Method for evaluating working memory and enhancing information transfer in nervous system

Abstract

The invention discloses a method for judging working memory and enhancing information transfer in a nervous system. The method comprises the steps that temporary storage neurons corresponding to all information units are positioned; in the process that the nervous system processes the random event, a temporary storage neuron corresponding to an information unit associated with the random event is activated through a neural electrode; a plurality of first temporary storage neurons activated by random events and a plurality of second temporary storage neurons activated by electrodes jointly act on a plurality of output neurons to obtain an information processing result. The methods and systems may assess working memory, and are used to enhance information transfer within the nervous system.

Description

Technical Field

[0001] The present application relates to the field of brain neuroscience technology, and in particular to a method for evaluating working memory and enhancing information transfer within a nervous system. Background Art

[0002] Imagine you are playing cards with friends. You are actually implementing a sorting algorithm in your brain: the order of the cards is constantly rearranged according to number, suit, etc. to form a winning hand. Sequential sorting, or the manipulation of multiple items with different ordinal ranks in working memory (WM), is essential for daily activities and can be used in aspects such as planning, navigation, verbal communication or solving arithmetic problems.

[0003] Working memory refers to the short-term maintenance and control of information required for ongoing cognitive tasks. Although the mechanisms of WM maintenance have been the focus of research over the past few decades, the control or manipulation of WM is a more complex issue and has therefore been relatively less studied. Previous studies have suggested that WM "gating" serves as a control mechanism that regulates the entry (and exit) of information into (and out of) WM. However, WM control is not just a general gating mechanism, as memory information can be transformed, sorted, and controlled in a variety of flexible ways, requiring the nervous system to implement more sophisticated operations, such as creating and reordering neural assemblies representing individual items in WM. On the other hand, the prefrontal cortex is believed to be involved in the maintenance and control of WM. In the sequential working memory (SWM) task, the items in each sequence are gated into and stored in a decoupled low-dimensional sequence subspace in the prefrontal cortex. Although sequences can be maintained in SWM in a combinatorial manner, there is little evidence that animals can manipulate sequences on the basis of working memory, and the neural and computational mechanisms of working memory sorting in SWM in the prefrontal cortex remain largely unstudied at the single-neuron and population levels.

[0004] In addition to learning mental operations, humans are able to transform these operations into discrete internal variables, thereby constructing complex symbolic systems ranging from natural language to mathematical or musical notation. For example, when you learn simple arithmetic operations such as addition or subtraction, you assign arbitrary symbols '+' or '-' to these operations, a process known as indicator association. Many animals can also understand similar abstract associations; Nieder et al. found association neurons in the prefrontal cortex of macaques and the caudal cingulum of crows that link task-relevant visual cues to numerical values. However, a complete conceptual framework is still lacking to explain how the primate brain combines learned associations into actionable discrete variables for further manipulation. Summary of the invention

[0005] The purpose of the present application is to provide a method for assessing working memory and enhancing information transfer within a nervous system. The method and system can assess working memory and be used to enhance information transfer within a nervous system.

[0006] A first aspect of the present application provides a method for evaluating working memory, comprising:

[0007] (a) providing a sample sequence to a subject for learning working memory, wherein the sample sequence includes a plurality of sequence items, and the subject needs to sequentially learn the plurality of sequence items for working memory;

[0008] (b) After the first delay period, provide the subjects with a rule reminder;

[0009] (c) after a second delay period, instructing the subject to report a result sample sequence, wherein the result sample sequence is obtained by the subject following the prompt of the rule and processed by the subject's nervous system;

[0010] (d) judging the working memory of the subject based on the result sample sequence.

[0011] In another preferred embodiment, the first delay period and / or the second delay period is 500-700 ms.

[0012] In another preferred example, the result sample sequence includes multiple result sequence items.

[0013] In another preferred embodiment, the neural activity signals of the subject's neurons or group neurons in the first delay period and the second delay period are simultaneously recorded and stored by neural electrodes.

[0014] In another preferred embodiment, a plurality of sequence item working memory subspaces for representing information of the plurality of sequence items are obtained by analyzing neural activity signals of neurons or groups of neurons in the first delay period.

[0015] In another preferred embodiment, the sequence item is an external stimulus.

[0016] In another preferred embodiment, the sequence items include information selected from the following group: text, numbers, images, sounds, smells, spatial location information, or a combination thereof.

[0017] In another preferred embodiment, the representation structure of the sequence item working memory subspace faithfully reflects the structure of external stimuli.

[0018] In another preferred embodiment, the working memory subspace is based on neurons.

[0019] In another preferred embodiment, in step (c), when the subject's nervous system is processing, the information in the multiple sequence item working memory subspaces is transferred to the corresponding multiple sequence item temporary subspaces, and the information in the multiple sequence item working memory subspaces is cleared.

[0020] In another preferred embodiment, the information of the multiple temporary subspaces is then transferred to the multiple sequence item working memory subspaces, thereby achieving exchange.

[0021] In another preferred example, in the sorting process (i.e., the rule prompt is sorting), when the sample sequence includes sequence item a and sequence item b, when the rule prompt is to report the sample sequence in reverse order, the result sample sequence is sequence item b and sequence item a.

[0022] In another preferred example, the first sequence item memory subspace is used to represent the information of the sequence item a, and the second sequence item memory subspace is used to represent the information of the sequence item b. After the rule prompt (reverse sorting) occurs, in the first step, the information of the first sequence item memory subspace is transferred to the first temporary subspace. At this time, the first temporary subspace stores the information representing the sequence item a, and the information of the second sequence item memory subspace is transferred to the second temporary subspace. The second temporary subspace stores the information used to represent the sequence item b; in the second step, the information stored in the first temporary subspace flows into the second sequence item memory subspace, and the information stored in the second temporary space flows into the first sequence item memory subspace, thereby realizing the exchange of reverse order output.

[0023] In another preferred example, the rule prompt includes reporting the sample sequence in a forward order or reporting the sample sequence in a reverse order.

[0024] In another preferred example, the rule prompt is to present sequence items in a predetermined manner.

[0025] In another preferred example, the predetermined method includes forward reporting, reverse reporting, interval reporting (for example, giving the 1st sequence item, the 3rd sequence item, the 5th sequence item; or reporting according to equal or unequal numbers of intervals), sum reporting (for example, giving the sum of the 1st sequence item and the 2nd sequence item, and the sum of the 2nd sequence item and the 3rd sequence item), or a combination thereof.

[0026] In another preferred example, the length of the sample sequence is 2 or 3 or longer, for example, n, where n is a number greater than 2. In another preferred example, when the length of the sample sequence is 2, the sample sequence includes a sequence item a and a sequence item b.

[0027] In another preferred example, when the length of the sample sequence is 3, the sample sequence includes sequence item a, sequence item b and sequence item c.

[0028] In another preferred example, the forward sequential reporting sample sequence is sequence item a, sequence item b and sequence item c.

[0029] In another preferred example, the reverse order reporting sample sequence is sequence item c, sequence item b, and sequence item a.

[0030] In another preferred embodiment, the method is a non-diagnostic and non-therapeutic method.

[0031] A second aspect of the present application provides a system for evaluating working memory, comprising:

[0032] A working memory sequence item input module, wherein the working memory sequence item input module is configured to input, transmit or provide a sequence for learning working memory to the subject, wherein the sequence includes ≥2 sequence items;

[0033] A rule prompt module, wherein the rule prompt module is configured to send a rule prompt to the subject; and

[0034] A working memory test module, wherein the working memory sequence item input module is configured to feedback the learning working memory results of the subject based on the rule prompts; and

[0035] An optional timing control module is configured to: after the working memory sequence items are input, set a predetermined delay period I, a rule prompt period, and a predetermined delay period II, wherein the subject begins to provide feedback on the learning working memory results during the predetermined delay period II.

[0036] In another preferred embodiment, the learning working memory result is a sample sequence of report results.

[0037] In another preferred example, the system further includes: a neural signal acquisition module, wherein the neural signal acquisition module is configured to acquire (record and store) the neural signals of the subject.

[0038] In another preferred embodiment, the neural signal includes a neural activity signal of neurons (preferably a group of neurons).

[0039] In another preferred embodiment, the neural signal acquisition module acquires neural activity signals of one or more neurons (preferably group neurons) of the subject during delay period I and delay period II.

[0040] In another preferred embodiment, the neural signal acquisition module further acquires neural activity signals of one or more neurons (preferably group neurons) of the subject during the rule prompt period.

[0041] In another preferred example, the neural signal acquisition module includes neural electrodes.

[0042] In another preferred example, the system further includes: a working memory analysis module, wherein the working memory analysis module is configured to: analyze the learning working memory results fed back by the subject and the subject's neural signals, so as to obtain an evaluation result of the subject's working memory.

[0043] In another preferred embodiment, the evaluation result includes: the quality of working memory, the mode of working memory, or a combination thereof.

[0044] The third aspect of the present application provides a method for enhancing information transfer within a nervous system, comprising:

[0045] (a) Locate the temporary storage neurons corresponding to each information unit;

[0046] (b) During the process of the neural system processing random events, the temporary storage neurons corresponding to the information units associated with the random events are activated through neural electrodes.

[0047] Preferably, the information unit associated with the random event comes from an artificial intelligence system processing the random event; or the information unit associated with the random event comes from a preset event.

[0048] Preferably, the first plurality of temporary storage neurons activated by the random event and the second plurality of temporary storage neurons activated by the electrodes act together on the plurality of output neurons, thereby obtaining an information processing result.

[0049] In another preferred embodiment, the method is a non-diagnostic and non-therapeutic method.

[0050] A fourth aspect of the present application provides a method for enhancing information transfer within a nervous system, comprising:

[0051] (a) Connecting neural electrodes to the nervous system enables reading and regulating the activation status of neurons in various regions of the nervous system;

[0052] (b) inputting each information unit in the information unit set into the nervous system through the sensory organs of the organism in sequence, thereby locating the activated neurons corresponding to each information unit input through the neural electrodes;

[0053] (c) inputting a first event into the nervous system through a sensory organ of the organism, wherein the first event is a subset of the information unit set; and then reading and locating temporary storage neurons of the nervous system for storing each information unit describing the first event through the neural electrodes;

[0054] (d) repeating step (c), sequentially inputting a plurality of events different from the first event into the nervous system through the sensory organs of the organism, each of the plurality of events being a subset of the set of information units;

[0055] (e) When a random event in the external environment is input into the nervous system through the sensory organs of the organism and activates the first plurality of temporary storage neurons, the unactivated second plurality of temporary storage neurons of the nervous system are activated through the neural electrodes.

