A multi-object hyper-veector probability decomposition method, device and system

Abstract

This invention discloses a method, apparatus, and system for probabilistic decomposition of multi-object hyperdimensional vectors, belonging to the fields of neural symbolic computation, hyperdimensional computation, and in-memory computation technology. It can improve the decomposition accuracy, efficiency, and hardware energy efficiency of complex hyperdimensional representations. The method includes: obtaining a query hyperdimensional vector to be decomposed; performing unbinding processing on the query hyperdimensional vector to obtain similarity results corresponding to each candidate hyperdimensional item under the category to be decomposed; performing candidate item selection based on the similarity results; continuing to perform unbinding processing and candidate item selection on the remaining categories to obtain decomposition results corresponding to multiple categories under the same object; reconstructing a single-object hyperdimensional vector based on the decomposition results corresponding to multiple categories under the same object to obtain a residual vector; repeatedly performing unbinding, candidate item selection, object reconstruction, and residual update on the residual vector until the residual vector meets a preset termination condition.

Description

Technical Field

[0001] This invention belongs to the fields of neural symbolic computation, hyperdimensional computation and in-memory computation technology, and specifically relates to a method, device and system for multi-object hyperdimensional vector probabilistic decomposition. Background Technology

[0002] Neurosymbolic artificial intelligence, by combining the perceptual capabilities of neural networks with the reasoning abilities of symbolic systems, can outperform purely data-driven methods in scenarios such as structured cognition, combinatorial reasoning, and interpretable intelligence. Hyperdimensional computing utilizes high-dimensional hyperdimensional vectors to represent symbols, relations, and combinatorial semantic information, and achieves reasoning through binding, superposition, debinding, and similarity calculation. Therefore, it is considered an important implementation paradigm adapted to neurosymbolic computing.

[0003] In existing hyperdimensional reasoning tasks, an increasing number of applications require the simultaneous processing of multiple objects and their hierarchical attribute information. For example, in structured visual reasoning tasks, a single image often contains multiple objects, each with attributes such as category, location, color, size, type, or even deeper levels. To describe these complex relationships, existing technologies have proposed hyperdimensional representation models oriented towards multi-object, multi-subclass hierarchical structures, capable of uniformly encoding multiple objects and their attribute combinations in hyperdimensional space.

[0004] However, when multiple objects are superimposed onto the same query hyperdimensional vector, significant superposition noise arises between the hyperdimensional items, making the decomposition process susceptible to the "curse of super-superposition" and the problem of duplicate items appearing repeatedly. As the number of objects, categories, or hierarchical depth increases, the separability between candidate items decreases, and the accuracy and efficiency of traditional decomposition methods both decline significantly.

[0005] Existing decomposition methods typically employ candidate selection based on a fixed threshold or a search strategy based on exhaustive combination. On the one hand, fixed thresholds often need to be readjusted under different task scales, numbers of categories, and numbers of objects, resulting in poor adaptability. On the other hand, exhaustive combination or a large number of iterative similarity searches introduce high latency and energy consumption, especially for multi-object composite hyperdimensional vectors.

[0006] Furthermore, existing hyperdimensional decomposition algorithms are typically implemented in von Neumann architectures, requiring frequent movement of hyperdimensional vector data between memory and processing units. Since unbinding and similarity calculation involve numerous multiply-accumulate or equivalent multiply-accumulate operations, this type of implementation faces a significant memory wall problem. In neural symbolic systems, even if the front-end perceptual model has output the query hyperdimensional vector, subsequent decomposition stages still consume substantial memory access and computational resources, resulting in limited overall system energy efficiency.

[0007] Therefore, how to propose a decomposition method and implementation structure that can improve the accuracy of multi-object composite hyperdimensional vector decomposition, reduce fixed threshold dependence, and further adapt to low-power hardware execution has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention aims to provide a method, apparatus, and system for multi-object hyperdimensional vector probabilistic decomposition, which can reduce the dependence of the multi-object hyperdimensional vector decomposition process on fixed threshold adjustment and high-precision numerical recovery, and improve the decomposition accuracy, decomposition efficiency, and hardware energy efficiency of complex hyperdimensional representations.

