A brain-like chip based on content-addressable memory and a connection representation method thereof

Abstract

The application discloses a kind of brain-like chip intersection based on content addressing memory and its connection representation method, including multiple neuron nodes and multiple corresponding on-chip routing units;The on-chip routing unit is used to select and determine the transmission path of the pulse data transmitted between neuron nodes and forward pulse data;Each neuron node includes multiple neurons, axon structure based on content addressing memory and dendrite structure based on content addressing memory, and the connection relationship between each neuron and the node where target neuron is located is constructed in the axon structure of source neuron, and the connection relationship between source neuron and neuron in node is constructed in the dendrite structure of target neuron, so that the flexible connection relationship between neurons and its connection characteristics can be realized, and the efficient routing between neurons is improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial neural network technology, specifically relating to a neuromorphic chip based on content-addressable memory and its connection representation method. Background Technology

[0002] Content-Addressable Memory (CAM) is a special type of memory that addresses data based on its content. CAM differs from ordinary SRAM (Static Random-Access Memory) lookup tables in that it employs a different addressing method. In CAM, the main working mechanism involves automatically and simultaneously comparing the input data item with all data items stored within it to determine if the input data item matches a specific data item in the CAM, and then outputting the corresponding match information. This content-addressing method makes CAM perform exceptionally well in high-speed lookup and matching tasks. Ordinary SRAM lookup tables, on the other hand, typically use address-based addressing, directly retrieving the corresponding data based on the memory address.

[0003] Traditional deep neural networks (DNNs) employ dense connections, where each neuron is connected to all neurons in the previous layer. This fully connected structure offers limited flexibility. Traditional DNNs are deep learning models based on optimization methods such as weight matrices and gradient descent, which present several shortcomings and challenges. They require significant computational resources; training DNNs demands substantial computational resources, leading to a dependence on high-performance hardware, which may be impractical or uneconomical in some environments. They also have high data requirements, typically requiring large amounts of labeled data for training to learn effective feature representations. In some domains, obtaining large-scale labeled data is extremely difficult, limiting the model's generalization ability. Furthermore, they are highly dependent on environmental conditions; deep neural networks are sensitive to differences between training and testing environments, and even slight environmental changes can cause performance degradation. Finally, they suffer from poor interpretability; due to the complex structure of deep neural networks, explaining the model's decision-making mechanisms becomes very difficult.

[0004] As a third-generation neural network, spiking neural networks possess high biological interpretability and exhibit unique advantages in several aspects, including few-shot learning, low power consumption, and robustness. Inspired by biological neural networks, their network structure and pulse transmission methods demonstrate unique strengths in real-world learning tasks. With the deepening research into neuromorphic computing chips, spiking neural networks have become a research hotspot, and the industry has released several neuromorphic chips based on spiking neural networks. These chips achieve tasks such as perception, learning, and decision-making by simulating pulse transmission, bringing new possibilities to the field of artificial intelligence.

[0005] On average, each neuron in the human brain is connected to 1,000 surrounding neurons, providing powerful computational and learning capabilities. Similarly, the design of neuromorphic chips requires handling and simulating such complex connections to construct vast networks from the intricate connections of a massive number of simple neurons, enabling complex functions such as perception, thought, and control. Therefore, the design of neuromorphic chips must consider the number of connections, weight adjustment mechanisms, and the formation and disappearance of dynamic connections. Flexible connection representation methods are a key factor in neuromorphic chip design, used to describe whether a connection exists between two neurons and its characteristics. This flexibility is a necessary condition for simulating brain structure and one of the key factors contributing to the outstanding performance of neuromorphic chips in various tasks. Summary of the Invention

[0006] In view of the above, the purpose of this invention is to provide a neuromorphic chip based on content-addressable memory and its connection representation method, so as to realize efficient connection and routing between multiple neuron nodes in the neuromorphic chip.

[0007] To achieve the above-mentioned objectives, this invention provides a neuromorphic chip based on content-addressable memory, comprising multiple neuronal nodes and their corresponding multiple on-chip routing units.