[0056] In another preferred embodiment, in the step (e), the second plurality of temporary storage neurons activated by the neural electrodes are temporary storage neurons corresponding to the information units describing the association of the random events.

[0057] In another preferred example, the information unit associated with the random event comes from an artificial intelligence system processing the random event; or the information unit associated with the random event comes from a preset event.

[0058] In another preferred embodiment, the neural electrode is connected to a control system so that information can be transmitted between the neural electrode and the control system.

[0059] In another preferred example, the control system includes an artificial intelligence system or a computer storage control system.

[0060] In another preferred embodiment, the control system includes a memory, and the memory is used to store the correspondence information between the information units and neurons obtained through the neural electrodes in the steps (c) and (d).

[0061] In another preferred embodiment, the nervous system is selected from the brain of a primate.

[0062] In another preferred embodiment, the neural electrodes can cover various areas of the nervous system.

[0063] In another preferred embodiment, the method is a non-diagnostic and non-therapeutic method.

[0064] It should be understood that within the scope of the present invention, the above-mentioned technical features of the present invention and the technical features specifically described below (such as embodiments) can be combined with each other to form a new or preferred technical solution. Due to space limitations, they will not be described one by one here. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. It should be understood that the drawings described below are only some implementation examples of the present invention, and ordinary technicians in this field can also obtain other implementation examples based on these drawings without paying creative work.

[0066] Figure 1 are paradigms, behavioral, and neuronal recordings;

[0067] (A) (left) shows the task structure, where on each trial, a sequence of two or three consecutive spatial locations was presented. (A) (right) The monkeys were instructed to memorize the sequence and sort it according to the rule cue (cucumber picture: forward; apple picture: backward) in the order of presentation (forward) or in the reverse order (backward);

[0068] (B) shows a plot of accuracy as a function of ordinal rank. Five example sessions for two monkeys are shown for each sequence length. Behavioral performance is conditional accuracy for each ordinal—e.g., accuracy for ordinal 3 depends on accuracy for ordinal 1 and 2;

[0069] (C) Plot of the serial response pattern as a function of the ordinal rank, averaged across two monkeys (error bars represent SEM). In the reverse condition, monkeys were more likely to confuse ordinal 1 with ordinal 2 than with ordinal 3 (P < 0.05, Wilcoxon signed-rank test);

[0070] (D) Recording locations. (Left) Segmented model of the skull and microactuator; (Middle) Magnetic resonance imaging (MRI) image of monkey O; (Right) Schematic diagram of electrode locations in monkey O (top) and monkey L (bottom);

[0071] (E) Proportion of position-selective neurons during the delay period. ANOVA was performed separately for the firing rate (FR) of neurons at each delay period (delay period 1: FR calculated during the period from -300 to 0 ms before the onset of the regular cue, delay period 2: FR calculated during the period from -300 to 0 ms before the onset of the cue).

[0072] (F to H) Show that in each panel, raster plots and mean firing rates are color-coded according to the spatial position of a given ordinal level (labeled as stimuli 1 and 2) for different tasks [shown in the upper right panel of (F)]. Vertical gray lines indicate stimulus onset times and vertical gray bars indicate regular presentation windows.

[0073] Example neurons showing item selectivity in delay period 1 (F), delay period 2 (G), and both delay periods (H).

[0074] Figure 2 is to recruit two ordinal WM subspaces for mental categorization;

[0075] (A and B) Low-dimensional representations of stimulus 1 and stimulus 2 are found in the two subspaces shown in (B) during the first delay period of the forward trial (-400 to 0 ms before the rule cue onset) (labeled S1D1F - for stimulus 1, delay period 1, forward - and S2D1F); in (B), locations are color-coded and compared with Figure 2 (D) Sharing in. Points near the origin represent data before the onset of the first stimulus;

[0076] VAF between S1D1F (marked as 1) and S2D1F (marked as 2) in (C);

[0077] (D) Graphical summary of SWM representation in neural state space. S1D1F and S2D1F are oriented in a nearly orthogonal manner in neural state space. The neural representation of a sequence can be decomposed into the sum of its constituent terms in the ordinal subspace, the q1, q2 axes of the subspaces S1D1F and S2D1F;

[0078] (E and F) Schematic representation of the mapping (E) and manipulation (F) hypotheses: Items in delay period 1 are either (H1) mapped to the new subspace (E) or (H2) swapped within the original subspace (F). Note that in the bottom panel matrix, the order of the subspaces is the same for both rows and columns.

[0079] (G) is the cross-conditional decoding generalization score matrix between the eight subspaces of monkey O and monkey L;

[0080] (H) is the coefficient of determination (R) for template regression and model comparison under different assumptions 2 ) and relative Bayesian information criterion (BIC) (H3, shared ordinals in delay period 1 and different ordinals in delay period 2; H4, different ordinals in delay period 1 and delay period 2), with error bars indicating SD between regressions of 100 bootstrap estimates of the cross-decoding generalization score matrix;

[0081] (I) Accuracy of decoding the spatial location of each stimulus on correct trials across time using pseudo-trials, for the ordinal 1 and ordinal 2 subspaces. For (I) and (J), lines and shading represent mean ± 95% confidence intervals. Vertical gray shaded areas represent the cue-present period. Horizontal bars represent decoding above chance (P < 0.025, one-sided t-test; thin, uncorrected; thick, Bonferroni-corrected). Dashed gray lines represent chance levels for label shuffling.

[0082] (J) Decoding accuracy across time for single error trials (FB, forward response expected but backward response reported; BF, backward response expected but forward response reported).

[0083] Figure 3 are 2 temporary variables—psychological exchanges in the 2-step model (2temp–2step);

[0084] Wherein, (A and B) show an example neuron exhibiting position-selective transient activity for one stimulus (only after the onset of the backward rule cue) (A), and an example neuron exhibiting position-selective transient activity for two stimuli (only after the onset of the backward rule cue) (B);

[0085] (C) Decoding accuracy across time for stimulus 1 (in temporary variable 2) and stimulus 2 (in temporary variable 1) for monkey O using pseudo-trials;

[0086] For (C) and (H), lines and shading represent mean ± 95% confidence interval. B, corrected backward trial; F, corrected forward trial. Horizontal bars represent the significance level between the two trial types: P < 0.025, one-sided t test; thin, uncorrected; thick, Bonferronic correction;

[0087] (D) Schematic diagram of the 2-step model with 2 temporary variables; the information flow of stimulus 1 (a) and stimulus 2 (b), respectively, from the original ordinal (OR) subspace to the temporary subspace, and then to the target ordinal (TR) subspace;

[0088] (E) (a) and (b) Signal above chance in the relative subspace indicates decoding accuracy above chance. Lines and shading represent mean ± SD. Horizontal bars indicate signal above chance greater than 0, P < 0.025, Bonferronic correction;

[0089] (F) Bootstrap distribution of the time when information appears in the temporary space (P < 0.025, Bonferroni correction), disappears in the original ordinal-WM (OR) (P ≥ 0.025, Bonferroni correction), appears in the target ordinal-WM (TR), and disappears in the temporary subspace in the information flow (a) and (b). ****P < 0.0001 indicates a significant difference in means between the two distributions, two-sided t test;

[0090] (G) shows the signal intensity in the downstream subspace [quantified as radius score (RS)] based on the upstream intensity of the two steps in (a) and (b), with significant differences between the high (H) group and the low (L) group (separated by the median). Arrows show the potential information flow from the upstream subspace to the downstream subspace.

[0091] (H) Decoding accuracy across time for FB single error trials;

[0092] (I) For correct backward trials and FB error trials, significant times of occurrence of information after the rule cue in the ordinal or temporal subspace (P < 0.025, Bonferroni corrected).

[0093] Figure 4 The distributed single-neuron ordinals and the temporal subspace basis are shown;

[0094] Among them, (A) shows how to normalize the response vector (e) of each single neuron i , the unit vector of the i-th neuron) is projected into the corresponding subspace (the axis of subspace r: ). Projection value Defined as the contribution of a single neuron to the subspace.

[0095] (B) Normalized engagement rates (NPR) of ordinal WM and temporal subspaces in monkey O (top) and monkey L (bottom). Error bars represent SEM across recording sessions.

[0096] (C) Distribution of neuronal contributions to ordinal WM and temporal subspace across all recording sessions in monkey O (top, n = 36) and monkey L (bottom, n = 138). Only neurons with significant contributions are shown.

[0097] (D) Histogram of the distribution of NSA indices (top row) between subspace pairs across all recording sessions. The difference in neuronal contribution between subspace pairs in calculating NSA is as follows: rank minus temp and 1 minus 2. Cyan and lavender mark overlapping (|NSA| ≤ 0.2) and disjoint (0.8 ≤ |NSA| ≤ 1) neurons. The bottom row shows a comparison of the proportion of overlapping and disjoint neurons within and between rank-temporal pairs for the corresponding columns. The results for both monkeys were averaged, and each individual was labeled with a different marker.

[0098] Figure 5 The rule-cue subspace and neural state landscape transitions are shown;

[0099] (A) An example of a neuron that is tuned to a forward cue but not to spatial location. Raster plots and mean firing rates are color-coded according to spatial location given an ordinal rating (labeled as stimuli 1 and 2);

[0100] (B) Neural activity is projected onto the rule-cue subspace (nPC = 3). Thick lines show the average activity projection for the corresponding trial type. Thin lines show the projection for 30 individual sequences (one pseudo-trial was sampled for each sequence). Triangles mark the onset of the rule cue. Points in the red curve mark the time points at which decoding accuracy in the temporal subspace differed significantly between forward and backward trials;

[0101] (C) Distribution of the time when information appeared significantly ( P < 0.025, Bonferroni corrected) in the rule-cue subspace, temporary space 1, and temporary space 2;

[0102] (D) The probability of a given trial backward was decoded by neural activity in the rule-cue subspace (nPCs = 2) using correct backward and error single trials. Colored lines and shading represent mean ± SD. Horizontal lines indicate significance above chance; gray lines indicate chance level.

