Freeman W J
Department of Molecular and Cell Biology, University of California at Berkeley 94720-3200.
Prog Brain Res. 1994;102:319-33. doi: 10.1016/S0079-6123(08)60549-X.
Mathematical models are essential for the study of complex neural systems at all levels of the hierarchy from macromolecules through neurons to brain systems. ANN are readily available, but most of them are inappropriate for modeling brain function in normal behavior, because they stem from studies of neural systems in anesthetized or paralyzed animals which are capable only of reflux output. That class of models lacks the goal-directed, self-organizing properties of neural systems in behaving animals. In contrast to the stability of ANN and their reliance on asymptotic convergence to steady states (point attractors) and periodic oscillations (limit cycle attractors), BNN are intrinsically unstable. They continuously generate 'spontaneous' aperiodic activity that manifests the operation of chaotic dynamics undergoing repeated state transitions. Observations on the activity patterns of sensory cortex reveal that the perceptual outputs of BNN are by construction of spatial patterns and dynamic trajectories and not by computation using symbolic representations. Chaotic dynamics plays essential roles both in the construction of perceptions and in the continuing update of cortical populations, which requires selective synaptic modification during associative learning and habituation. Simultaneous multichannel recording from the olfactory bulb and cortex has given the following experimental results. (1) The cortical activity that relates to the perception of a sensory stimulus is carried macroscopically by a smaller number of single neurons e.g. 'units', 'feature detectors'. (2) The macroscopic activity reflects the meaning and significance of the stimulus for the experimental subject and not the stimulus as it is known to the observer. (3) The activity carries the meaning in spatial patterns, not in time series (the difference between a phonograph or radio and movie or TV). (4) The spatial patterns of activity that accompany previously learned stimuli or responses are changed by the introduction of new stimuli and also by modifications in reinforcement contingencies. There is no invariance in the memory store within the populations. (5) The patterns of activity are created by dynamic neural interactions in sensory cortex, not by registration or filtering of stimuli. There is no evidence for storage, retrieval, cross-correlation or logical tree search. (6) The dynamics is chaotic, not merely noisy, so that each act of perception involves a new construction by the cortex not by mere information processing. From these findings we infer that chaotic dynamics plays a crucial role in the construction of the associational contexts comprising the memory systems of experimental subjects.
数学模型对于研究从大分子到神经元再到脑系统等各级层次的复杂神经系统至关重要。人工神经网络很容易获得，但它们中的大多数不适用于对正常行为中的脑功能进行建模，因为它们源于对麻醉或瘫痪动物的神经系统的研究，这些动物只能进行反馈输出。这类模型缺乏行为动物神经系统的目标导向、自组织特性。与人工神经网络的稳定性及其对渐近收敛到稳态（点吸引子）和周期性振荡（极限环吸引子）的依赖相反，生物神经网络本质上是不稳定的。它们持续产生“自发的”非周期性活动，这体现了经历重复状态转换的混沌动力学的运作。对感觉皮层活动模式的观察表明，生物神经网络的感知输出是通过空间模式和动态轨迹的构建，而不是通过使用符号表示的计算。混沌动力学在感知的构建以及皮层群体的持续更新中都起着至关重要的作用，这需要在联想学习和习惯化过程中进行选择性的突触修饰。从嗅球和皮层进行的同步多通道记录得出了以下实验结果。（1）与感觉刺激感知相关的皮层活动在宏观上由较少数量的单个神经元携带，例如“单元”“特征检测器”。（2）宏观活动反映的是刺激对实验对象的意义和重要性，而不是观察者所知道的刺激本身。（3）活动通过空间模式而非时间序列来承载意义（这是留声机或收音机与电影或电视的区别）。（4）伴随先前学习的刺激或反应的活动空间模式会因新刺激的引入以及强化条件的改变而改变。群体内的记忆存储不存在不变性。（5）活动模式是由感觉皮层中的动态神经相互作用产生的，而不是由刺激的记录或过滤产生的。没有证据表明存在存储、检索、互相关或逻辑树搜索。（6）动力学是混沌的，而不仅仅是有噪声的，因此每次感知行为都涉及皮层的新构建，而不仅仅是信息处理。从这些发现中我们推断，混沌动力学在构成实验对象记忆系统的联想情境的构建中起着关键作用。