[0009] To achieve the above objectives, the present invention may adopt the following specific technical solutions: The primary objective is to provide a method for probabilistic decomposition of multi-object hyperdimensional vectors, comprising the following steps: Step 1: Obtain the query hyperdimensional vector to be decomposed, which is used to represent the composite semantic information corresponding to multiple objects; Step 2: For the category to be decomposed, perform unbinding processing on the query hyperdimensional vector to obtain the similarity results corresponding to each candidate hyperdimensional item under the category to be decomposed; Step 3: Perform candidate selection based on the similarity results. In the first round of decomposition, deterministic selection is adopted. Starting from the second round of decomposition, the similarity results are mapped to a candidate probability distribution, and target candidates are selected based on the candidate probability distribution. Step 4: Replace the corresponding category label with the selected target candidate, and continue to perform unbinding and candidate selection on the remaining categories to obtain the decomposition results corresponding to multiple categories under the same object; Step 5: Reconstruct the single-object hyperdimensional vector based on the decomposition results corresponding to multiple categories under the same object, and remove the single-object hyperdimensional vector from the current query hyperdimensional vector to obtain the residual vector; Step 6: Repeat the unbinding, candidate selection, object reconstruction and residual update processes on the residual vector until the residual vector meets the preset termination condition.

[0010] Furthermore, the object represented by the query hyperdimensional vector contains category, subcategory, or multi-level attribute information. The method performs candidate selection in order from high level to low level, and narrows the search range of the lower level based on the decomposition results of the upper level.

[0011] Furthermore, in step 3, the candidate probability distribution is obtained by normalizing the similarity results of the candidate hyperdimensional items through a normalized mapping, a competitive mapping, or an activation mapping.

[0012] Furthermore, in step 4, in the subsequent decomposition stage, the selected target candidates are used to replace the corresponding category labels, including: For categories that have been decomposed, the hyperdimensional vectors of the corresponding target candidates are used to participate in the subsequent unbinding operations of the categories; For categories that have not yet been decomposed, the hyperdimensional vectors of the corresponding category labels continue to be used in the unbinding operation.

[0013] Furthermore, in step 6, the preset termination condition includes at least one of the following: The sum of the absolute values ​​of the elements of the residual vector is lower than a preset residual threshold; The similarity between the residual vector and the noise reference vector satisfies the noise determination condition; The number of objects obtained from the decomposition has reached the preset limit; When the residual vector still does not meet the termination condition after reaching the upper limit of the number of objects, the initial query hyperdimensional vector is restored, and at least one round of iterative decomposition based on the candidate probability distribution is re-executed in the unbinding, candidate selection, object reconstruction and residual update process of steps 2-6 to achieve automatic error correction.

[0014] Furthermore, the similarity calculation is performed through a ferroelectric in-memory computational structure, which includes: Multiple ternary storage units, each of which includes a first ferroelectric transistor and a second ferroelectric transistor, wherein the first ferroelectric transistor and the second ferroelectric transistor store positive weights, negative weights or zero weights respectively through different threshold state combinations; An input driving circuit is used to input the signed input corresponding to the query hyperdimensional vector into multiple three-valued storage units; The positive accumulation path and the negative accumulation path are used to accumulate the positive product result and the negative product result, respectively; A differential sensing circuit is used to obtain the similarity results of candidate hyperdimensional items based on the outputs of the positive and negative accumulation paths.

[0015] Furthermore, the ternary storage cell is a 2FeFET-4T structure, with multiple storage cells forming an array, and multiple columns of storage cells sharing some switching transistors; the input driving circuit inputs integer-type hyperdimensional vectors in a bit-division time-division manner, enabling the ferroelectric in-memory computing structure to perform signed phased similarity calculations on non-truncated hyperdimensional vectors after the superposition of multiple objects.

[0016] Furthermore, the ferroelectric in-memory computing structure also includes a probability selection peripheral circuit, which includes: Competitive networks are used to enhance the differentiation of candidates; A random signal generation circuit is used to generate random comparison thresholds; The comparison selection circuit is used to determine the target candidate based on the comparison result between the candidate output and the random comparison threshold.