[0103] (E) shows the correlation between the time when S-2 enters the ordinal 1 subspace (switching process) and the time when the decoding probability (regularity cue signal) becomes significant in (D) in correct backward and error trials. Gaussian noise (s = 20 ms) was added to each point for visualization.

[0104] (F and G) Schematic representation of dynamic portraits indicating the relationship system and exchange process.

[0105] Task-related neural activity can be decomposed into an ordinal temporal subspace (F) and a rule-cue subspace [(G), left].

[0106] The exchange process is triggered only in subregions of the rule-cue subspace, supporting the reconfiguration of information flow between ordinal temporary subspaces [(G), right]. DETAILED DESCRIPTION

[0107] How the brain mentally orders a series of items in a specific order in working memory (WM) remains largely unknown. The inventors used high-throughput electrophysiological recordings to study mental ordering in the frontal cortex of macaques, which memorize and order spatial sequences in forward or reverse order based on visual cues. The inventors unexpectedly found that items at each ordinal level in WM are encoded in a separate level (ordinal)-WM subspace, which is then maintained or reordered between subspaces based on cues, accompanied by two additional temporary subspaces, in two operational steps.

[0108] Furthermore, the cue activity acts as an instructive signal that triggers the sorting process. Thus, a complete conceptual framework is proposed in which neural landscape transitions in frontal lobe neural states underlie the symbolic system underlying sequential WM mental programming.

[0109] In the following description, many technical details are provided to help readers better understand the present application. However, those skilled in the art can understand that the technical solution claimed in the present application can be implemented even without these technical details and various changes and modifications based on the following embodiments.

[0110] the term

[0111] As used herein, the terms "rule hint" and "rule clue" are used interchangeably.

[0112] As used herein, the terms "rank," "ordinal," and "rank" are used interchangeably.

[0113] As used in this article, working memory refers to a short-term memory. For example, the phenomena seen and the words heard will first enter the working memory and then be processed according to the current purpose. For example, the subconscious understanding of language syntax and mental arithmetic of addition, subtraction, multiplication and division all rely on the processing of information stored in working memory.

[0114] It should be noted that in the application documents of this patent, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one" do not exclude the existence of other identical elements in the process, method, article or device including the elements. In the application documents of this patent, if it is mentioned that an action is performed according to an element, it means that the action is performed at least according to the element, which includes two situations: performing the action only according to the element, and performing the action according to the element and other elements. Expressions such as multiple, multiple, and multiple include 2, 2 times, 2 kinds, and more than 2, more than 2 times, and more than 2 kinds.

[0115] In the present invention, all directional indications (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the components in a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0116] To make the purpose, technical solution and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below in conjunction with the accompanying drawings. It should be understood that these are only some examples that the present invention may take for the reader, but are not intended to limit the scope of the present invention.

[0117] Example 1

[0118] Experimental paradigm and behavioral results

[0119] This application trained two macaques to perform a delayed sequence sorting task (see Figure 1 A). In each trial, a spatial sequence of 2 or 3 items in length was presented during the presentation phase while the monkeys fixed their gaze on a dot in the center of the screen. Each ordinal position was randomly drawn (without replacement) from the spatial positions of the six hexagonal rings. The monkeys had to keep the sequence in memory during the first delay period. After the delay, a regular cue appeared in the center of the screen. Depending on the cue, the monkeys had to keep (forward) or reverse (reverse) the positional order of the sequence and maintain the corresponding outcome in working memory (SWM) during the second delay period. After a 'report start' cue (fixation dot turned blue), the monkeys reported the sequence by touching the screen positions in the appropriate order (see Figure 1 A, right; see Methods below for details).

[0120] At each ordinal level, the average accuracy was significantly above chance level ( Figure 1 B, all P values ​​<< 0.001, two-tailed t-tests). Monkeys were able to successfully repeat or reverse length-2 and length-3 sequences on a trial-by-trial basis. In both the forward and reverse conditions, monkeys were more likely to select items that were spatially close to the target when errors occurred and to confuse sequential with adjacent sequences ( Figure 1 C) For sequences of length 3, it is possible that the monkeys may have swapped the first and third items on reverse trials rather than ordering a single item. However, the pattern of errors across ordinals seems incompatible with this view, as very few swap errors were observed between the first and third ordinals ( Figure 1 C) There were no or very small differences in the monkeys' accuracy and reaction times in the forward and reverse tasks.

[0121] Electrophysiological recordings and single-neuron results

[0122] A 157-channel microdrive electrode ( Figure 1D), mainly covering the prefrontal and premotor cortices, with a total of 4191 neurons recorded (monkey O: 3219 neurons; monkey L: 972 neurons). To ensure that there were enough trials for each level and spatial position combination, only neural activity collected during the performance of the sequence task of length 2 was performed for subsequent analysis.

[0123] We focused on neural activity during two delay periods (delay period 1: after the second stimulus presentation and before the start of the regular cue; delay period 2: after the end of the regular cue and before the appearance of the "start" cue). Neuronal responses showed diverse temporal patterns during the delay periods. A significant number of neurons showed position selectivity during the first delay (11%) or the second delay (18%); the distribution of cell selectivity during different delay periods was relatively consistent between the two monkeys ( Figure 1 E). Figure 1 FG shows some example neurons where selectivity during the first delay period tended to be present in both the forward and reverse tasks because the monkeys did not know the sorting rule before seeing the cue ( Figure 1 F). In contrast, neural activity during the second delay period may be selective for one of the tasks ( Figure 1 G, Stimulus 2 was conditioned in the reverse trial). Approximately 32% of neurons showed location-selective neural activity during both delay periods ( Figure 1 E). An example neuron ( Figure 1 H) shows selective neural activity for positions 1 and 6 (on ordinal 1) and positions 3 and 4 (on ordinal 2) during the first delay period. The selectivity for these positions was maintained until the second delay period in the forward task, but was reversed immediately after the rule cue in the reverse task.

[0124] The geometric structure of sequential working memory representation in frontal lobe population neural activity

[0125] Next, we focused on the activity of frontal cortical cells at the population level to identify neuronal ensembles that represent single position information in spatial working memory (SWM). We first determined the geometric characteristics of SWM by analyzing the high-dimensional state of frontal neuronal activity during the first delay period of forward trials. We used a generalized linear model (GLM) to quantify the effects of spatial position and ordinal rank on the neural responses of individual neurons during delay period 1—fitting the firing rates of individual neurons with spatial position and ordinal rank as variables (6 positions × 2 ranks = 12 combinations). It should be noted here that although the task design did not include repeated positions in the sequence (e.g., the sequence ‘11’, ‘22’, etc.), the GLM can account for the bias caused by the lack of these repeated trials (see Methods for more details). The selectivity of each neuron for each variable was measured using the regression coefficient obtained by regressing all correct trials ( Figure 2 A).

[0126] Therefore, the regression coefficients of all neurons can be concatenated together to obtain a vector representation of the population state for the 12 position-level combinations. These 12 vectors were then divided into two groups (ordinal level 1 and level 2), and principal component analysis (PCA) was performed on each group to obtain the axes that capture the main response variance due to position changes (the first two principal components). Consistent with previous studies, this analysis produced a highly reliable state space characterization: first, the subspaces of different ordinal levels (ordinal level 1 and level 2) are nearly orthogonal in the neural state space, which is clearly manifested by their low variance explanation; second, a common ring-like geometric structure can be found in each ordinal subspace, which is very similar to the ring structure of the spatial positions presented to the monkeys ( Figure 2 B and 2C). Thus, a factored representation of SWM was identified by a population of frontal cortical neurons—SWM employs an additive combination of the vector representations of each ordinal rank position ( Figure 2 D).

[0127] Neural dynamics in ordinal subspace reflect sequence rearrangements

[0128] We then examined whether the neural dynamics of activity in the ordinal subspace during the two delay periods reflected reordering of short-term working memory (SWM). Using the subspace analysis described above, the same neuronal population was used to examine neural activity during the two delay periods in both trial types (forward / backward), resulting in eight conditions (2 delay periods × 2 trial types × 2 ranks). The corresponding eight ordinal subspaces were found in each monkey (see Methods); the SWM representations obtained under the different subspace computation conditions were labeled with the following abbreviations when necessary: ​​S1D1F represents memory of ordinal rank 1 (S1) at time period delay 1 (D1) in forward report trials (F), S2D2B represents memory of rank 2 at time period delay 2 in backward report trials (B) ... ( Figure 2 B, EG). We assume two possibilities for SWM reordering after the rule prompt appears:

[0129] First, in delay period 1, sequence items in separate ordinal working memory subspaces (Ordinal 1 and Ordinal 2) are mapped into other ordinal working memory subspaces (Ordinal 1' and Ordinal 2') in the second delay for subsequent motor instructions, thus involving four separate ordinal working memory subspaces and reflecting a "mapping" process from perception to motor preparation ( Figure 2 E top). In this case, since the order-1 subspace in delay period 1 is different from the order-1' subspace in delay period 2, the cross-conditional decoding performance at different delay periods (training on D1 and testing on D2 or vice versa) should still be lower, even though S1D1F, S1D1B and S1D2F, S2D2B retain the same information. The same reasoning applies to the order-2 and order-2' subspaces, so a 4-block diagonal matrix representing the expected cross-conditional decoding results can be obtained ( Figure 2 E bottom).

[0130] Second, the maintenance or exchange of sequential items involves only two ordinal working memory subspaces (e.g., in backward trials, information in the ordinal 1 subspace is transferred to the ordinal 2 subspace, while the ordinal 2 information is moved to the ordinal 1 subspace). This hypothesis reflects the intrinsic exchange process of WM ( Figure 2 F top). In this case, since the subspaces for each ordinal number are the same for delay period 1 and delay period 2, the cross-conditional decoding performance at different delay periods should remain high, resulting in a 2-block diagonal matrix ( Figure 2 F bottom).

