况。机器学习的应用现在已经很普遍了--想想谷歌或亚马逊上的个性化搜索、语音识别系统或Cortana和Siri等数字助理。这些应用程序是关联的-它们寻找变量之间的一致关系。虽然它们令人印象深刻，但它们仍然缺乏人类使用先验知识和直觉的能力，以及检查新情况和想象替代结果的能力。人工智能的进一步发展可能会停滞不前，而不会突破到更像人类的学习。该提案的目标是开始训练机器像人类儿童一样学习的过程，从而大大推进人工智能领域。人类通过反复接触相似但略有不同的事件来学习。对于机器来说，我们相信同样的结果可以通过新一代的深度神经网络来实现--机器可以通过反复接触类似但有些不同的数据集来训练，以学习和预测结果。我和我的团队将努力实现这一目标，并尝试将人工智能超越联想模型，转向类似人类的推理。有三个具体的成分将成为我们研究的重点：首先，我们将努力训练一台具有基础物理学深度学习的机器，看看它是否可以在各种场景中应用这种训练，接近人类的常识;第二，我们将尝试训练一台机器建立因果关系，这将使它能够预测其他结果，人类对反事实的看法--人类天生就有能力想象无限多的“如果......会怎样？”'情景;第三，我们将探索在特定场景之外训练机器进入全新领域的可能性，就像人类在遇到新情况时使用经验和想象力一样。有许多领域的工作与这个项目相关，我们将使用两个应用程序来说明驱动这项研究的动机，并验证我们的算法。这两个应用程序本身具有重要意义，并解决了一个具有挑战性的工业问题（自动驾驶汽车）和一个重要且非常有影响力的生物信息学开放问题（肽测序）。

Mathematicians and scientists have been pursuing Artificial Intelligence (AI) for over sixty years. A significant discipline within AI is machine learning, where a system is trained through exposure to enormous amounts of data to learn new situations. Applications of machine learning are now commonplace - think of personalized searches on Google or Amazon, voice recognition systems, or digital assistants such as Cortana and Siri. These applications are associational - they look for consistent relationships between variables. While they are impressive, they still lack a human's ability to use prior knowledge and intuition and the ability to examine a new situation and imagine alternative outcomes. Further development of AI could plateau without a breakthrough into more human-like learning. The goal of this proposal is to begin the process of training a machine to learn as a human child learns, thereby significantly advancing the field of AI. Humans learn through repeated exposure to similar but slightly different events. For a machine, we believe the same result can be accomplished through a new generation of deep neural networks - where a machine can be trained through repeated exposure to similar but somewhat different data sets to learn and anticipate outcomes. My team and I will work towards this goal and attempt to move AI beyond associational models and towards human-like reasoning. There are three specific ingredients which will be the focus of our research: first, we will work to train a machine with deep learning of basic physics and see if it can apply this training across a wide variety of scenarios, approaching what would be common sense in a human; second, we will attempt to train a machine in cause-and-effect relationships which should allow it to predict alternative outcomes, the way a human considers counterfactuality - the innate human ability to imagine an infinite number of `what if?' scenarios; third, we will explore the possibilities of conditioning the machine to use its training outside of specific scenarios into completely new fields, the way a human can use experience and imagination when encountering new situations. There are many areas of work associated with this project and we will use two applications to illustrate the motivation driving this research and to validate our algorithms. These two applications are of significant importance by themselves and address a challenging industrial problem (autonomous vehicles) and an important and very influential bioinformatic open question (peptide sequencing).