Real-time learning of predictive recognition categories that chunk sequences of items stored in working memory

Abstract

How are sequences of events that are temporarily stored in a cognitive working memory unitized, or chunked, through learning? Such sequential learning is needed by the brain in order to enable language, spatial understanding, and motor skills to develop. In particular, how does the brain learn categories, or list chunks, that become selectively tuned to different temporal sequences of items in lists of variable length as they are stored in working memory, and how does this learning process occur in real time? The present article introduces a neural model that simulates learning of such list chunks. In this model, sequences of items are temporarily stored in an Item-and-Order, or competitive queuing, working memory before learning categorizes them using a categorization network, called a Masking Field, which is a self-similar, multiple-scale, recurrent on-center off-surround network that can weigh the evidence for variable-length sequences of items as they are stored in the working memory through time. A Masking Field hereby activates the learned list chunks that represent the most predictive item groupings at any time, while suppressing less predictive chunks. In a network with a given number of input items, all possible ordered sets of these item sequences, up to a fixed length, can be learned with unsupervised or supervised learning. The self-similar multiple-scale properties of Masking Fields interacting with an Item-and-Order working memory provide a natural explanation of George Miller's Magical Number Seven and Nelson Cowan's Magical Number Four. The article explains why linguistic, spatial, and action event sequences may all be stored by Item-and-Order working memories that obey similar design principles, and thus how the current results may apply across modalities. Item-and-Order properties may readily be extended to Item-Order-Rank working memories in which the same item can be stored in multiple list positions, or ranks, as in the list ABADBD. Comparisons with other models, including TRACE, MERGE, and TISK, are made.

Keywords: Adaptive Resonance Theory; Magical Number 7; Masking Field; Time Invariant String Kernel; capacity limits; category learning; speech perception; working memory.

Figure 1

Macrocircuit of the list chunk learning model simulated in the current article. An Item-and-Order working memory for the short-term sequential storage of item sequences activates a Masking Field network through an adaptive filter whose weights learn to selectively activate Masking Field nodes in response to different stored item sequences and to thereby convert them into list chunks.

Figure 2

In an Item-and-Order working memory, acoustic item activities C^(I)₁, C^(I)₂, C^(I)₃, are stored in working memory by a gradient of activity. A correct temporal order is represented by a primacy gradient, with the most active cell activity X_i corresponding to the first item presented, the second most active corresponding to the second item presented, and so on. (Reprinted with permission from Grossberg and Kazerounian, 2011).

Figure 3

(A) A primacy gradient is stored in response to the sequence of items “1-2-3” is shown, with activities in a solid line corresponding to “1,” activities in dashed lines corresponding to “2,” and activities in dotted lines corresponding to “3.” (B) A primacy gradient is stored in response to the sequence “3-2-1.”

Figure 4

Neurophysiological data and simulations of monkey sequential copying data. (A) Plots of relative strength of representation (a complex measure of cell population activity, as defined by Averbeck et al., 2002) vs. time for four different produced geometric shapes. Each plot shows the relative strength of representation of each segment for each time bin (at 25 ms) of the task. Time 0 indicates the onset of the template. Lengths of segments were normalized to permit averaging across trials. Plots show parallel representation of segments before initiation of copying. Further, rank order of strength of representation before copying corresponds to the serial position of the segment in the series. The rank order evolves during the drawing to maintain the serial position code. At least four phases of the Averbeck et al. (; Figure 9A) curves should be noted: (1) presence of a primacy gradient; that is, greater relative activation corresponds to earlier eventual execution in the sequence during the period prior to the initiation of the movement sequence (period −500–400 ms); (2) contrast enhancement of the primacy gradient to favor the item to be performed (greater proportional representation of the first item) prior to first item performance (period ~100–400 ms); (3) reduction of the chosen item's activity just prior to its performance and preferential relative enhancement of the representation of the next item to be performed such that it becomes the most active item prior to its execution (period ~400 ms to near sequence completion); and (4) possible re-establishment of the gradient just prior to task completion. (Reproduced with permission from Averbeck et al., 2002). (B) Simulations of item activity across the motor plan field of the LIST PARSE model for 3, 4, and 5 item sequences vs. simulation time. In both (A) and (B), line colors correspond to representations of segments as follows: yellow, segment 1; green, segment 2; red, segment 3; cyan, segment 4; magenta, segment 5. (Reproduced with permission from Grossberg and Pearson, 2008).