[0017] The second objective is to provide a multi-object hyperdimensional vector probabilistic decomposition apparatus for performing the aforementioned multi-object hyperdimensional vector probabilistic decomposition method, including: The query vector acquisition module is used to obtain the query hyperdimensional vector to be decomposed; The unbinding module is used to perform unbinding processing on the query hyperdimensional vector for the category to be decomposed; The similarity calculation module is used to calculate the similarity between candidate hyperdimensional items and unbinding results; The probability selection module is used to perform deterministic selection in the initial decomposition stage and to select target candidates after mapping the similarity results to candidate probability distributions in subsequent decomposition stages. The object reconstruction module is used to reconstruct a single object's hyperdimensional vector based on the decomposition results corresponding to multiple categories. The residual update module is used to remove the single-object hyperdimensional vector from the current query hyperdimensional vector to form a residual vector. The termination judgment module is used to determine whether the preset termination conditions are met.

[0018] The third objective is to provide a multi-object hyperdimensional vector probabilistic decomposition system, comprising: a perceptual front end, the aforementioned multi-object hyperdimensional vector probabilistic decomposition device, and a ferroelectric in-memory computing structure. The perceptual front end is used to output a query hyperdimensional vector, which originates from image samples in a structured visual reasoning task. The multi-object neural symbolic reasoning device is used to perform probabilistic hyperdimensional vector decomposition. The ferroelectric in-memory computing structure is used to perform debinding-related similarity calculations and candidate probability selection.

[0019] Compared with the prior art, the present invention has the following advantages: (1) This invention reduces the strong dependence of the decomposition process on a fixed threshold by mapping the similarity results to a candidate probability distribution and performing probabilistic candidate selection; reduces the subsequent search space and improves the decomposition efficiency by replacing the category labels with the decomposed items through a class-by-class shrinkage mechanism; and improves the decomposition accuracy in complex multi-object scenarios by forming an automatic error-correcting iterative decomposition process through object reconstruction and residual elimination.

[0020] (2) This invention performs signed three-valued similarity calculation through ferroelectric in-memory computing structure, which reduces data transfer overhead and the need for high-precision numerical recovery; by utilizing the sparse features of query hyperdimensional vectors, it achieves lower power consumption and higher hardware execution efficiency.

[0021] (3) This invention is applicable to scenarios such as structured visual reasoning, neural symbolic reasoning and edge intelligent reasoning, and has good versatility and scalability. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall process of the multi-object neural symbol reasoning method of the present invention; Figure 2 This is a schematic diagram of the probabilistic hyperdimensional vector decomposition process and probability mapping function of the present invention; Figure 3 This is a schematic diagram of the 2FeFET-4T ternary in-memory computing structure of the present invention, wherein (a) is a structural diagram of the 2FeFET-4T in-memory computing unit, and (b) is a structural diagram of the 2FeFET-4T in-memory computing array composed of units; Figure 4 The diagram shows the peripheral circuit of the in-memory computation structure for probabilistic hyperdimensional vector decomposition of the present invention, wherein (a) is the driving circuit structure diagram, (b) is the accumulation circuit structure diagram, (c) is the sensing circuit structure diagram, and (d) is the probability selection circuit structure diagram. Figure 5 This is an application block diagram of the neural symbolic reasoning system of the present invention; Figure 6 This is a schematic diagram comparing the decomposition accuracy and decomposition time of the present invention with other methods as relevant parameters change. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] Example 1

[0025] like Figure 1 and Figure 2 As shown in the figure, this embodiment provides a multi-object hyperdimensional vector probabilistic decomposition method.

[0026] In one embodiment, the probabilistic hyperdimensional vector decomposition method includes the following steps: Step 1: Obtain the composite query hyperdimensional vector Q and determine the set of categories to be decomposed. ,in, This represents the total number of categories. For the th... There are 1 category, and the candidate set is 1 , This represents the maximum number of candidate options corresponding to a single category. The query hyperdimensional vector Q contains combined semantic information corresponding to multiple objects. Objects contain category, subcategory, or multi-level attribute information.

[0027] Step 2: For the category to be decomposed, perform unbinding processing on the query hyperdimensional vector to obtain the similarity results corresponding to each candidate hyperdimensional item under the category to be decomposed.

[0028] (1) Select the current category to be decomposed The query hyperdimensional vector Q is unbound using the category labels other than the current category to obtain the candidate response results corresponding to each candidate option under the current category. In one embodiment, the unbinding operation can be expressed as: , in, This indicates element-wise binding operations. Indicates the first The hyperdimensional vector of category labels corresponding to each category Indicates the first The first category The hyperdimensional vector corresponding to each candidate.