[0008] Each neuron node includes multiple neurons, an axonal structure based on content-addressable memory, and a dendritic structure based on content-addressable memory. The axonal structure of the source neuron contains the connection relationship between each neuron and the node where the target neuron is located, and the dendritic structure of the target neuron contains the connection relationship between the source neuron and the neurons within the node.

[0009] The on-chip routing unit is used to select and determine the pulse transmission and forwarding between neuron nodes.

[0010] Preferably, in the neuromorphic chip, the axonal structure is shared by all neurons within a neuronal node.

[0011] Preferably, in the neuromorphic chip, the axon structure includes a content-addressable memory, a first decoding module, and a target node identifier memory;

[0012] The content addressing memory stores the neuron identifier of the current neuron node. Different hit / missing information can be obtained by querying the content addressing memory through different neuron identifiers.

[0013] The first decoding module is used to decode the hit / miss information output by the content-addressable memory to obtain the access address of one or more target node identifier memories;

[0014] The target node identifier memory outputs the stored target node identifier according to the access address output by the first decoding module.

[0015] Preferably, in the neuromorphic chip, the content addressing memory compares the input index neuron representation with each bit of the stored neuron identifier using an NAND gate circuit, and outputs the final hit or missing information using an AND gate circuit.

[0016] Preferably, in the neuromorphic chip, the dendritic structure is shared by multiple neurons within a neuronal node, supporting both inference and learning modes. In inference mode, the backward connectivity is obtained using the representation method of the dendritic structure, i.e., the connection relationship and connection strength of the source neuron that triggered the pulse in one or more neurons in the current neuronal node are obtained. In learning mode, the forward connectivity is obtained using the representation method of the dendritic structure, i.e., the connection relationship and connection strength of the neuron that received or triggered a pulse in the current node in the current time period with one or more neurons in the current neuronal node or other neuronal nodes are obtained.

[0017] Preferably, in the neuromorphic chip, the dendritic structure includes a backward-connected content-addressable memory, a forward-connected content-addressable memory, a second decoding module, and a connectivity feature memory;

[0018] The backward-connected content addressing memory is used in inference mode to store source node identifiers and source neuron identifiers. Different hit / missing information is obtained by querying the backward-connected content addressing memory using the source node representation and source neuron identifier contained in the received pulse information.

[0019] The forward connection content addressing memory is used in learning mode to store the current neuron identifier. Different hit / missing information can be obtained by querying the content addressing memory through different neuron identifiers.

[0020] The second decoding module is used to decode the hit / miss information output by the backward join content addressing memory and the forward join content addressing memory to obtain the query address of one or more join feature memories;

[0021] The connectivity feature memory outputs the corresponding target neuron identifier and connectivity strength based on the query address output by the second decoding module.

[0022] To achieve the above objectives, the present invention provides a connection representation method for neuromorphic chips based on content-addressable memory, comprising:

[0023] In neuromorphic chips, an axonal structure based on content-addressable memory and a dendritic structure based on content-addressable memory are constructed for each neuron node. In the axonal structure of the source neuron, the connection relationship between each neuron and the node where the target neuron is located is constructed, and in the dendritic structure of the target neuron, the connection relationship between the source neuron and the neurons within the node is constructed.

[0024] Preferably, in the connection method, the constructed axonal structure is shared by all neurons within a neuronal node, and includes a content-addressable memory, a first decoding module, and a target node identifier memory;

[0025] The content addressing memory is queried by different neuron identifiers, and compared with the neuron identifier of the current neuron node stored therein to obtain different hit / missing information;

[0026] The first decoding module decodes the hit / miss information output by the content-addressable memory to obtain the access address of one or more target node identifier memories;

[0027] The target node identifier is output by the target node identifier memory based on the access address output by the first decoding module.

[0028] Preferably, in the connection method, the constructed dendritic structure is shared by multiple neurons within a neuron node, supporting both inference and learning modes. In inference mode, the backward connectivity is obtained using the representation method of the dendritic structure, i.e., the connectivity and connection strength of the source neuron that triggered the pulse in one or more neurons in the current neuron node are obtained. In learning mode, the forward connectivity is obtained using the representation method of the dendritic structure, i.e., the connectivity and connection strength of the neurons in the current node that received or triggered pulses in the current time period and one or more neurons in the current neuron node or other neuron nodes are obtained.

