A High-Level Model of Neocortical Feedback Based on an Event Window Segmentation Algorithm

The author previously presented an event window segmentation (EWS) algorithm [5] that uses purely statistical methods to learn to recognize recurring patterns in an input stream of events. In the following discussion, the EWS algorithm is first extended to make predictions about future events. Next, this extended algorithm is used to construct a high-level, simplified model of a neocortical hierarchy. An event stream enters at the bottom of the hierarchy, and drives processing activity upward in the hierarchy. Successively higher regions in the hierarchy learn to recognize successively deeper levels of patterns in these events as they propagate from the bottom of the hierarchy. The lower levels in the hierarchy use the predictions from the levels above to strengthen their own predictions. A C++ source code listing of the model implementation and test program is included as an appendix. In a previous paper [5], the author presented a segmentation algorithm that uses purely statistical methods to learn to recognize recurring patterns in an input stream of events. This algorithm will be referred to here as the event window segmentation (EWS) algorithm. An event is the arrival of an integer value (for example, a character code) in the input stream. For the purposes of this paper, the events in the input stream can be considered as ordered in time, but with possibly nonuniform intervals between successive events. A recurring pattern is an ordered set of adjacent events that occurs at a statistical frequency that is significantly higher than that expected for a random ordering of events. The initial presentation of the EWS algorithm in [5] made no mention of the potential to use the statistical information acquired by the algorithm to make predictions about future events. The following discussion will focus on methods for extending the EWS algorithm to make such predictions. This work was inspired by the description of the memory-prediction framework by Hawkins and Blakeslee [4], who theorize that the basis of human intelligence is the ability of each small region of the neocortex to learn to recognize recurring patterns. They propose that the primary function of the neocortex is to use the patterns stored in these regions to continuously make reliable predictions about future events. Furthermore, by connecting several such regions to form a hierarchy, successively higher regions in the hierarchy learn to recognize successively deeper levels of patterns in the event streams that enter the regions located at the bottom of the hierarchy. Each region in the hierarchy then uses the predictions from the regions above and below it to strengthen its own predictions. The EWS algorithm is a convenient tool for experimenting with various types of feedback that might be exchanged between the levels in a hierarchy of regions, and for tuning this feedback to improve the predictive power at each level. This paper describes the operation of a high-level model in which each region in the hierarchy is implemented as an instance of the EWS algorithm. To support this model, the  Send correspondence to: Jerry Van Aken, Microsoft Corporation, One Microsoft Way, Redmond, WA 98052. original algorithm is first adapted to predict future events in an input stream of events. These predictions are based on statistical information that is already gathered by the original algorithm. Next, the algorithm is extended to communicate with other regions in a hierarchy. The EWS algorithm in each region sends its local predictions to other regions, and uses feedback from these regions to strengthen its own predictions. The hierarchy discussed by Hawkins and Blakeslee [4] is a tree-like hierarchy, as shown in Figure 1(a). For the sake of simplicity, the following discussion will describe a model hierarchy, as shown in Figure 1(b), in which each level in the hierarchy consists of a single region. Further extension of the EWS algorithm to support more complex hierarchies, such as the one in Figure 1(a), is a potential area for future work. Figure 1. Hawkins and Blakely [4] describe a tree-like hierarchy of regions, as shown in (a). However, the discussion here will consider only a simple hierarchy of regions, as shown in (b), and will focus on the interface between two adjacent regions in the hierarchy. All processing performed by the model hierarchy in Figure 1(b) is initiated by the arrival of an event in the input stream to level 1, at the bottom level of the hierarchy. This event always triggers processing activity in level 1, and this activity may or may not trigger activity in level 2, the next-higher level. If activity is triggered in the level 2, this activity may or may not trigger activity in level 3, and so on. If a prediction made at the top level is accurate, this prediction remains invariant for a relatively long period, during which processing activity is mostly confined to the lower levels in the hierarchy. If an unexpected event occurs at a lower level, activity quickly propagates to the top of the hierarchy. The EWS algorithm does not try to simulate the operation of the biological neurons in the neocortex . However, if the basic operation performed by a human neocortical region is indeed to recognize recurring sequences, as proposed in [4], then the neocortex has at least this much in common with the EWS algorithm. Thus, understanding how to use feedback to improve the predictive power of a hierarchy of instances of the EWS algorithm might provide insights into the role of feedback in the