[0131] We performed a cross-condition decoding analysis for each pair between the eight ordinal working memory subspaces identified previously. We trained the decoder using the neural activity projected onto one subspace and tested it using the corresponding neural activity from the other subspace projected onto the same subspace (see Methods). We found that the eight subspaces were primarily derived from two ordinal subspaces (Ordinal 1 and Ordinal 2 subspaces), encoding information about stimulus 1 and stimulus 2, respectively, consistent with our second hypothesis ( Figure 2 G). The decoding accuracy between subspace pairs within the same sequence array is high, close to the positive control group (the diagonal line represents the decoding accuracy between two different estimates of the same subspace), while subspace pairs between different sequence arrays show very low decoding performance ( Figure 2 G). Template regression and model comparison were further performed on the cross-condition decoding accuracy matrix; the results showed that the two-subspace model was superior to other models and could explain the most variance ( Figure 2 H). As a control, these analyses were also performed with more principal components (PCs) included (e.g., 5 PCs), and the results were similar to the original 2PC case. It is important to note that although the two-subspace model is more prominent than the others, these models are not mutually exclusive and multiple models can coexist.

[0132] It is crucial to identify the number of ordinal working memory subspaces during the SWM operation because they may involve mechanistically different neural dynamic patterns. Based on the results of the two ordinal subspaces, it is predicted that in the forward task, the item information in the ordinal subspaces during the two delay periods will be maintained and exchanged immediately after the rule prompt. The neural dynamics in these two ordinal subspaces should be able to reflect the information exchange in WM. By observing the neural activities of the two ordinal subspaces during the two delay periods, it was indeed observed that the memory signals in the ordinal subspaces could be continuously decoded and read out during the two delay periods of the forward experiment, as well as the information exchange between the two ordinal subspaces in the reverse experiment ( Figure 2 I). In error trials, when the monkeys confused the two rules, opposite neural patterns were found in the two ordinal subspaces, i.e., the ordinal information of delay period 1 was exchanged in delay period 2 of the forward trial, while the ordinal information of delay period 1 was maintained in delay period 2 of the reverse trial ( Figure 2 J) - Although the two maintained signals almost exchanged in the reverse error trials, the information exchange (i.e., the time when the signal re-entered the two ordinal subspaces) occurred approximately 191 ± 35 ms (mean ± SD) after the regular cue appeared in the error trials, which was significantly earlier than the 259 ± 7 ms in the correct trials (P < 0.0001, one-sided t test, Figure 2I, bottom). Sometimes the monkeys also made localization errors, i.e. they followed the rule correctly but remembered the wrong location. In these trials, the corresponding incorrectly located lobe could be decoded from the ordinal subspace in both tasks.

[0133] Memory sorting via temporary working memory space

[0134] So far, the whole picture of frontal lobe neural activity supporting memory swapping between ordinal working memory subspaces has been identified, but no explanation has been provided for how this swapping occurs through neural computation. In computer programming, the simplest and probably most widely used way to swap two variables (e.g., 'a' and 'b') is to use a third temporary variable (e.g., 'temp') since the two variables cannot be swapped directly due to mutual interference. The swap function can be defined as follows:

[0135] defineswap(a,b)

[0136] Step-1temp:=a

[0137] Step-2a:=b

[0138] Step-3b: =temp

[0139] The structure of the frontal cortex has multiple subpopulations and therefore may also support parallel computation using multiple processors. Two or more operations can be performed simultaneously using multiple temporary variables (e.g., 'temp_1' and 'temp_2'). The above swap algorithm using one temporary variable requires three instructions to be executed sequentially, however, using two temporary variables, the swap can be completed in two steps:

[0140] Step-1

[0141] Processor1:temp_1:=a

[0142] Processor2:temp_2:=b

[0143] Step-2

[0144] Processor1:a:=temp_2

[0145] Processor2:b:=temp_1

[0146] If additional temporary variables are recruited in the frontal cortex, temporary position signals should be observed in temporary subspaces - which are different from the previously shown "same across delay periods" ordinal subspaces, but can exchange information with them - acting as buffer memories (or "caches") during the memory exchange process. To test this hypothesis, the changes in neural activity from the first to the second delay period were examined at both the level of individual neurons and populations of neurons. First, it was found that a subset of neurons showed temporary position-selective activity after the rule cue, but these selective neural firings only appeared in backward trials where an exchange was required. Most of these neurons showed temporary selectivity for only one stimulus (example neurons, Figure 3 A), while some neurons were selective for both stimuli (example neuron, Figure 3 B) Still other neurons showed selectivity for both the delay period and the period after the rule cue.

[0147] Next, we examined neuronal populations to search for temporal subspaces (“temporal variables”) and tested whether their dynamics reflected the sequential execution of the swap. We excluded neural activity in the stimulus space during both delay periods, as well as in the ordinal 1 and 2 working memory subspaces, and used the remaining neural activity to search for additional temporal subspaces in which position information (stimuli 1 and 2) should be temporally encoded (see Methods). Two temporal (temp-1 and temp-2) subspaces were identified in the reversed trials, each encoding the position information of an ordinal ( Figure 3 C) When projected using the activity of the forward trials, very limited positional information is found in these subspaces ( Figure 3 C), and the temporary ordinal subspace was not found in the forward test. Therefore, a computational model is proposed in which the frontal cortex uses two additional temporary subspaces to exchange sequence working memory in a two-step manner (the "2-temp-2-step" model, Figure 3 D): In the first step, the stimulus information will be transferred from order 1 (or 2) to the temp2 (or 1) subspace; in the second step, the information will flow from temp2 (or 1) to the order 2 (or 1) subspace to realize the exchange function.

[0148] This model was tested by three additional analyses. First, the temporal relationship between neural activities in the temporary subspace and the ordinal working memory subspace was examined, and it was found that their temporal dynamics perfectly conformed to the "2-temp-2-step" model. That is, the position information first appeared in its corresponding temporary subspace (step 1, the information appeared in the temporary subspace at 176±8 ms), then the information was displayed in its target ordinal working memory (TR) subspace (step 2, the information appeared in TR at 259±7 ms), and gradually disappeared in its original ordinal working memory (OR) subspace (step 1, the information disappeared at 260±3 ms) ( Figure 3 EF). The information in the temporary subspace is kept until the end of the exchange process (step 2, the information disappears at 442±21 milliseconds).

[0149] Second, if the sequence relationship in function holds, an important prediction can be derived from the model: higher memory-related activity transferred from the ordinal subspace should be accompanied by higher activity in the temporary subspace (step 1). A similar relationship should hold for step 2. The analysis confirmed this prediction: during delay period 1, strong ordinal working memory activation led to a corresponding increase in activity in the temporary subspace. This relationship also held during delay period 2 (step 2). Figure 3 G).

[0150] Finally, macaques made errors when they confused the rule cue and the working memory of the exchange sequence in the forward task. Therefore, if the two temporary subspaces are involved in the exchange function, temporary position information should also be observed in these error (forward) trials. The neural activity of a single error trial in the forward trial was projected onto the temporary subspace identified using the correct backward trial. The results showed that there was significant position-selective activity in the temporary subspace in the error trial ( Figure 3 H), indicating the involvement and generalization of the temporary subspace in error (out-of-distribution) trials. Furthermore, activity into the temporary subspace occurred significantly earlier in error trials compared to correct trials ( Figure 3 I, 107 ± 67 ms / error vs. 176 ± 8 ms / correct, one-sided t-test P < 0.0001). This finding is consistent with the swapping time of the working memory subspace on error trials ( Figure 3 IJ), thus establishing a temporal relationship between the temporary subspace and the memory subspace: the early rise of activity in the temporary subspace leads to the early appearance of exchange information in the corresponding ordinal working memory subspace.

[0151] Single neuron distribution of neural substrates in temporal and ordinal subspaces

[0152] How are the temporary and ordinal working memory subspaces implemented at the level of individual frontal neurons as independent states of frontal neuron population activity? For each neuron, the unit vector on its axis is projected onto the ordinal subspace and the temporary subspace. The projection vector Ari reflects the degree of alignment between the axis of the individual neuron i and the subspace r ( Figure 4 A). We can further ask: what fraction of neurons contribute to each subspace, how these neurons are spatially organized in the prefrontal cortex, and whether individual neurons align with multiple subspaces.

[0153] The normalized participation ratio assesses the proportion of neurons contributing to each subspace (Methods). Values ​​close to 1 indicate that the corresponding subspace is distributed throughout the recorded neuronal population, whereas values ​​close to 0 indicate that it is localized to only a few neurons. On average, about 29% of neurons contribute to the ordinal 1 subspace, 31% to the ordinal 2, 38% to the temporal 1, and 36% to the temporal 2 ( Figure 4 B). Ordinal working memory and temporary memory are widely distributed in the frontal cortex, covering the prefrontal and premotor areas ( Figure 4 C).

[0154] In order to determine whether a single neuronal axis is mainly aligned with one specific subspace or is equally aligned with multiple subspaces, a quantity called the neuron-to-subspace alignment (NSA) index was defined. Based on the NSA index, disjoint and overlapping neurons were classified (see Methods). That is, NSA values ​​close to 1 (or -1) indicate that the neuron is mainly aligned with one subspace (called "disjoint neurons"). In contrast, when NSA is close to 0, this indicates that the neuron is equally aligned with two subspaces (called "overlapping neurons"). Histograms of the NSA index for different subspace pairs are shown ( Figure 4 D, top). There are more disjoint neurons between the WM and temporary subspaces than within both WM (or both temporary) subspaces ( Figure 4 D , bottom), which may provide separate functions for the single neuronal basis between WM and the temporal subspace.

[0155] Relationship between the sorting process and the associated symbols

[0156] When monkeys learned the trained working memory sorting task, they were also able to gradually associate discrete visual cues with different response rules, suggesting that the function of "memory sorting" has been transformed into different symbols. This process is often called "indicative association" and is part of symbolic reference. Although only one pair of picture rule signals (here apple or cucumber) was used in the neural recordings, the monkeys could perform the classification task using multiple pairs of visual signals, which suggests the abstract indicative association involved in SWM sorting.

