Computations and Unsupervised Learning in Recurrent Neural Networks

描述（由申请人提供）：一般来说，人类的大脑，特别是大脑皮层，仍然是人类已知的最复杂的计算系统。阐明大脑皮层处理信息能力的机制对于理解正常的皮层处理和由异常皮层功能产生的无数神经系统疾病都是至关重要的。目前的研究开始于这样一个假设，即理解皮层处理的必要步骤将是阐明两个相关问题：时间处理和神经动力学。首先，时间处理（对时间信息的解码）对大多数形式的感觉处理至关重要，也许最明显的是语言。然而，即使是简单的时间计算形式背后的神经机制也不为人所知。其次，神经动力学（神经活动时空模式的演变变化）被认为在许多形式的神经计算中起着关键作用，包括时间处理。然而，考虑到由尖峰神经元组成的循环网络中神经活动固有的复杂性和高度非线性性质，我们还不明白(1)动态是如何出现和控制的（例如，避免癫痫活动），以及(2)如何以依赖经验的方式调整动态以允许学习。一般认为突触可塑性（以及其他基因座的可塑性）最终是皮层网络控制和行为的基础。然而，尽管许多形式的可塑性在前馈网络中实现时提供了强大的学习规则，但在循环网络中实现有效的学习规则已经被证明是非常棘手的。在当前的建议中，我们将研究并行的多个学习规则，协同操作，是否可以导致循环网络中的有效学习。我们将特别关注兴奋性和抑制性突触的同时控制，这被认为是循环网络的关键。该提案包括两个目标。在第一个目标中，我们将使用一套突触学习规则来训练一个大规模的循环网络，以发展对口语音素的稳定和强大的反应-换句话说，以时空尖峰模式编码复杂的刺激。第二个目标将使用一种新的学习规则来教输出单元选择性地对音素类做出反应——也就是说，解码皮层网络的模式。如果在这个方向上取得进展，这里描述的研究不仅会增强我们对正常皮层处理的理解，而且会增强我们对以异常的时间处理和神经动力学为特征的神经疾病中出现的异常皮层状态的理解；这类疾病包括自闭症、脆性X染色体、阅读障碍和癫痫。行为和认知最终是嵌入在复杂网络中的数十万神经元动态相互作用的一种涌现特性。虽然在孤立地理解细胞和突触特性方面取得了重大进展，但阐明数十万神经元的活动如何构成皮层计算的基础仍然是神经科学中难以捉摸的基本目标。这里描述的研究将使用皮质网络的大规模模拟来定义突触学习规则，这些规则允许计算从循环的皮质网络中出现。具体来说，人工神经网络学习辨别语音音素的能力将被检验。理解计算是如何从相互关联的元素的大量网络中产生的，是理解正常皮层功能以及在无数神经疾病中观察到的病理性皮层异常的必要步骤。

DESCRIPTION (provided by applicant): The human brain in general, and the cerebral cortex in particular, remains the most sophisticated computational system known to man. Elucidating the mechanisms underlying the cerebral cortex's ability to process information is critical for understanding both normal cortical processing, and a myriad of neurological disorders produced by abnormal cortical function. The current research begins with the assumption that a necessary step towards understanding cortical processing will be to elucidate two related problems: temporal processing and neural dynamics. First, temporal processing (the decoding of temporal information) is of fundamental importance for most forms of sensory processing, perhaps most notably speech. Yet the neural mechanisms underlying even simple forms of temporal computations are not known. Second, neural dynamics (evolving changes in the spatial-temporal patterns of neural activity) is thought to play a pivotal role in many forms of neural computation, including temporal processing. However, given the inherently complex and highly nonlinear nature of neural activity in recurrent networks composed of spiking neurons, we do not yet understand (1) how dynamics emerges and is controlled (e.g. to avoid epileptic activity), and (2) how dynamics is tuned in an experience- dependent manner to allow learning. It is generally assumed that synaptic plasticity (as well as plasticity at other loci) ultimately underlies the control and behavior of cortical networks. However, while a number of forms of plasticity have provided powerful learning rules when implemented into feed-forward networks, the implementation of effective learning rules in recurrent networks has proven largely intractable. In the current proposal we will examine whether multiple learning rules in parallel, operating synergistically, can lead to effective learning in recurrent networks. We will pay particular attention to the simultaneous control of excitatory and inhibitory synapses, which is hypothesized to be critical in recurrent networks. The proposal consists of two Aims. In the first Aim we will use a set of synaptic learning rules to train a large-scale recurrent network to develop stable and robust responses to spoken phonemes - in other words to encode complex stimuli in a spatial-temporal pattern of spikes. The second Aim will use a novel learning rule to teach output units to respond selectively to classes of phonemes - that is, to decode the patterns of the cortical network. If progress is made in this direction the research described here will enhance not only our understanding of normal cortical processing, but in the understanding of abnormal cortical states that arise in neurological disorders characterized by abnormal temporal processing and neural dynamics; such diseases include autism, Fragile X, dyslexia, and epilepsy. Behavior and cognition are ultimately an emergent property of the dynamic interaction of hundreds of thousands of neurons embedded in complex networks. While significant progress has been made towards understanding cellular and synaptic properties in isolation, elucidating how the activity of hundreds of thousands of neurons underlie cortical computations remains an elusive and fundamental goal in neuroscience. The research described here will use large-scale simulations of cortical networks to define the synaptic learning rules that allow computations to emerge from recurrent cortical networks. Specifically, the ability of artificial neural networks to learn to discriminate spoken phonemes will be examined. Understanding how computations emerge from massive networks of interconnected elements is a necessary step towards understanding normal cortical function, as well as the pathological cortical abnormalities observed in a myriad of neurological disorders.