EFRI BRAID: Using Proto-Object Based Saliency Inspired By Cortical Local Circuits to Limit the Hypothesis Space for Deep Learning Models

这个新兴前沿研究和创新（EFRI）项目将通过使用大脑的计算模型来帮助人工智能系统更有效地利用数据和电力，从而缩小自然智能（NI）和人工智能（AI）之间的差距。具体来说，该项目的灵感来自于哺乳动物大脑的能力，即只存储和处理视觉系统传达的信息的适当选择的子集。如果没有这种被称为选择性注意或“显著性”的特征，大脑很快就会被大量的感官信息淹没。该项目将把视觉注意力的神经科学模型转化为用于深度神经网络学习的新算法。这些新算法将大大减少在学习新模式时必须更新的变量数量。这些受大脑启发的算法的好处将通过在定制的计算硬件上实现来放大，这些硬件旨在模仿哺乳动物大脑结构的形式和功能。其结果将使新的人工智能设备具有变革性的新功能和性能，适用于从自动驾驶汽车到医疗诊断的应用。尽管现有的人工智能系统是革命性的，但它们远远落后于自然世界中的生物体，例如一个年轻的动物从它的父母那里学习如何生存，这需要识别捕食者并学习有效的逃避行动。对当前人工智能硬件和软件的推断预测，达到这些性能水平将需要大量的能量和训练数据。像这样的项目将导致下一代人工智能，通过新的、神经启发的学习策略克服这些预期的障碍。该项目将通过培养从K-12到博士后的各种人工智能培训人员来支持未来的人工智能劳动力，并将为人工智能研究和开发社区提供创新的算法，硬件和数据集。深度学习在多个任务中取得了令人印象深刻的性能，这是由反向传播的能力驱动的，可以为大量参数“分配信用”。典型的网络具有非常复杂的计算图，有许多选项可以为每次计算分配积分。这种大量的选项带来了学习非常灵活的好处，但也带来了大量能源消耗的成本以及需要非常大的数据集进行学习。对重要（显著）特征的预先选择会导致学习中的归纳偏差，但当适当地调节时，这种偏差可以被最佳地选择;这发生在通过进化或发展的生物信息处理中。对于这个项目，这些偏见可以受到生物学的启发或学习，并可以在软件和硬件中实例化。该项目的目标是创建一个混合架构，其中局部电路实现了一种注意力机制，该机制提供了一个“门”或调制，用于为具有卷积架构的全局学习网络选择特征。注意力机制通过包括学习到的重要特征的先验知识，大大减少了用于推理和学习的特征的数量。这项研究的起点将是现有的注意力模型，这些模型符合生物学数据，但这将通过允许注意力机制的元分析来扩展。期望的是，在确定和实施一组任务/输入统计的最佳注意力机制之后，推理和学习的功率需求将大大降低，并且学习将基于比传统方法少得多的示例来实现。该项目还将提供大量机会，从多学科受训人员中培训高素质的人工智能工作者，包括代表性不足的少数民族和妇女，从K-12到博士后奖学金。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。

This Emerging Frontiers in Research and Innovation (EFRI) project will close the gap between natural intelligence (NI) and artificial intelligence (AI), by using computational models of the brain to help AI systems make more efficient use of both data and power. Specifically, the project takes inspiration from the ability of mammalian brains to store and process only an appropriately chosen subset of the information conveyed by the visual system. Without this feature, called selective attention or “saliency,” the brain would soon be overwhelmed by the sheer volume of incoming sensory data. This project will translate neuroscience models of visual attention to new algorithms for learning in deep neural networks. These new algorithms will greatly reduce the number of variables that must be updated while learning new patterns. The benefits of these brain-inspired algorithms will be amplified by implementation on customized computing hardware designed to mimic the form and function of structures from the mammalian brain. The result will enable new AI devices with transformative new capabilities and performance for applications from self-driving cars to medical diagnosis. As revolutionary as existing AI systems are, they fall well short of living organisms in the natural world, such as a young animal learning from its parent how to survive, which requires the recognition of predators and learning of effective evasive actions. Extrapolation of current AI hardware and software predicts that reaching these levels of performance would require prohibitive amounts of energy and training data. Projects such as this one will lead to the next generation of AI, overcoming these anticipated obstacles through new, neuro-inspired, learning strategies. This project will support the AI workforce of the future by educating a diverse cadre of AI trainees, from K-12 to Postdocs, and it will make innovative algorithms, hardware and datasets available to the AI research and development community.Deep learning has achieved impressive performance in multiple tasks, driven by the capacity for backpropagation to “assign credit” to a vast array of parameters. Typical networks have immensely complex computational graphs, with many options to assign credit for every computation. This large number of options comes with the benefits of being very flexible in learning, but also with the costs of large energy consumption and the need for very large datasets for learning. A preselection of important (salient) features will cause inductive biases in learning, but such biases, when appropriately conditioned, can be optimally selected; this occurs in biological information processing via evolution or development. For this project, these biases can be inspired by biology or learned and can be instantiated in software and hardware. This goal of this project is creation of a hybrid architecture, where local circuits implement an attentional mechanism that provides a “gate” or modulation for selecting features for a global learning network with a convolutional architecture. The attentional mechanism dramatically decreases the number of features considered for inference and for learning by including a learned prior of what features are important. The starting point for the research will be existing attentional models that fit biological data, but this will be expanded by allowing a metasearch over the attentional mechanisms. The expectation is that after determining and implementing optimal attentional mechanisms for a set of tasks/input statistics, power requirement for both inference and learning will be substantially reduced, and learning will be enabled based on considerably fewer examples than traditional methods. This project will also provide substantial opportunities to advance training of highly qualified artificial intelligence workers, from a pool of multi-disciplinary trainees, including under-represented minorities and women, at all levels from K-12 to Postdoctoral Fellowships. Furthermore, the results will be made available in the form of databases and published system designs.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.