直播时间：2022年2月20日（周日）20:00-21:00

讲者介绍

黄佳妮：宾夕法尼亚大学博四学生，导师是Mayur Naik。主要研究方向是机器学习和编程语言的交叉领域：运用PL的方法，以及神经符号方法，使学习的过程更加强健，数据的使用更加高效。目前在NeurIPS和ICML会议上发表论文2篇。

报告题目：Scallop: From Probabilistic Deductive Databases to Scalable Differentiable Reasoning

报告摘要：Deep learning and symbolic reasoning are complementary techniques for an intelligent system. However, principled combinations of these techniques are typically limited in scalability, rendering them ill-suited for real-world applications. We propose Scallop, a system that builds upon probabilistic deductive databases, to bridge this gap. On synthetic tasks involving mathematical and logical reasoning, Scallop scales significantly better without sacrificing accuracy compared to DeepProbLog, a principled neural logic programming approach. Scallop also scales to a real-world Visual Question Answering (VQA) benchmark that requires multi-hop reasoning, achieving 84.22% accuracy and outperforming two VQA-tailored models based on Neural Module Networks and transformers by 12.42% and 21.66% respectively.

论文题目：Scallop: From Probabilistic Deductive Databases to Scalable Differentiable Reasoning

分享亮点：

1. We introduce the notion of top-k proofs which generalizes exact probabilistic reasoning, asymptotically reduces computational cost, and provides relative accuracy guarantees.

2. We design and implement a framework, Scallop, which introduces a tunable parameter k and efficiently implements the computation of top-k proofs using provenance in Datalog, while retaining the benefits of neural and symbolic approaches.

3. We empirically evaluate Scallop on synthetic tasks as well as a real-world task, VQA with multi-hop reasoning, and demonstrate that it significantly outperforms baselines.

论文推荐

1. DeepProbLog（DPL）系列work：

1.1. Problog: A probabilistic prolog and its application in link discovery. [DE RAEDT, L., KIMMIG, A., AND TOIVONEN, H.] (2007)

这一篇paper是这个系列工作的第一篇paper，要清晰地了解这系列作品的core就看这篇。它清晰地阐述了weighted model counting如何能够被用来计算一个logic program的不同查询结果的概率。Problog有个很有意思的网站：https://dtai.cs.kuleuven.be/problog/editor.html 大家可以自己写点program试着玩一下。

1.2 DeepProbLog: Neural Probabilistic Logic Programming. [R. Manhaeve, S. Dumancic, A. Kimmig, T. Demeester and L. De Raedt.] (2018)

这一篇paper是这个系列工作里把Problog这个engine改造成一个可以被用于learning的engine。它提出的主要改变是把weighted model counting增加一个带gradient的semiring，使得整个reasoning过程differentiable。

2. Tensorlog: A differentiable deductive database. [COHEN, W. W.] (2016)

这篇paper的program execution过程是soft的。作者首先把一个program转换成一个factor graph，然后再利用message passing去代替logic program 执行的过程。

3. Neural Theorem Prover (NTP) 系列work：

3.1 End-to-end differentiable proving. (NTP) [ROCKTÄSCHEL, T., AND RIEDEL, S.] (2017)

这篇paper引入了两个neural module，一个是Unification Module，另一个是Or module去替代deterministic execution中的unification和or operation。这篇paper 的方法非常expensive，需要考虑所有possible reasoning paths。

3.2 Differentiable reasoning on large knowledge bases and natural language (Greedy NTP) [MINERVINI, P., BOSNJAK, M., ROCKTÄSCHEL, T., RIEDEL, S., AND GREFENSTETTE, E.] (2019).

这篇paper在NTP的基础上使用了beam search，让reasoning path selection的过程变得更加efficient·。

3.3 Learning reasoning strategies in end-to-end differentiable proving (CTP) [MINERVINI, P., RIEDEL, S., STENETORP, P., GREFENSTETTE, E., AND ROCKTÄSCHEL, T.]

这篇paper在NTP的基础上使用了一个neural module，来选择需要考虑的reasoning paths，使得reasoning过程更加efficient·。

回放视频+推荐论文+讲者PPT 登录后获取查看地址