[0157] Next, we sought to answer the question of how this indicative relationship is represented by the neural state of the frontal lobe. At the level of single neurons, we found that a subset of neurons (monkey O: 7.8%; monkey L: 7.6%) selectively represented visual rule cues rather than spatial locations (a neuron that selected the forward rule was seen in Figure 5 A). We also searched for a neural subspace representing the rule cue by performing principal component analysis on the average neural activity of forward / backward trials, and used the first three principal components to describe the neural dynamics of the rule cue information during the delay period. In the rule cue subspace, the neural trajectories of forward and backward trials diverged immediately after the onset of the rule cue (130 ± 18 ms), significantly earlier than the later exchange process (the emergence of information in the temporary subspace) with an inflection point in backward trials (176 ± 8 ms) ( Figure 5 BC). In addition, the rule-cue subspace is almost orthogonal to the previously determined memory subspaces (ordinal and temporal). The low VAF rate between subspaces indicates that the rule symbol information is represented by different neural subspaces, which can better control the rearrangement of information in the memory-related subspace.

[0158] To investigate whether the visual cue signal triggered the sequence ordering process as expected if the association was indeed occurring, we further analyzed the neural activity in error trials, focusing on those error trials that should have been reported using the forward rule but were reported using the reverse rule (“FB errors”). In these trials, the neural response to the “reverse” cue appeared earlier than in correct trials, even before the rule cue appeared ( Figure 5 D, orange curve). Consistent with the earlier appearance of the rule cue signal, the exchange response in the ordinal subspace appeared much earlier in FB error trials than in correct trials (191 ± 35 ms, Figure 5 E, orange) (259 ± 7 ms, P < 0.0001, one-sided t test, Figure 5 E, red). The early appearance of the rule cue representation may be due to the generation of an internal rule signal. For example, the monkey may have assumed from the beginning of the trial that the current trial was reverse. Similarly, in some error trials, the monkey did not perform the reverse operation sequence as expected (“BF error”), and in these trials, the rule cue signal was significantly delayed (196 ± 25 ms, P < 0.0001, one-sided t test) ( Figure 5 D, green curve), which may be due to insufficient attention to the rule cue. Moreover, in the same BF error trial, the late rise of the rule cue signal led to a delay in the “approach exchange” process ( Figure 5 E, green).

[0159] Finally, with neural evidence for mental sorting processes and indicative visual signals, we propose a conceptual framework for an algorithmic working memory sorting system implemented in the same frontal lobe neural state space: the abstract association between discrete rule cues and the complex dynamics of memory sorting is represented by neural activity and their dynamics in rule cue space, an ordinal working memory subspace, and a temporal subspace. In rule cue space, the influx of external (sensory) or internal signals (e.g., on error trials) pushes neural states along different trajectories ( Figure 5 G, left), in which each point can be further expanded into a vector sum of neural representations of ordinal working memory and temporal signals that have trajectory-dependent dynamics ( Figure 5 F reverse trials). Specifically, after the reverse cue, the trajectory in the regular cue space is pushed to a subregion where the information flow between the ordinal subspace and the temporal subspace is reconfigured to achieve the exchange process ( Figure 5 G, subregion in red box). In contrast, following the positive cue, the state trace was kept outside the switching subregion, preventing the switching process during the entire delay period ( Figure 5 G, gray box). Neural landscape transitions in macaque frontal lobe neural states

[0160] Thousands of neurons were recorded in the frontal cortex of macaque monkeys performing a rule-cued SWM sorting task. The spatial location memory of each ordinal was encoded in SWM in a disjoint ordinal working memory subspace and then maintained or swapped following a rule cue. Working memory sorting was performed in two parallel operational steps using two additional temporary subspaces and indicated by neural activity in independent visual cue subspaces. It was concluded that the neural representation of the abstract association between discrete cues and memory sorting, implemented in the same frontal lobe, could be decomposed into the sum of components in the rule-cued, temporary, and ordinal working memory subspaces. Activity in the subspaces accurately predicted the location or order information that the monkeys recalled, even on single error trials. Thus, neural landscape transitions in the neural state of the frontal lobe of macaque monkeys underlie the symbolic system underlying the operation and control of SWM.

[0161] Break down complex spatial ordering into simple steps

[0162] A fundamental question in cognitive neuroscience is to elucidate the brain events underlying intelligent operations. The computational power of the brain may be derived from the sequential execution of relatively simple computational steps. However, how the brain decomposes complex intelligent operations (abstract rules) into discrete neural units or steps remains an unsolved mystery. Previous studies have shown that neurons represent abstract rules by describing the interrelationships between objects, events, and responses, such as opposing eye movements, spatial locations, categories or boundaries, matches and mismatches, "greater than" and "less than", and rotations of numerical rules. In the Wisconsin Card Sorting Test, which requires flexible rule changes, prefrontal neurons encode the relationship between visual objects and motor responses. However, the manipulation of abstract rules or WM involves not only simple mapping associations between sensory and motor, but also neural mechanisms that create and rotate or combine different components in memory or action. Here, the results show that componentized neural representations and their dynamics in frontal circuits underlie the programmed control of spatial sequences in SWM, and the decomposed information flow is accurately controlled to the appropriate subspaces and operated at a specific moment based on internal (long-term memory) or external (sensory) cues. In sequencing, serial and parallel computations in frontal lobe circuits perform multistep control processes. This control is important because it suggests that neural algorithms may be broadly applicable to a variety of cognitive operations based on sequence sets (e.g., language, music, or mathematics), where different stimuli, tasks, and memories may potentially be generalized.

[0163] From the perspective of saving cognitive resources, using one temporary variable is the simplest way to exchange two variables. So, what is the advantage of using multiple temporary subspaces in the macaque frontal cortex? Here, it is assumed that if only one temporary variable is used, the exchange of variables requires three sequential steps, and the timing of these steps needs to be precisely controlled to avoid interference between variables. Using two temporary variables only requires two steps, and the timing requirement is not as strict. In other words, the precision in time can be traded for the stability of the exchange by increasing the number of temporary subspaces. In addition, multiple temporary variables may be beneficial for generalizing the exchange function to longer (e.g., lengths 3 and 4) sequences, because the behavioral results ( Figure 1 B) suggests that the monkeys might need a more complex sorting algorithm to handle longer sequences (e.g., a sequence of length 3 cannot be reversed with just one swap).

[0164] Switching of neural state landscapes indicates indicative associations

[0165] In number or matching / non-matching rule tasks, two classes of prefrontal cells have been shown: one encoding visual features of shapes or pictures (via input from the inferior temporal cortex) and the other representing numerical or “same / different” information. After instruction cues were given, transient neurons in the dorsal premotor cortex were found to mediate the generation of motor programs from maintained sequence memories. However, the model proposed in this study outlines a more complete conceptual framework to determine the representational geometry and neural dynamics of abstract rules in frontal neural states and further explain the behavior of memory sequencing in SWM. Creating a mapping or indicative relationship between abstract algorithmic representations or operations and their associated labels is crucial because these labels can drive the redescription of representations in humans or animals and help them organize mental space, which may also be reflected back to the external world.

[0166] Clearly, the brain does not work the same way as a computer program. How such cognitive operations are implemented in a complex landscape of specific subregions of neural states and shaped in prefrontal circuits remains incompletely answered. However, based on the neural data, it can be hypothesized that: first, neuronal subpopulations with potential multiple mixed selectivity in prefrontal circuits (e.g., memory and transient neurons) may provide a mechanistic basis for information transformation in different neural hyperspaces; and second, multiple subspaces interact with each other through gain modulation mechanisms enabled by the population structure. Only when the system is pushed into a region where information can be exchanged is the gated information flow from the working memory subspace of a specific order to the temporary subspace opened (see Figure 5 J); Third, the gating or interaction process is predetermined by the anatomical connections between neuronal subpopulations and may be formed during the animal's task learning. This theoretical framework requires further causal intervention for each decomposed component to be verified (only correlations are shown in the current results), and further theoretical and experimental research is needed in the broad field of neural symbolic computation.

[0167] John von Neumann once asked how the analog, parallel, and fallible brain could perform multi-step computations without being drowned out by biological noise. Here, we propose that different classes of neural populations and their dynamics in the macaque frontal cortex play a key role in the ordering that constitutes working memory by decomposing complex cognitive tasks into discrete sequences of operations, each of which likely involves parallel computations that underlie relatively slow and serial "production" systems.

[0168] Methods

[0169] Two male rhesus monkeys, aged 7 to 8 years and weighing 9 to 10 kg, were trained to perform this task. During the task, the monkeys sat in a primate chair facing a 21.5-inch light-emitting diode (LED) display, and their eye positions were tracked using an infrared eye-tracking system (Eyelink, SR Research Ltd.) with a sampling rate of at least 250 Hz.

[0170] During the experiment, the monkeys had to start a trial by pulling a lever and fixating a yellow square in the center of the screen. Once they did so, the square turned white and 100 ms later six circles arranged in a hexagon appeared on the screen. The circles had a diameter of 2 degrees and were angled 11 degrees from the center. After the monkeys had fixed their gaze on the white square for 500 ms, two or three stimuli (white dots selected from the six circles; each position was selected only once per trial) were displayed sequentially during the sampling period. Each stimulus was displayed for 250 ms and separated by a random interval of 300 to 500 ms. After the offset of the second stimulus, two delay periods would occur, each lasting 500 to 700 ms. The first delay period would end at the onset of the rule cue image, which indicated the reporting rule that the monkeys had to follow on the current trial, and the second delay period would begin at the offset of the rule cue. The rule cue itself remained on the screen for 250 ms—for the forward task, the cue was a cucumber; for the backward task, the cue was an apple. At the end of the second delay period, the “go” cue would appear (the center white square turned blue), prompting the animal to report. The monkey then touched the screen to report the sample sequence in the order presented (forward rule) or in reverse order (reverse rule). However, if the monkey released the lever or looked outside the fixation window (2.7 degrees) before the prompt appeared, the trial would terminate immediately. When the monkey touched correctly, the touched location would flash white. If they touched the wrong location, the touched location would flash blue and the trial would end. If the monkey reported the entire sequence correctly, he would receive a water or juice reward. The spatial sequence and rule type were pseudo-randomly selected on each trial. The analysis included 36 sessions (monkey O) and 138 sessions (monkey L).

