RI: Small: Embracing Deep Neural Networks into Probabilistic Answer Set Programming

低级感知与高级推理的集成是人工智能的基本问题之一。今天，随着深度神经网络的兴起，这个话题又被重新讨论。虽然深度学习在许多感知任务中表现出色，但神经网络如何计算常识推理的多个方面，如因果关系，默认值，溯因推理和反事实推理，并不明显。这些主题在知识表示（KR）领域已经得到了很好的研究，包括答案集编程（ASP），但大多数KR形式主义都是面向逻辑的，并且不像深度学习那样包含高维特征空间和预先训练的视觉和文本模型，这限制了KR在许多涉及不确定性的实际应用中的适用性。拟议的研究的目标是调查的知识表示，推理和学习的原则性组合，通过集成答案集编程与神经网络，这将使表示，推理和学习在符号和子符号的水平。该项目将研究两种不同的整合方法。一种是基于神经原子概念的松耦合，神经原子作为神经网络输出和概率答案集编程参数之间的接口。另一种是更紧密耦合的方法，从神经网络中获得模糊值原子事实，并将模糊答案集语义应用于向量化表示。这些方法不仅允许在神经网络感知结果上应用符号推理，而且还允许在训练神经网络时使用逻辑规则，使得神经网络不仅从来自数据的隐式相关性中学习，而且还从由ASP规则表达的显式复杂语义约束中学习。该项目的成功将有助于确定弥合知识表示和机器学习之间差距的基本问题。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The integration of low-level perception with high-level reasoning is one of the fundamental problems in artificial intelligence. Today, the topic is revisited with the recent rise of deep neural networks. While deep learning excels in many perception tasks, it is not obvious how multiple aspects of commonsense reasoning, such as causality, defaults, abductive reasoning, and counterfactual reasoning, can be computed by neural networks. These subjects have been well-studied in the area of knowledge representation (KR) including answer set programming (ASP) but most KR formalisms are logic-oriented and do not incorporate high-dimensional feature space and pre-trained models for vision and text as in deep learning, which limits the applicability of KR in many practical applications involving uncertainty. The goal of the proposed research is to investigate a principled combination of knowledge representation, reasoning, and learning by integrating answer set programming with neural networks, which will enable representation, inference, and learning in both symbolic and sub-symbolic levels. The project will investigate two different approaches to integration. One is a loose coupling that is based on the concept of neural atoms which serves as an interface between the neural network output and the parameters for probabilistic answer set programming. The other is a tighter coupling method that obtains fuzzy-valued atomic facts from the neural network and applies the fuzzy answer set semantics on the vectorized representation. Not only these methods allow for applying symbolic reasoning on the neural network perception result but also allow for making use of logical rules in training a neural network so that a neural network not only learns from implicit correlations from the data but also from the explicit complex semantic constraints expressed by ASP rules. The success of the project will contribute to identifying fundamental issues in bridging the gap between knowledge representation and machine learning.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.