[0029] (2) Calculate the similarity between each candidate hyperdimensional vector and the candidate response result to obtain the candidate similarity set. .

[0030] Step 3: Perform candidate selection based on similarity results. In the first round of decomposition, the candidate with the highest similarity is selected as the target candidate for the current category; starting from the second round of decomposition, probability mapping is performed on the candidate similarity set to obtain the candidate probability distribution. Depending on the specific task requirements, the similarity of candidate options can be normalized, enhanced in competition, or activated to ensure that candidates with higher similarity have a higher probability of being selected.

[0031] Based on the candidate probability distribution, target candidates are determined through probability sampling, random threshold comparison, or probability competition.

[0032] Step 4: For categories that have been decomposed, use the corresponding target candidate to replace the original category label in the subsequent category unbinding operation; for categories that have not yet been decomposed, continue to execute steps 2 to 3 until the decomposition results corresponding to multiple categories under the same object are obtained. Through this substitution mechanism, the search space of subsequent categories can be continuously reduced during the category-by-category decomposition process, and interference caused by the superposition of multiple objects can be suppressed.

[0033] Step 5: After all categories corresponding to an object have been decomposed, reconstruct the object-level hyperdimensional vector of the object based on the target candidates corresponding to these categories. And subtract the object-level hyperdimensional vector from the current query hyperdimensional vector Q. To obtain the residual vector R.

[0034] Step 6: Determine whether the residual vector R meets the termination condition. If the sum of the absolute values ​​of the elements of the residual vector R is lower than the preset residual threshold, or if the residual vector R meets the noise determination condition, then the decomposition is terminated; otherwise, the residual vector R is used as a new query hyperdimensional vector, and the next round of decomposition is continued.

[0035] If the residual vector still does not meet the termination condition after reaching the upper limit of the number of objects, the initial query hyperdimensional vector is restored, and at least one round of iterative decomposition process including unbinding, candidate probabilistic selection, object reconstruction, and residual update is re-executed to achieve automatic error correction. Since subsequent rounds use a probabilistic approach to select candidates and repeatedly execute the above decomposition process through residual update, early erroneous selections can be corrected over multiple rounds, resulting in an automatic error correction effect.

[0036] In a further embodiment, the preset residual threshold is based on the dimension of the query hyperdimensional vector. and number of categories The preset residual threshold is determined and can be set by combining the residual statistics at the correct decomposition termination in the training samples, validation samples, or prior samples. Specifically, the preset residual threshold can be expressed as the ratio of the hyperdimensional vector dimension to the residual threshold. Proportional parameter form ,in, This is an empirical coefficient. In one implementation, A value of 0.03 can be used; in other embodiments, Adjustments can be made based on the statistical results of pre-experiments or validation sets, depending on the task size and number of categories. (The last sentence appears to be incomplete and unrelated to the preceding text.) As the value increases, the preset residual threshold can increase linearly accordingly.

[0037] In this embodiment, a single object not only contains multiple peer categories but also multi-level attribute relationships. For example, a certain category may also contain corresponding subcategories or secondary attribute items. In this case, the present invention performs decomposition in order from high-level to low-level. Specifically, after selecting target candidates for a certain upper-level category, the target candidate is used as a priori constraint for the lower-level decomposition, and similarity calculation and probability selection are performed only within the lower-level candidate range corresponding to that upper-level category. In this way, the low-level search space can be significantly reduced, avoiding the need to perform a full comparison of all lower-level items.

[0038] Therefore, although this invention allows for compatibility with query hyperdimensional vector inputs with hierarchical structures, its main inventiveness lies not in the hierarchical encoding itself, but in achieving a more robust decomposition process through probabilistic decomposition, class-by-class substitution, and residual error correction under complex hierarchical and multi-object superposition conditions.

[0039] Example 2

[0040] like Figure 3 and Figure 4 As shown, this embodiment provides a ferroelectric in-memory computing structure for performing similarity calculations based on Embodiment 1. The ferroelectric in-memory computing structure includes: multiple ternary storage units, an input driving circuit, a positive accumulation path and a negative accumulation path, a differential sensing circuit, and a probability selection peripheral circuit.