[0171] Electrophysiological recordings

[0172] A 157-channel microdrive (Gray Matter Research, USA, LS-157) (32) was implanted in the frontal cortex of two monkeys. The microdrive has tungsten electrodes (AlphaOmega) with an electrode spacing of 1.5 mm and an impedance of approximately 1 megohm. Electrophysiological signals were recorded at 40 kHz using a neural recording data acquisition system (OmniPlex, Plexon Inc.). Eye position signals and event markers were also collected in real time and stored together with the electrophysiological signals. After recording, high-frequency signals (300 to ~8000 Hz) were automatically classified offline by Ironclust and manually sorted by an offline classifier (Plexon Inc.). Single and multiple units were merged for data analysis, and units with an average firing rate below 1.2 Hz were excluded. During recording, the depth of the electrode was adjusted as needed to maximize the number of units recorded simultaneously.

[0173] Single neuron selectivity

[0174] For each neuron, mean firing rates were calculated over multiple time intervals and then subjected to analyses of variance (ANOVA) across these time periods. Analysis of position selectivity during the delay period involved examining two delay intervals: 300 to 0 ms before the onset of the rule cue, and 300 to 0 ms before the onset cue. Neurons that exhibited transient item selectivity after the rule cue were characterized by position selectivity within a 300 ms window (from 50 ms after the onset of the rule cue to 100 ms after the end of the rule cue) but were non-selective during the delay period. Neurons identified as rule-selective exhibited selectivity for the rule but not for position within the same 300 ms time window. Neuronal selectivity was determined based on P values ​​(P < 0.001 for O-type monkeys and P < 0.05 for L-type monkeys).

[0175] Generalized Linear Regression

[0176] GLMs were used to determine how each neuron was affected by various task variables over time. For each neuron, trials were divided into forward trials and backward trials according to the rule cue, and GLM analyses were then performed separately for the different trial types.

[0177] Each sequence of length 2 can be represented as a combination of two one-hot vectors. For example, the sequence [4 2] can be represented as (000100 010000) T Therefore, the task variable for the sequence [l1,l2] is defined as two one-hot vectors T of length 6 (six positions or two levels of items). r (l r )(r=1 or 2), whose elements follow

[0178]

[0179] Where l = 1,…,6.

[0180] Neuronal activity was obtained by counting spikes using non-overlapping time windows of 50 ms. The number of spikes of neuron i in time window t for the sequence [l1,l2], and a given rule type (rule = F for forward trials, rule = B for backward trials), is described by the GLM as follows

[0181]

[0182] Where Δt = 50s, which is the width of the time window. FR is the corresponding firing rate, which follows the following formula

[0183]

[0184] in are the regression coefficients, and is the intercept that is only related to the neuron identity and rule type. To avoid overfitting, the Lasso regularization term was added, and the regularization amplitude with the least bias was selected using cross-validation. For each neuron, the fitting process was repeated 50 times using different trial sets obtained by bootstrapping. As a result, 50 estimates of each regression coefficient were obtained to confirm that the GLM analysis provides stable results. The GLM fit can compensate for the bias caused by missing conditions and trial imbalance (there is no trial like l1=l2).

[0185] GLM analyses were performed independently at different time points, including the stimulus 1 (S-1) display period (SD-1, or 0 to 250 ms from the onset of S-1), the interstimulus interval (ISI, or 250 to 550 ms from the onset of S-1), the stimulus 2 (S-2) display period (SD-2, or 0 to 250 ms from the onset of S-2), the first delay period (D-1, or -500 to 0 ms before the onset of the rule cue), the rule cue display period (RD, or 0 to 250 ms from the onset of the rule cue), and the second delay period (D-2, or 250 to 750 ms from the onset of the rule cue).

[0186] After obtaining the coefficients, the estimated firing rate of each neuron was reconstructed. All neurons with the highest estimated firing rates below the threshold TH in all time windows and sequences were discarded, where TH for monkey O was 5 spikes per second (spikes·s -1 ), monkey L’s TH=0.

[0187] Regression ordinal rank subspace ( Figure 2 B and C)

[0188] For a given time point t, the 12 coefficients obtained for a particular neuron are (r=1,2;l=1,…,6) (average of 50 estimates) concatenated into a 12-dimensional vector The first six elements correspond to S-1, and the last six elements correspond to S-2. The response of a neuron to a specific stimulus in a specific period is estimated by selecting the corresponding vector elements and averaging all time points in a given period, denoted as Where S is the stimulus and T is the time period. The six-dimensional vectors of all selected neurons are finally stacked together to form a population-level response matrix M(S, T, rule) (size 6×N n , where N n is the number of selected neurons, for monkey O N n =2898, for monkey L's N n =972) for a given combination of stimulus, time period, and rule type (referred to as S / T / R conditions).

[0189] Perform PCA on the matrix M and transform the first N PC PC S (Length N n ) is retained as the subspace basis W. By projecting the data matrix into the subspace, a low-dimensional representation of the position information can be obtained ( Figure 2 B) Select N PC =2 for result analysis, but different N PC The choice does not qualitatively change the results. Since there are 8 S / T / R conditions [2 stimuli (S) × 2 delay periods (D) × 2 rules (F for forward, B for backward), such as M(S-1, D-1, F)], 8 hierarchical subspaces are obtained accordingly, named S / T / R conditions (such as S1D2B, based on W S1D1F ; Figure 2 B).

[0190] Geometric relations between subspaces

[0191] The geometric relationship between subspaces is quantified by calculating the VAF ratio. Given two matrices M a and M b (Size is 6×N n , where N n is the number of neurons, a and b are S / T / R conditions), and their corresponding subspace basis W a (Size is N n ×k a ) and W b (Size is N n ×k b ), subspace pair (W a ,Wb ) is defined as

[0192]

[0193] in If the subspaces are orthogonal, the VAF is 0; if the subspaces are perfectly aligned, the VAF is 1. The ratio of the VAFs of different estimates of the same subspace was also calculated to test the stability of the subspace identification process and as a control for the VAFs between different subspaces.

[0194] Preprocessing of neural activity for decoding of memory items

[0195] To facilitate subsequent decoding analysis, firing rate data were first obtained by counting neuronal spikes with 20 ms non-overlapping time windows and smoothing them using a half-Gaussian kernel (s = 60 ms, including only earlier time points). For each time point, the smoothed firing rate of a given neuron was z-scored by subtracting the mean response of all correct trials and then dividing by the corresponding standard deviation. For a recording session k, the responses of all trials, all selected time periods (see the “Generalized Linear Regression” section) and all recorded neurons were combined to obtain a population response matrix X (k) [Size is Where K (k) is the number of trials recorded, T is the number of selected time points, is the number of recorded neurons in recording session k]. The responses of different experimental conditions and time points are connected to X (k) The first dimension.

[0196] The label of a trial was defined as a combination of the position of stimulus 1 (positions 1 to 6), the position of stimulus 2 (positions 1 to 6), and the trial type (four types of trials were considered: F for correct forward trials; B for correct backward trials; FB for incorrect trials with forward cues but reported backward sequences; BF for incorrect trials with backward cues but reported forward sequences; other incorrect trials were excluded from further analysis). For each session k, 70% of the correct trials under each label were randomly selected for the training set. The remaining correct trials and all incorrect trials are used for the test set For the analysis of different questions, the corresponding label of each trial is assigned based on the item identity and / or position or trial type, which will be explained in the following sections.

[0197] To generate pseudo-trials for a given label for pseudo-population decoding analysis, one trial under this label was randomly selected in each session, the corresponding population responses were pooled, and these trials across all sessions were concatenated along the neuronal axis to form a pseudo-trial matrix. Since every incorrect trial label was not guaranteed in every session, pseudo-population decoding analysis was only applied to correct trials. For each correct label, the corresponding population responses from A trial of size (N) generates a pseudo trial for each session k. 标签 ·T)×N n The pseudo-group training matrix X 训练 , where N 标签 = 60 is the number of labels [(F or B) × 30 sequences per rule], N n is the total number of neurons in all sessions. Similarly, A pseudo trial is generated for each label, forming a size (N 标签 ·T)×N n The test matrix X 测试 .

[0198] Comparison of models under different assumptions ( Figure 2 E to H)

[0199] Several hypotheses have been proposed for the sorting mechanism, focusing on how many different rank subspaces are involved in the sorting process ( Figure 2 E and F). A pipeline is developed to test these different hypotheses. First, a metric is proposed to describe whether multiple subspaces are "similar" to each other and whether they can be replaced by a common subspace. Since each of the subspaces identified above captures a large variance and maintains the geometry of the neural data for a specific time period and a specific regular type (i.e., a specific S / T / R condition, e.g., S1D1F captures the stimulus-1 position geometry of the neural activity for forward trials in Delay 1), it is inferred that two subspaces can be considered similar if their corresponding neuronal activities are similar when projected into either subspace simultaneously. Taking subspaces S1D1F and S2D2B as examples (with subspace cardinalities W, respectively), S1D1F and W S2D2B , size is Nn×k, where k is the subspace dimension and Nn is the number of neurons). First, collect the pseudo-population activity corresponding to S1D1F: pick X 训练 The responses of all positive trials in , only the time points in delay 1 are retained, forming the submatrix X S1D1F , the size is ( Where T delay1 is the number of time points in delay 1. Then, X is adjusted according to the position of stimulus 1 (S-1, 1 to 6).S1D1F Mark each data point in and get a target vector (length is Y S1D1F A low-dimensional logistic regression classifier is trained to associate this pair (X S1D1F W S1D1F ,Y S1D1F ). In order to obtain cross-conditional generalization accuracy, from X to S1D1F test The pseudo-group activity X corresponding to S2D2B is collected in S2D2B , mark the position of S-2, and use the data set (X S2D2B W S1D1F ,Y S2D2B ) to test the decoder trained above. If the performance is high, it indicates that the two datasets have similar representations in S1D1F. Similarly, it is also possible to test whether the two datasets have similar representations in S2D2B. If the two datasets have similar representations in both subspaces, then a common subspace of S1D1F and S2D2B can be constructed to capture the geometry in the two datasets. In this way, the cross-(S / T / R) conditional generalization accuracy of each pair in the eight hierarchical subspaces is calculated, and the accuracy matrix D is obtained. The diagonal of the accuracy matrix (self-generalization performance) is then normalized to obtain the cross-conditional generalization score matrix GS ( Figure 2 G), that is

[0200]

[0201] Where C = 1 / 6 is the chance level of decoding accuracy. From the first row (column) to the eighth row (column), the elements of GS correspond to the training (testing) subspace in the following order: S1D1F, S1D1B, S1D2F, S2D2B, S2D1F, S2D1B, S2D1F, and S1D2B. According to this specific row and column order, GS will show different characteristics under different hypotheses H1 to H4 ( Figure 3 E and F). To assess which hypothesis best fits the data, template regression analysis was used. For ease of use, the following “diagonal” matrix I is defined a,b (size is 8×8, a≤b), that is

[0202]

[0203] In this way, the template regression from H1 to H4 can be expressed as

[0204] H l :

[0205] H2:

[0206] H3:

[0207] H4:

[0208] Flatten the 8×8 matrices on both sides into vectors of length 64 and minimize the Frobenius norm ||GS-GS'|| F , we can get the regression coefficient c. The regression score is defined as

[0209]

[0210] The relative Bayesian Information Criterion (BIC) is defined as

[0211]

[0212] Where kc is the number of regression coefficients, is the average of all entries in R. 2 The hypothesis with a high value and a low relative BIC value fits the data better.

