Machine Networks of Attention from Human Networks of Attention

最近，机器学习中增加的“注意力”改进了许多算法，这些算法有可能改变各种机器学习方法（transformers，perceptivers）。简单地说，注意力允许算法将更多的权重分配给与某些任务相关的输入，而将更少的权重分配给不相关的输入。我们认为，随着输入数据量的增加，机器学习中的注意力机制将变得越来越重要，即使是高效的算法也必须做出明智的选择，决定哪些输入应该获得优先级或被主动抑制。我们将应用当前人类注意力的理论来改进机器学习算法，以适应智能体的目标。生物注意力已经被研究了100多年，它包括多个重叠的网络，这些网络帮助生物体将神经处理分配给对给定任务很重要的感觉输入。例如，定向网络使用眼球运动和空间注意力转移来检查我们环境的重要区域。执行控制网络根据我们不断发展的目标调整感官优先级。我们将在各种任务中使用高质量的眼动跟踪数据作为人类注意力的代理，并使用这些数据为机器学习提供新的注意力机制。这项研究的主要目标是改进机器学习算法中现有的注意力机制，使它们能够更好地为给定任务优先考虑输入数据。由此产生的提高的效率将减少现有任务所需的资源，并提高这些算法的范围，用于更昂贵的计算任务。通过明确测试我们从人类视觉处理中了解到的机制，我们将更好地了解机器和人类的注意力如何在联合任务中协同工作。

The addition of 'attention' to machine learning has recently improved many algorithms with the potential to transform a wide range of machine learning approaches (transformers, perceivers). Attention, simply, allows an algorithm to allocate more weight to input that is relevant for certain tasks, and less weight to the irrelevant. We propose that attention mechanisms in machine learning will become increasingly important as the volume of input data increases, and even efficient algorithms will have to make informed choices about which input should receive priority or actively inhibited. We will apply current theories of human attention to improve machine learning algorithms that adjust to the goals of the agent. Biological attention has been studied for more than 100 years and comprises multiple overlapping networks that help an organism allocate neural processing to sensory input that is important to a given task. For example, the orienting network uses eye movements and shifts of spatial attention to inspect important areas of our environment. The executive control network adjusts sensory priority for our evolving goals. We will use high quality eye tracking data in various tasks as a proxy for human attention and use these data to inform novel attentional mechanisms for machine learning. The key objectives of the research will be to improve existing attention mechanisms in machine learning algorithms such that they are better able to prioritize input data for a given task. The resulting improved efficiency will reduce resources needed for existing tasks and and improve the scope of these algorithms for more computationally expensive tasks. By explicitly testing mechanisms that we know from human visual processing, we will gain a better understanding of how machine and human attention might work together in joint tasks.