[0041] Each ternary memory cell includes a first ferroelectric transistor and a second ferroelectric transistor. The first and second ferroelectric transistors store positive, negative, or zero weights respectively through different threshold state combinations. The ternary memory cell is preferably a 2FeFET-4T structure, with multiple memory cells forming an array, and some switching transistors shared between multiple columns of memory cells. Sharing switching transistors among multiple memory cells reduces the array area. For highly sparsity input hyperdimensional vectors, the proportion of zero input is high, which can significantly reduce effective readout and accumulation power consumption.

[0042] The ternary storage cell can represent three storage states—positive, negative, and zero—by controlling the threshold state combinations of the first and second ferroelectric transistors. During the write phase, threshold state programming is completed by applying positive and negative voltage pulses to the corresponding control terminals. During the read or calculation phase, the input drive circuit inputs the signed input from the query hyperdimensional vector to the array, and the array generates positive and negative contribution currents based on the input and storage state. All positive contributions are accumulated in the positive accumulation path, and all negative contributions are accumulated in the negative accumulation path. The candidate similarity result is then obtained through a differential sensing circuit. Since the decomposition algorithm focuses more on the relative magnitude relationship between candidate options rather than necessarily recovering high-precision values, a simplified sensing and comparison structure can be adopted, reducing the requirement for a high-precision analog-to-digital converter.

[0043] The input driving circuit is used to input the signed input corresponding to the query hyperdimensional vector into multiple three-value storage units; the input driving circuit inputs the integer hyperdimensional vector in a bit-division time manner, so that the ferroelectric in-memory computing structure can perform signed phased similarity calculation on the non-truncated hyperdimensional vector after the superposition of multiple objects.

[0044] The positive and negative accumulation paths are used to accumulate the positive and negative product results, respectively.

[0045] A differential sensing circuit is used to obtain the similarity results of candidate hyperdimensional items based on the outputs of the positive and negative accumulation paths.

[0046] The probabilistic selection peripheral circuit includes a competition network, a random signal generation circuit, and a comparison selection circuit. The competition network is used to enhance the differences between candidate options; the random signal generation circuit is used to generate a random comparison threshold; and the comparison selection circuit is used to determine the target candidate option based on the comparison result between the candidate option output and the random comparison threshold.

[0047] With the above structure, the probability selection process corresponding to the candidate probability distribution in the algorithm can be implemented at the hardware level without having to recover the high-precision numerical probability for each candidate, thus making it more suitable for low-power hardware implementation.

[0048] Example 3

[0049] This embodiment provides a multi-object hyperdimensional vector probabilistic decomposition apparatus for executing the multi-object hyperdimensional vector probabilistic decomposition method of Embodiment 1, including: The query vector acquisition module is used to obtain the query hyperdimensional vector to be decomposed; The unbinding module is used to perform unbinding processing on the query hyperdimensional vector for the category to be decomposed; The similarity calculation module is used to calculate the similarity between candidate hyperdimensional items and unbinding results; The probability selection module is used to perform deterministic selection in the initial decomposition stage and to select target candidates after mapping the similarity results to candidate probability distributions in subsequent decomposition stages. The object reconstruction module is used to reconstruct a single object's hyperdimensional vector based on the decomposition results corresponding to multiple categories. The residual update module is used to remove the single-object hyperdimensional vector from the current query hyperdimensional vector to form a residual vector. The termination judgment module is used to determine whether the preset termination conditions are met.

[0050] Example 4

[0051] This embodiment provides a multi-object hyperdimensional vector probabilistic decomposition system, including: a sensing front end, a multi-object hyperdimensional vector probabilistic decomposition device of embodiment 3, and a ferroelectric in-memory computing structure of embodiment 2.

[0052] The perceptual front end is used to output the query hyperdimensional vector, which is derived from image samples in the structured visual reasoning task; the multi-object neural symbol reasoning device is used to perform probabilistic hyperdimensional vector decomposition; the ferroelectric in-memory computing structure is used to perform unbinding-related similarity calculation and candidate probability selection.

[0053] like Figure 5 The diagram shows an application block diagram of the neural symbolic reasoning system based on this invention. The perception front end receives image input and outputs the corresponding query hyperdimensional vector. The query hyperdimensional vector represents the combined attribute information of multiple objects in the image.

