RI: Small: Modules for neural computation

人工智能不需要复制人类大脑，但可以从人类大脑中获得建设性的灵感，大脑是分层组织的，从大型大脑结构到较小的区域，皮质柱，一直到由几个相互作用的神经元组成的微电路。这些模块可以通过将问题分解为局部可解的子问题来提高我们的大脑效率。这反过来可能会让我们通过找出大脑中的错误来更快地学习，并让我们在新问题上做得更好。模块还可以使大脑更有效率，允许它使用更少的神经元和突触。在那些大脑似乎受益于模块化的领域，现代深度学习系统似乎更弱。将模块构建到深度神经网络中，有望大大提高泛化能力、可解释性、学习中的信用分配、计算成本，并使它们对对抗性刺激更具弹性。该项目的结果将是改进现代人工智能系统的性能和理解。该项目将通过暑期学校，教学和出版相结合的方式，广泛地传播计算结果到神经科学和神经科学结果到计算社区。为了实现这些目标，该项目使一个广泛的跨学科的方法，既生产系统的模块化和剖析其模块化方面。该研究旨在建立网络，在学习的同时，通过逐步开发结构模块来分解训练任务。这是通过最小化评估神经元连接的社区结构的成本函数来完成的。它将设计学习算法，通过在模块级别执行学分分配来鼓励模块化。这是通过在神经元和模块级别进行门控学习来完成的。最后，项目团队将开发用于解释模块化网络的工具。这是通过对人类受试者进行心理物理实验和计算分析来完成的。所有这些组件相互关联，提供了一个统一的图像，说明模块化如何改善当前的机器学习方法，并在神经科学和深度学习之间建立了一座新的桥梁。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估来支持。

Artificial intelligence need not replicate the human brain, but may constructively take inspiration from it. The brain is organized hierarchically, from large brain structures to smaller regions, cortical columns, all the way down to microcircuits made of a few interacting neurons. These modules may make our brains more efficient by dissecting problems into locally solvable subproblems. This in turn may make us learn faster by figuring out where in the brain a mistake was made and allow us to do better on new problems. Modules may also make the brain more efficient, allowing it to use fewer neurons and synapses. In those areas where brains seem to benefit from modularity, modern deep learning systems appear to be weaker. Building modules into deep neural networks promises to greatly improve generalization, interpretability, credit assignment in learning, computational cost and make them more resilient to adversarial stimuli. The result of this project will be improvements to the performance and understanding of modern artificial intelligence systems. The project will contribute in a broad way to the dissemination of computational results to neuroscience and of neuroscience results to the computational community through a combination of summer schools, teaching, and publishing. To meet these goals, this project enables a broadly interdisciplinary approach both to produce systems with modularity and to dissect their modular aspects. The research aims to build networks that, while learning, dissect training tasks by incrementally developing structural modules. This is done by minimizing cost functions that evaluate the community structure of neuron connectivity. It will design learning algorithms that encourage modularity by performing credit assignment at the level of modules. This is done by gating learning at both the neuron and the module level. Finally, the project team will develop tools for interpreting modular networks. This is done by performing psychophysical experiments with human subjects and by computational analysis. All of these components interrelate, provide a unified picture of how modularity can improve current machine learning approaches, and build a new bridge between neuroscience and deep 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.