RI: Small: Dynamic Attractor Computing: A Novel Computational Approach Applied Towards Temporal Pattern and Speech Recognition

利用大脑的计算策略一直是计算神经科学和机器学习长期追求的目标。这一目标一直难以实现的一个原因是，大多数神经计算框架没有有效地捕获大脑的一个基本计算功能：无缝编码、表示和处理时间信息的能力。目前的项目试图通过使用递归神经网络固有的神经动力学来生成时间模式和处理时间信息来解决这一缺点。产生弹钢琴或分析复杂的语音时间结构所需的精细运动模式的能力，只是人类大脑产生和处理复杂时间模式的复杂能力的两个例子。值得注意的是，这两个例子还说明了大脑计算能力的另一个特点：“时间扭曲”。我们可以以不同的速度演奏同一首乐曲，或者理解以慢速或快速说出的语音。大脑以灵活和时间不变的方式处理时间信息的能力背后的机制是当前提案的关键焦点。最近的理论和实验研究都支持这样的观点，即大脑没有采样率、时间仓或显式延迟线，而是通过递归神经网络的内部动力学来编码时间和刺激的时间特征。然而，由于两个原因，这些递归神经网络模型的计算潜力一直是有限的：1)虽然众所周知神经电路的递归连接是可塑性的，但事实证明，将可塑性纳入模拟的递归神经网络模型具有挑战性；2)具有最大计算潜力的动态区域正是那些也表现出混沌的动态区域--避免了它们的大部分计算潜力，因为动力学不能在试验中重现。该项目建立在一个新的框架之上，以一种“驯服”递归网络的混沌动态的方式来调整递归连接的权重。该方法创建了局部稳定的轨迹(动态吸引子)，提供了一种新颖且潜在强大的计算方法，可以优雅地编码时间信息，并保留对最近事件的内部记忆。与当前项目特别相关的是，证明这些网络可以产生以不同速度流动的类似神经轨迹家族，从而允许网络在不同速度下产生相同的复杂运动模式。该项目还将确定动态吸引子和时间扭曲的原理是否可以应用于感觉处理领域，将语音识别作为大脑区分复杂时空刺激能力的试验台。

Harnessing the brain's computational strategies has been a long sought objective of computational neuroscience and machine learning. One reason this goal has remained elusive is that most neurocomputational frameworks have not effectively captured a fundamental computational feature of the brain: the ability to seamlessly encode, represent, and processes temporal information. The current project seeks to address this shortcoming by using the neural dynamics inherent to recurrent neural networks to generate temporal patterns and process temporal information. The ability to generate the fine motor patterns necessary to play the piano or parse the complex temporal structure of speech, are but two examples of the human brain's sophisticated ability to generate and process complex temporal patterns. Notably, both these examples also illustrate an additional feature of the brain's computational abilities: "temporal warping." We can play the same musical piece at different speeds, or understand speech spoken at slow or fast rates. The mechanisms underlying the brain's ability to process temporal information in a flexible and temporally invariant fashion are a key focus of the current proposal. Recent theoretical and experimental studies have favored the view that the brain does not have sampling rates, time bins or explicit delay lines; but rather encodes time and the temporal features of stimuli through the internal dynamics of recurrent neural networks. The computational potential of these recurrent neural network models, however, has been limited for two reasons: 1) while it is well established that the recurrent connections of neural circuits are plastic, it has proven challenging to incorporate plasticity into simulated recurrent neural network models; 2) the dynamic regimes with the most computational potential are precisely those that also exhibit chaos--voiding much of their computational potential because the dynamics is not reproducible across trials. Building on a novel framework, this project tunes the weights of the recurrent connections in a manner that "tames" the chaotic dynamics of recurrent networks. The approach creates locally stable trajectories (dynamic attractors) which provide a novel and potentially powerful computational approach that can elegantly encode temporal information, and retain internal memories of recent events. Of particular relevance to the current project is to demonstrate that these networks can produce families of similar neural trajectories that flow at different speeds, thus allowing the network to generate the same complex motor pattern at different speeds. This project will also determine if the principles of dynamic attractors and time warping can be applied in the domain of sensory processing, using speech recognition as a test bed for the brain's ability to discriminate complex spatiotemporal stimuli.