[0054] The probabilistic hyperdimensional vector decomposition module performs unbinding, similarity calculation, probability selection, object reconstruction, and residual update on the query hyperdimensional vector to recover the category and attribute information of each object in the image. The recovered object-level structure results can be further input into the subsequent symbolic reasoning module to complete tasks such as rule judgment, matrix reasoning, or structure completion. In this way, the structured visual reasoning task avoids the problem of poor interpretability in conventional neural network computation.

[0055] In this implementation, the sensing front end can be a neural network or other sensing modules that can output hyperdimensional vector representations; while the similarity calculation and candidate selection process at the back end can be performed by the aforementioned ferroelectric in-memory computing structure to reduce overall latency and energy consumption.

[0056] Example 5

[0057] To verify the effectiveness of the hyperdimensional vector decomposition method of the present invention, a comparative analysis was conducted on the present invention and related decomposition methods under the same experimental settings. Figure 6 The comparison results of different methods in terms of decomposition accuracy and decomposition time are shown. Figure 6 As shown, the present invention can shorten the time required for the decomposition process while ensuring a high decomposition accuracy, and can still maintain a good efficiency advantage, especially when the number of objects increases or the decomposition complexity increases.

[0058] Table 1

[0059] Table 2

[0060] Table 1 presents a comparison of the object decomposition results of this invention on the RAVEN dataset, and Table 2 presents a comparison of the inference results of this invention on the RAVEN dataset. Combined with... Figure 6 As can be seen from Tables 1 and 2, this invention can effectively improve the decomposition stability and task accuracy of multi-object composite hyperdimensional vectors by combining probabilistic selection, class-by-class substitution and residual iterative update. At the same time, by combining the ferroelectric in-memory computing structure, the data transfer overhead and hardware execution energy consumption in the similarity calculation stage can be further reduced.

[0061] When the preset upper limit for the number of objects is reached and the residual has not disappeared, the present invention can restore the original query hyperdimensional vector and re-execute the probabilistic decomposition of subsequent rounds, thereby utilizing the search diversity brought about by probabilistic selection to achieve automatic error correction and further improve the reliability of the final decomposition result.

[0062] This invention provides an efficient technical solution for multi-object composite hyperdimensional vector decomposition by combining probabilistic candidate selection, decomposed term replacement, object reconstruction, residual iterative update, and ferroelectric in-memory computation hardware. This solution not only improves the decomposition accuracy and robustness of hyperdimensional representations in neural symbolic reasoning tasks but also significantly reduces memory access and computational overhead at the hardware implementation level, demonstrating promising application prospects.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-object hyper-veector probability decomposition method, characterized in that, Includes the following steps: Step 1: Obtain the query hyperdimensional vector to be decomposed, which is used to represent the composite semantic information corresponding to multiple objects; Step 2: For the category to be decomposed, perform unbinding processing on the query hyperdimensional vector to obtain the similarity results corresponding to each candidate hyperdimensional item under the category to be decomposed; Step 3: Perform candidate selection based on the similarity results. In the first round of decomposition, deterministic selection is used. Starting from the second round of decomposition, the similarity results are mapped to a candidate probability distribution, and target candidates are selected based on the candidate probability distribution. Step 4: Replace the corresponding category label with the selected target candidate, and continue to perform unbinding and candidate selection on the remaining categories to obtain the decomposition results corresponding to multiple categories under the same object; Step 5: Reconstruct the single-object hyperdimensional vector based on the decomposition results corresponding to multiple categories under the same object, and remove the single-object hyperdimensional vector from the current query hyperdimensional vector to obtain the residual vector; Step 6: Repeat the unbinding, candidate selection, object reconstruction and residual update processes on the residual vector until the residual vector meets the preset termination condition.

2. The multi-object hyper-veector probability decomposition method according to claim 1, wherein, The object represented by the query hyperdimensional vector contains category, subcategory, or multi-level attribute information. The method performs candidate selection in order from high level to low level, and narrows the search range of the lower level based on the decomposition results of the upper level.

3. The multi-object hyperdimensional vector probabilistic decomposition method according to claim 1, characterized in that, In step 3, the candidate probability distribution is obtained by normalizing the similarity results of the candidate hyperdimensional items through a normalized mapping, a competitive mapping, or an activation mapping.

