Bridging Computer Science with Neuroscience towards a new understanding of reasoning

这一博士研究项目旨在搭建自动推理和神经科学之间的桥梁，为设计更像人类的计算系统提供启示。传统的机器推理通常被设计为遵循一组预定义的逻辑规则，对于数学定理证明等逻辑符号推理表现良好。然而，这种基于符号逻辑规则的系统往往不能处理包含多模式异质信息(包括视觉、听觉信息等)的推理任务。另一方面，人脑在处理这种非正式、异质和近似推理任务时非常有效地进行推理。通过研究人们对任务进行主动推理时的大脑神经活动，我们可以更深入地理解人类推理的本质，并开发模拟这种推理过程的系统。许多研究已经开发了模拟大脑功能的机器学习模型，如人工神经网络(ANN)和分层隐马尔可夫模型(HMM)。卷积神经网络(CNN)是一种特殊类型的神经网络，最近在图像处理和语音识别任务中特别成功。美国有线电视新闻网在研究视觉皮质的神经结构方面受到启发。由于并行和分布式计算技术的进步，CNN最近变得流行起来，主要是GPU并行计算技术。并行计算允许在可行的时间内训练更深层次的神经网络。研究人员用CNN模拟人的视皮层完成物体识别任务。然而，关于模拟推理引擎--前额叶皮质(PFC)的研究很少。该项目旨在研究人类进行推理时PFC中的神经活动，为开发一种新型的用于自动推理的人工神经系统提供参考。推理的类型包括语言推理、视觉(图解)推理和符号推理。不同类型的推理共同激活了PFC和特定的体感皮质。例如，PFC和视觉皮质在为视觉关系建立心理模型的过程中被激活。视觉推理是具有新皮质的哺乳动物最常见的推理类型。可以说，视觉推理在进化论意义上比其他类型的推理更重要。此外，视觉推理在神经科学界也得到了广泛的研究。因此，在这个项目中，我计划在第一阶段对视觉推理进行研究，然后扩展到其他类型的推理。功能磁共振成像(FMRI)使我们能够通过测量血氧水平依赖(BOLD)反应来监控大脑内的神经活动。大脑某些区域的神经活动水平越高，该区域的血液耗氧量就越高。有了功能磁共振成像，我们可以监控承担推理任务的人PFC内的神经激活模式。然后可以分析这些神经激活的模式，以阐明推理任务的神经计算过程。有了神经计算过程的知识，我们可以修改现有的神经网络(如深度信念网络、递归神经网络和卷积神经网络)，以允许更好地映射到PFC内部的神经电路，或者提出更准确地捕获神经过程的新型神经计算模型。我们还可以使用这些见解以启发式的形式指导自动推理系统的推理过程--这将更多地反映人类的推理，以及使自动系统更像人类，从而使其更容易获得。

This PhD research project aims to bridge automated reasoning and neuroscience to shed light on designing more human-like computing system. Traditional machine reasoning, usually designed to follow a set of pre-defined logical rules, performs well for logical symbolic reasoning such as mathematical theorem proving. However, such symbolic logic rule based systems are often not able to deal with reasoning tasks that contain multi-modal heterogeneous information (including visual, audial information, etc.). Human brains, on the other hand, are very efficient at reasoning with such informal, heterogeneous and approximate reasoning tasks. By studying brain neural-activities when people are actively reasoning about a task, we can gain a deeper understanding of the nature of human reasoning, and develop systems that emulate such reasoning processes.Many researches have been developing machine learning models such as Artificial Neural Networks (ANN) and Hierarchical Hidden Markov Models that emulate the functioning of the brain. One particular type of ANN, the Convolutional Neural Network (CNN), has been particularly successful recently in image processing and speech recognition tasks. CNN is inspired in studying the neural structures of the visual cortex. CNN has become popular recently because of improvement in parallel and distributed computing technology, mainly the GPU parallel computing technology. Parallel computing allows much deeper neural networks to be trained in feasible time. Researches have simulating human visual cortex in tasks of object recognition with CNN. However, little research has been done on emulating the reasoning engine, the pre-frontal cortex (PFC). This project aims to study neural activities in PFC while people are undertaking reasoning, and shed light on developing a new type of artificial neural system for automated reasoning.There are many types of reasoning, such as verbal reasoning, visual (diagrammatic) reasoning and symbolic reasoning. Different types of reasoning activate PFC and specific somatosensory cortex together. For example, PFC and visual cortex are activated in the process of creating mental models for visual relations. Visual reasoning is the most common type of reasoning in mammals with neo-cortex. Visual reasoning is arguably more primary in an evolutional sense than other types of reasoning. Moreover visual reasoning has been widely studied in the neuroscience community. Therefore, in this project I plan to conduct research into visual reasoning in the first phase, and then extend to other types of reasoning. Functional Magnetic Resonance Imaging (fMRI) allows us to monitor neural activities inside the brain by measuring Blood Oxygen Level Dependent (BOLD) response. Higher level of neural activities in certain area of the brain corresponds to increased level of blood oxygen consumption in that area. With fMRI we can monitor patterns of neural activations inside PFC of people undertaking reasoning tasks. These patterns of neural activations can then be analyzed to shed light on the neural computational processes of reasoning tasks. With knowledge of the neural computational processes, we can modify existing neural networks (such as Deep Belief Network, Recurrent Neural Network, and Convolutional Neural Network) to allow better mapping on to the neural circuitries inside PFC, or propose new types of neural computational models that more accurately captures the neural processes. We can also use these insights to guide automated reasoning systems' reasoning processes in the form of heuristics - these would reflect human reasoning more, as well as make automated systems more human-like and thus accessible.