[0213] WM subspace identification and pseudo-population cross-temporal decoding analysis of correct trials ( Figure 2 I)

[0214] Figure 2 The results in H suggest that it is more likely that only two different rank-WM subspaces are involved in the mental sorting process. To obtain a shared low-dimensional representation of multiple group-level response matrices, PCA analysis was performed on them simultaneously. In this case, the matrices were merged into a new matrix along the stimulus axis, and then PCA was performed on it. The first rank-WM subspace rank 1 was obtained by performing PCA on the set of matrices [M(S-1,D-1,F);M(S-1,D-1,B);B(S-1,D-2,F);M(S-2,D-2,B)] (size 24×Nn), and the second rank-WM subspace rank 2 was obtained by performing PCA on the set [M(S-2,D-1,F);M(S-2,D-1,B);M(S-2,D-2,F);M(S-1,D-2,B)] (size 24×Nn).

[0215] In order to examine the dynamics of the information intensity of two stimuli in the two hierarchical subspaces ( Figure 2 I), the accuracy of cross-temporal decoding was also tested. For rank 1 (based on W1), a network consisting of four different parts {(X i W i ,Y i)|i=S1D1F;S1D1B;S1D2F;S2D2B} to train the decoder. The derivation process of each model is similar to that described above (see the “Comparison of models under different assumptions” section). Then use X test The decoders were tested on W1, with each data point labeled according to its position in S-1 or S-2. For each subspace, 1000 decoders were trained and tested using randomly sampled training and test sets. The decoding accuracy of each decoder under different stimuli and rules was first averaged. The mean and standard deviation SD of all decoders were calculated. For rank 2, the decoding accuracy across time was obtained in a similar way.

[0216] Error single decoding analysis ( Figure 2 J)

[0217] To confirm the previously found important decoding results and the corresponding subspace dynamics are related to behavior, the decoding accuracy of error trials in the identified subspace was calculated. First, the subspace basis W (size Nn×ks, where Nn is the number of neurons in all sessions and ks is the dimension of the subspace) was split according to the recording sessions. Specifically,

[0218]

[0219] Among them, W (k) [Size is ] is the neuron basis in session k, is the number of neurons in session k, N SS is the number of logging sessions.

[0220] Similar to pseudo-group analysis, for session k, we can get The population response corresponds to S1D1F and the target vector, by keeping the correct forward trials (1 trial per randomly sampled label) and At the time point with a delay of 1, the position of S-1 is used as the classification label. For the rank-1 subspace (based on W1), use Train a logistic regression classifier. Then, use To test this classifier-decoder, each data point (for correct trials, 1 trial per label was randomly sampled; for incorrect trials, all trials were used for testing) was labeled according to its position in S-1 or S-2. 100 decoders were trained and tested in each subspace. For a specific correct or incorrect trial type ( Figure 2J, as defined in the section “Preprocessing of neural activity for decoding items in memory”), the decoding accuracy for S-1 or S-2 was averaged across all trials of a particular type and across all iterations of the decoder. The mean and standard deviation SD were calculated for each session (n=36). The decoding accuracy for error trials in other subspaces (e.g., rank 2) was obtained by a similar process.

[0221] To make a fair comparison, all time points ( Figure 2 I and J, and Figure 3 Decoding significance in both C and H) was tested using single-shot decoding accuracy.

[0222] Temporary subspace identification ( Figure 3 C)

[0223] To exclude prominent neural activities related to other events, the sensory subspace of each monkey was obtained from the connectivity matrix [M(S-1, SD-1, F); M(S-1, SD-1, B); M(S-2, SD-2, F); M(S-2, SD-2, B)] (size 24 × Nn).

[0224] To determine the potential temporal subspace of S-1, we first constructed the GLM coefficient matrix M(S-1, SP, B) (size 6Tsp×Nn, where SP is the exchange period, which is 50 to 550 milliseconds after the onset of the rule cue, and Tsp is the number of time points in SP). We subtracted the coefficients of the components aligned in rank 1, rank 2, and sensory subspaces from M to obtain a residual coefficient matrix

[0225]

[0226] Where I is the Nn-dimensional identity matrix, W i The size of is Nn×ki (ki=20), which is the basis of subspace i, Subs={sensory,rank1,rank2}. It should be noted that although only the first two or three dimensions of the sensory and rank subspaces are used for visualization and decoding analysis, the activities of the first 20 dimensions of these subspaces are also included here.

[0227] M′ with different starting points and time window lengths (different subintervals of SP) were calculated, and then PCA was performed on each M′ to obtain various potential temporary subspace bases. The basis that can capture most of the variance while maintaining the geometry of the spatial position of S-1 was selected and named temporary subspace 2. (temp 2). After determining temp 2, the M′ component consistent with temp 2 was subtracted from M′ and the same process described above was followed. By testing the error trial decoding accuracy of the candidate subspaces found after subtracting temp 2, it was confirmed that the temporary subspace of S-1 could no longer be found. The same temporary subspace identification process was used for M(S-2,SP,B) to identify the temporary subspace of S-2 (named temp1).

[0228] The above temporary subspace identification process is also applied to M(S-1, SP, F) and M(S-2, SP, F). This is also confirmed by testing the error decoding accuracy of the candidate temporary subspaces.

[0229] Characteristic time of mental sorting process ( Figure 3 E and F)

[0230] To make the dynamics of the same rank information in different subspaces comparable, the signal-above-chance (SAC) of the decoding accuracy of a specific rank in a specific subspace was calculated by subtracting its chance level and dividing by the maximum above-chance decoding accuracy ( Figure 3 E). The chance level of the hierarchical subspace is obtained by shuffling the labels, while the chance level of the temporary subspace is the accuracy of the forward test.

[0231] In order to determine the characteristic time that reflects the dynamic changes of the psychological sorting process ( Figure 3 F), when the decoding accuracy was significantly (P < 0.01) higher than the control level (this was also the time when the corresponding SAC was significantly greater than 0), the time point was marked as the level information appearing in a specific subspace; when the significance decreased (P ≥ 0.01), the time point was marked as the level information disappearing in that subspace.

[0232] To make a fair comparison, all feature times were obtained by single-shot decoding accuracy. Thirty-six sessions (n=36) were randomly sampled from all sessions and feature times were calculated. The bootstrapping process was repeated 1000 times to obtain the distribution of relevant feature times.

[0233] The signal strength correlation between subspaces ( Figure 3 G)

[0234] We tested whether the signal strength in the rank-WM subspace and the signal strength in the temporary subspace were correlated at the single-trial level. To exclude differences in the recording quality of different sessions, the neural activity of each session was first projected onto its subspace basis, and then the temporal changes in the ring radius of each stimulus in the memory subspace and the temporary subspace were calculated. The radius of each stimulus was depowered and normalized by a session-dependent factor. In this way, the radius scores calculated from different sessions were comparable while maintaining proportionality within the same session. To test the correlation in step 1(a) ( Figure 3 D, bold blue arrow in step 1), the single trials in each session are divided into two groups (the first 50% of the trials belong to group H and the rest of the trials belong to group L) according to the radius of the S-1 signal in the rank-1 subspace before the rule cue starts (upstream radius score), and it is tested whether the radius of S-1 in the temp-2 subspace during the exchange process (downstream radius score) of group H is still significantly larger than that of group L in all sessions. The same test is also applied to the pairings listed below. Step 1(b): the radius of S-2 in rank 2 before the rule cue starts and the radius of S-2 in temp 1 during the exchange process; Step 2(a): the radius of S-1 in temp 2 during the exchange process and the radius of S-1 in rank 2 after the exchange; Step 2(b): the radius of S-2 in temp 1 during the exchange process and the radius of S-2 in rank 1 after the exchange.

[0235] Neuron contribution to subspace and normalized participation ratio ( Figure 4 A to C)

[0236] In order to quantify the contribution of a single neuron to each rank-WM or temporary subspace, each single neuron e i The unit vector of the axis is projected onto the given subspace r (the axis is and ), and define the squared scalar projection as the contribution of the neuron to the corresponding subspace ( Figure 4 A) In other words

[0237]

[0238] For simplicity, the contribution of the i-th neuron to the subspace r is expressed as The normalized participation ratio was used to quantify the proportion of neurons that contributed significantly to subspace r per session ( Figure 4 B) For example, if in a given session NPR r =0.4, then A ri The neurons in the top 40% are considered to contribute significantly to the subspace r.

[0239] NSA Index( Figure 4 D)

[0240] Neuronal contribution measures the alignment between a single neuron and the corresponding subspace. Since different subspaces are estimated with the same set of neurons, the neuronal contribution of the same neuron in different subspaces can be compared. To this end, the NSA index is calculated as follows

[0241]

[0242] Among them, a and b represent subspaces a and b respectively. Represents the contribution of the ith neuron to subspaces a and b, respectively. After calculating the NSA index between subspaces a and b, neurons from all recording sessions were pooled together, and only neurons that contributed significantly to subspaces a or b were included.