4. The multi-object hyperdimensional vector probabilistic decomposition method according to claim 1, characterized in that, In step 4, during the subsequent decomposition stage, the selected target candidates are used to replace the corresponding category labels, including: For categories that have been decomposed, the hyperdimensional vectors of the corresponding target candidates are used to participate in the subsequent unbinding operations of the categories; For categories that have not yet been decomposed, the hyperdimensional vectors of the corresponding category labels continue to be used in the unbinding operation.

5. The multi-object hyperdimensional vector probabilistic decomposition method according to claim 1, characterized in that, In step 6, the preset termination condition includes at least one of the following: The sum of the absolute values ​​of the elements of the residual vector is lower than a preset residual threshold; The similarity between the residual vector and the noise reference vector satisfies the noise determination condition; The number of objects obtained from the decomposition has reached the preset limit; When the residual vector still does not meet the termination condition after reaching the upper limit of the number of objects, the initial query hyperdimensional vector is restored, and at least one round of iterative decomposition based on the candidate probability distribution is re-executed in the unbinding, candidate selection, object reconstruction and residual update process of steps 2-6 to achieve automatic error correction.

6. A multi-object hyperdimensional vector probabilistic decomposition method according to any one of claims 1-5, characterized in that, Similarity calculation is performed through a ferroelectric in-memory computational structure, which includes: Multiple ternary storage units, each of which includes a first ferroelectric transistor and a second ferroelectric transistor, wherein the first ferroelectric transistor and the second ferroelectric transistor store positive weights, negative weights or zero weights respectively through different threshold state combinations; An input driving circuit is used to input the signed input corresponding to the query hyperdimensional vector into multiple three-valued storage units; The positive accumulation path and the negative accumulation path are used to accumulate the positive product result and the negative product result, respectively; A differential sensing circuit is used to obtain similarity results of candidate hyperdimensional items based on the outputs of the positive and negative accumulation paths.

7. The multi-object hyperdimensional vector probabilistic decomposition method according to claim 6, characterized in that, The ternary storage cell is a 2FeFET-4T structure, with multiple storage cells forming an array, and multiple columns of storage cells sharing some switching transistors; the input driving circuit inputs integer hyperdimensional vectors in a bit-division time-division manner, enabling the ferroelectric in-memory computing structure to perform signed phased similarity calculations on non-truncated hyperdimensional vectors after the superposition of multiple objects.

8. The multi-object hyperdimensional vector probabilistic decomposition method according to claim 6, characterized in that, The ferroelectric in-memory computing structure further includes a probability selection peripheral circuit, which includes: Competitive networks are used to enhance the differentiation of candidates; A random signal generation circuit is used to generate random comparison thresholds; The comparison selection circuit is used to determine the target candidate based on the comparison result between the candidate output and the random comparison threshold.

9. A multi-object hyperdimensional vector probabilistic decomposition apparatus, used to execute the multi-object hyperdimensional vector probabilistic decomposition method of claim 6, characterized in that, include: The query vector acquisition module is used to obtain the query hyperdimensional vector to be decomposed; The unbinding module is used to perform unbinding processing on the query hyperdimensional vector for the category to be decomposed; The similarity calculation module is used to calculate the similarity between candidate hyperdimensional items and unbinding results; The probability selection module is used to perform deterministic selection in the initial decomposition stage and to select target candidates after mapping the similarity results to candidate probability distributions in subsequent decomposition stages. The object reconstruction module is used to reconstruct a single object's hyperdimensional vector based on the decomposition results corresponding to multiple categories. The residual update module is used to remove the single-object hyperdimensional vector from the current query hyperdimensional vector to form a residual vector. The termination judgment module is used to determine whether the preset termination conditions are met.

10. A multi-object hyperdimensional vector probabilistic decomposition system, characterized in that, include: The system comprises a perception front-end, a multi-object hyperdimensional vector probabilistic decomposition device as described in claim 9, and a ferroelectric in-memory computing structure. The perception front-end is used to output a query hyperdimensional vector, which is derived from image samples in a structured visual reasoning task. The multi-object neural symbolic reasoning device is used to perform probabilistic hyperdimensional vector decomposition. The ferroelectric in-memory computing structure is used to perform unbinding-related similarity calculations and candidate probability selection.

Images