Figure 5

A Masking Field is shown for three unitized lists which code the sequences “M,” “MY,” and “MYSELF.” Larger Masking Field cells code longer sequences. Larger cells also have stronger inhibitory connections that enable longer unfamiliar lists to overcome the salience of shorter familiar lists. These asymmetric inhibitory coefficients can arise from self-similar activity-dependent growth laws.

Figure 6

An example of Masking Field dynamics when two items are stored in working memory. List chunk activities are shown at various stages of input presentation. In each image, the lower frame shows the inputs to the working memory and the upper frame shows the masking field activities at that time. (A) Only one item is presented to working memory. A distributed activity pattern is generated across list chunks representing 1, 2, 3, and 4 items, with the most active cell a 1-chunk. (B) When two items are presented to working memory, a 2-chunk is most active. (C,D): As time continues, without the addition of any new inputs, one of the 2-chunks is selected through winner-take-all dynamics.

Figure 7

Same as in Figure 5, with three items stored in working memory. (A) One item in working memory and a 1-chunk is most active. (B) Two items in working memory, and a 2-chunk is most active. (C) Three items are stored in working memory, and a 3-chunk is most active. (D) As time goes on, a 3-chunk is chosen and all other chunks are inhibited.

Figure 8

Same as in Figure 6, with four items stored in working memory. (A) One item in working memory and a 1-chunk is most active. (B) Two items in working memory, and a 2-chunk is most active. (C) Three items are stored in working memory, and a 3-chunk is most active. (D) shows the winner-take-all choice of a 4-chunk.

Figure 9

Activities of the Masking Field through time as it responds to (A) sequence “1-2-3” (left) and (B) sequence “3-2-1.” In both cases, a 3-chunk is selected (second row). The random noise in the initial bottom-up filter values enable selection of different Masking Field cells in response to sequences of the same items in different orderings.

Figure 10

Winner-take-all chunk choices (first row) by the Masking Field to the sequences (A) “1-2-3-4” and (B) “4-3-2-1.” Different 4-chunks are chosen to represent the different sequences.

Figure 11

Adaptive filter weights during learning through time of the list chunks for (A) sequence “1-2-3” and (B) sequence “3-2-1.” The white bars represent the actual weights to these cells, while the black bars represent the ground truth weights that are expected after learning. At trial 1, the weights to the list chunks are essentially uniform with only the addition of small amounts of noise. Over time, these weights become parallel to the ground truth weights. For these simulations, there are 205 sequence presentations, before any sequence is presented again. By trial 16,000, where a trial is the presentation of a single sequence, each sequence had been presented a total of 77 times; by trial 32,000, 144 times; by trial 48,000, 221 times; by trial 54,000, 298 times; and by trial 90,000, 442 times.

Figure 12

Learned weights of list chunks that receive inputs from items “1,” “2,” “3,” and “4,” and that represent sequences whose first item is 1. The learned weights have all converged to the ground truth weights.

Figure 13

Activity through time of 4-chunks to the sequence “4-3-2-1” in successive presentation trials (successive rows). Gray bars denote reset events in which an incorrect list chunk is selected. Once the reset event occurs, the most active cell is shut down, and the remaining list chunks are allowed to compete for activity. On successive trials (A–F), fewer resets occur and the correct list chunk is chosen more quickly. See text for details.

Figure 14

Convergence over trials to ground truth weights for sequences (A) “1-2-3-4” and (B) “4-3-2-1.” Learning is much faster than in the unsupervised learning case. By trial 2100 each sequence had been presented 10 times by trial 4200, 20 times; by trial 6300, 30 times; and by trial 8400, 40 times.

Figure 15

(A) ARTWORD processing levels for speech perception. (Reprinted with permission from Grossberg and Myers, 2000). (B) cARTWORD laminar cortical model for conscious speech perception. (Reprinted with permission from Grossberg and Kazerounian, 2011).