[0243] The NSA index ranges between -1 and 1. NSA values ​​close to 1 (or -1; 0.8 ≤ |NSA| ≤ 1.0) indicate that the neuron is mainly aligned with one subspace, and is therefore called a disjoint neuron between the two subspaces. In contrast, when NSA is close to 0 (|NSA| ≤ 0.2), the neuron is evenly aligned with both subspaces and is therefore an overlapping neuron. The proportions of disjoint and overlapping neurons in each subspace pair were obtained, and the chi-square test was used to compare the proportions of disjoint and overlapping neurons in any two subspace pairs (ranktemp, rank1-rank2, or temp1-temp2).

[0244] Rule-cue subspace analysis Figure 5 B to E)

[0245] For correct forward and reverse trials, 20 trials of each sequence were sampled separately, and the average activity from 0 to 400 ms after the onset of the rule cue was calculated. The average activity of the two rule types and the average activity of all selected time points were merged into one dimension to form a matrix of size 2T×N, where T is the number of time points. The matrix was subjected to PCA analysis, and the first three PCs were selected as the rule-cue subspace. To test the stability of the above identification procedure, it was repeated 100 times, resulting in 100 estimates for each PC. The VAF ratios between the rule-cue subspace and the WM subspace and between the rule-cue subspace and the temporary subspace were calculated (see the “Geometric relations between subspaces” section). As a control, the VAF between different PC estimates is provided. The rule-cue subspace reported in the results is the average of all estimates ( Figure 5 B).

[0246] After obtaining the rule-cue subspace, single-trial analysis was performed (see the “Analysis of Error Single-Trial Decoding” section) to examine the decoding accuracy of the rule in correct and incorrect trials ( Figure 5 C) The rule prompts the time to enter the subspace ( Figure 5 C) was defined as the time point at which the decoding probability was significantly different from the chance level (1 / 2) (P < 0.01).

[0247] Mathematical representation of the index system ( Figure 5 G)

[0248] In the proposed index control and switching system model, x rule (l1,l2,t) (length Nn), that is, the task-related component of the population neural activity of the sequence [l1,l2] at time t, can be expressed as

[0249]

[0250] Among them, I rule (t) (length Nn) is the neural activity in the rule-cue subspace, Ws (size Nn×ks, where ks is the subspace dimension) is the subspace basis, H (size ks×6) is a commonly used low-dimensional representation of the hexagonal structure, Tr(lr) (length 6, 1 on the lrth element and 0 on the other elements) is the click vector for item l, is the encoding strength dynamics of rank r in subspace s for a given trial type, Subs = {rank 1, rank 2, temp 1, temp 2}.

[0251] discuss

[0252] In the present invention, the inventors unexpectedly discovered for the first time the group representation and operation calculation mechanism of the frontal lobe neurons of primates (such as macaques) during working memory operations by training macaques to complete the task of sorting spatial sequences in working memory.

[0253] Working memory refers to a short-term memory. The phenomena and words that mammals see and hear in their daily lives will first enter the working memory and then be processed according to the current purpose. For example, people's subconscious understanding of language syntax and mental arithmetic of addition, subtraction, multiplication and division are all inseparable from the processing of information stored in working memory. However, although the storage principle of working memory has been widely and deeply studied, there are few studies on the control of working memory processing.

[0254] In order to further study the processing and control mechanism of working memory, the inventors chose the sorting of memory content as the research object. The inventors trained macaques to memorize a spatial sequence - this spatial sequence consists of points that flash continuously at different positions, and sort the sequence (forward or reverse sorting) according to the subsequent rule prompts, and finally report after a certain memory time. The inventors used a high-throughput electrode array to record the frontal cortex of the macaques and collected data from a large number of neurons during the task.

[0255] First, the inventors analyzed the representation of working memory signals in group neurons before the rule prompt appeared (memory period 1) and after the rule appeared (memory period 2). There are hundreds of millions of neurons in the brain firing simultaneously, and their activity intensity varies. Therefore, the state of the brain at every moment can be simplified to a certain extent into a point in this billion-dimensional space. If you only focus on a specific moment (such as the memory period), the "points" under different experimental conditions will form a complex and elusive geometric structure in this space. Previous studies by the research group have shown (Xie et al., Science, 2022) that when memorizing a spatial sequence, the geometric structure of memory representation in the brain is regular-information in different orders is represented in different sub-planes (or subspaces), and its representation structure faithfully reflects the structure of external stimuli. Moreover, the subspaces of different orders are orthogonal to each other.

[0256] In this study, the inventors found that when the spatial sequence memory information is reordered, the position information in the subspaces of different orders in memory period 1 will disappear from the original subspace, and then the original information will be placed in the subspace of the corresponding order according to the new order in memory period 2 - just like two boxes containing red balls and green balls respectively, when the sorting occurs, the red and green balls disappear from their own boxes together, and then magically appear in the other's box. The inventors further verified through cross-time decoding analysis that in the subspaces of different orders, if no sorting operation occurs, the original spatial memory information of each remains stable over time, and if sorting occurs, the spatial memory information of each is exchanged during the process from memory period 1 to 2. The analysis of error trials also shows that if rearrangement occurs when sorting should not occur, these erroneous exchange signals can also be found in the corresponding space.

[0257] The inventors further discovered that when exchanging information between different subspaces, each subspace recruited an additional new subspace to temporarily store its own information. Each subspace first passed the original internal memory information to the new temporary subspace, and after its own internal memory was cleared, it passed the memory signal in the temporary subspace to the other party, thus completing the exchange of memory information between subspaces. The inventors proved that the temporary subspaces found were indeed used to store and exchange memory signals through the correlation analysis of signal intensity and dynamic time characteristics in temporary subspaces and sequential memory subspaces.

[0258] Finally, the inventors also discovered a subspace that represents the rules. In this subspace, the dynamic trajectories under different rules are significantly different. Through decoding generalization analysis and error trial analysis, it is shown that this space is indeed related to the sorting process. Therefore, it is speculated that the rule subspace may be the one that connects the specific rule signals from the outside world with the abstract dynamic trajectories inside, and then used to initiate and gate the exchange process. This space controls the flow of information between the order subspace and the temporary subspace, causing these subspaces to undergo different dynamic processes under different sorting rules.

[0259] All documents mentioned in this application are considered to be included in the disclosure of this application as a whole, so that they can be used as the basis for modification when necessary. In addition, it should be understood that after reading the above disclosure of this application, those skilled in the art can make various changes or modifications to this application, and these equivalent forms also fall within the scope of protection claimed in this application.

Claims

1. A method for evaluating working memory, characterized in that: include: (a) providing a sample sequence to a subject for learning working memory, wherein the sample sequence includes a plurality of sequence items, and the subject needs to sequentially learn the plurality of sequence items for working memory; (b) After the first delay period, provide the subjects with a rule reminder; (c) after a second delay period, instructing the subject to report a result sample sequence, wherein the result sample sequence is obtained by the subject following the prompt of the rule and processed by the subject's nervous system; (d) judging the working memory of the subject based on the result sample sequence.

2. The method according to claim 1, characterized in that At the same time, neural activity signals of the subject's neurons or group neurons during the first delay period and the second delay period are recorded and stored through neural electrodes.

3. The method according to claim 2, characterized in that By analyzing the neural activity signals of neurons or groups of neurons in the first delay period, a plurality of sequence item working memory subspaces for representing information of the plurality of sequence items are obtained.

4. The method according to claim 3, characterized in that In step (c), when the subject's nervous system is processing, the information in the plurality of sequence item working memory subspaces is transferred to the corresponding plurality of sequence item temporary subspaces, and the information in the plurality of sequence item working memory subspaces is cleared.

5. The method according to claim 4, characterized in that Then, the information of the multiple temporary subspaces is transferred to the multiple sequence item working memory subspaces, thereby achieving exchange.

6. A system for evaluating working memory, characterized in that include A working memory sequence item input module, wherein the working memory sequence item input module is configured to input, transmit or provide a sequence for learning working memory to the subject, wherein the sequence includes ≥2 sequence items; A rule prompt module, wherein the rule prompt module is configured to send a rule prompt to the subject; and A working memory test module, wherein the working memory sequence item input module is configured to feedback the learning working memory results of the subject based on the rule prompts; and An optional timing control module is configured to: after the working memory sequence items are input, set a predetermined delay period I, a rule prompt period, and a predetermined delay period II, wherein the subject begins to provide feedback on the learning working memory results during the predetermined delay period II.

7. A method for enhancing information transfer within a nervous system, characterized in that: include: (a) Locate the temporary storage neurons corresponding to each information unit; (b) During the process of the neural system processing random events, the temporary storage neurons corresponding to the information units associated with the random events are activated through neural electrodes.

8. A method for enhancing information transfer within a nervous system, characterized in that: include: (a) Connecting neural electrodes to the nervous system enables reading and regulating the activation status of neurons in various regions of the nervous system; (b) inputting each information unit in the information unit set into the nervous system through the sensory organs of the organism in sequence, thereby locating the activated neurons corresponding to each information unit input through the neural electrodes; (c) inputting a first event into the nervous system through a sensory organ of the organism, wherein the first event is a subset of the information unit set; and then reading and locating temporary storage neurons of the nervous system for storing each information unit describing the first event through the neural electrodes; (d) repeating step (c), sequentially inputting a plurality of events different from the first event into the nervous system through the sensory organs of the organism, each of the plurality of events being a subset of the set of information units; (e) When a random event in the external environment is input into the nervous system through the sensory organs of the organism and activates the first plurality of temporary storage neurons, the unactivated second plurality of temporary storage neurons of the nervous system are activated through the neural electrodes.

9. The method according to claim 8, characterized in that In the step (e), the second plurality of temporary storage neurons activated by the neural electrodes are temporary storage neurons corresponding to the information units describing the association of the random events.

10. The method according to claim 9, characterized in that The information unit associated with the random event comes from the artificial intelligence system processing the random event; or the information unit associated with the random event comes from a preset event.

Images