[0029] Preferably, in the connection method, the constructed dendritic structure includes a backward connection content addressing memory, a forward connection content addressing memory, a second decoding module, and a connection feature memory;

[0030] In inference mode, the memory is accessed by querying the backlink content addressing memory through the source node representation and source neuron identifier contained in the received pulse information, and compared with the stored source node identifier and source neuron identifier to obtain different hit / missing information;

[0031] In learning mode, the content addressing memory is queried through different neuron identifiers, and compared with the stored current neuron identifier to obtain different hit / missing information;

[0032] The second decoding module decodes the hit / miss information output from the backward join content addressing memory and the forward join content addressing memory to obtain the query address of one or more join feature memory.

[0033] The corresponding target neuron identifier and connection strength are output by connecting the characteristic memory according to the query address output by the second decoding module.

[0034] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0035] In the neuromorphic chip solution provided by this invention, an axonal structure based on content-addressable memory and a dendritic structure based on content-addressable memory are constructed for each neuron node containing multiple neurons. In the axonal structure of the source neuron, the connection relationship between each neuron and the node where the target neuron is located is constructed, and in the dendritic structure of the target neuron, the connection relationship between the source neuron and the neurons within the node is constructed. This can realize the flexible connection relationship and connection characteristics between neurons, and improve the efficient routing between neurons. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the structure of a neuromorphic chip based on content-addressable memory provided in the embodiment;

[0038] Figure 2 This is a schematic diagram of the content-addressable memory structure provided in the embodiment;

[0039] Figure 3 This is a schematic diagram of the axon structure provided in the embodiment;

[0040] Figure 4 This is a schematic diagram of the dendritic structure provided in the embodiment. Detailed Implementation

[0041] 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 embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0042] like Figure 1 As shown, the neuromorphic chip based on content-addressable memory (CAM) provided in the embodiment includes multiple neuron nodes and their corresponding multiple on-chip routing units. Each neuron node includes multiple neurons, a CAM-based axonal structure, and a CAM-based dendritic structure. In the axonal structure of the source neuron, the connection relationship between each neuron and the node where the target neuron is located is constructed. In the dendritic structure of the target neuron, the connection relationship between the source neuron and the neurons within the node is constructed.

[0043] The on-chip routing unit is used to select and determine the transmission direction of pulse data between neuron nodes and to forward the pulse data. Figure 1 A neuron located at the (0,0) node emits pulse data. The connection between the neuron and the target neuron at the (N,N) node can be obtained through the axonal structure in the neuron node. The corresponding pulse data is then forwarded to the (N,N) node through the on-chip routing unit. The connection between the source neuron and the neurons within the node can be obtained through the dendritic structure in the neuron node.

[0044] In the embodiments, both the axonal structure and the dendritic structure adopt the following... Figure 2 The CAM shown has a storage unit that stores identification information for indexing. An XOR gate is used to compare each bit of the index identifier with the stored identifier to see if they match. An AND gate outputs the final hit / miss result.

[0045] In the embodiments, the CAM-based axonal structure is shared by all neurons within a neuronal node, such as... Figure 3 As shown, the system includes modules such as a CAM (Cellular Array), a first decoding module, and a target node identifier memory. When a neuron fires a pulse, it queries the CAM using its own neuron identifier information. By comparing this information with the neuron identifier of the current neuron node stored in the CAM, one or more hit / missing information can be obtained. The first decoding module decodes the hit / missing information output by the CAM to obtain access addresses for one or more target node identifier memories. The target node identifier memory outputs the stored target node identifiers based on the access addresses output by the decoding module.

[0046] Figure 3 The existence of a connection between neuron A in the source node and a neuron in the target node X indicates that a neuron is connected to a neuron within a single node. Figure 3 Neuron B in the source node is connected to neurons Y and Z in the target node, indicating that a neuron has connections with neurons in multiple nodes. After querying the CAM using neuron A, only the first entry is matched. Therefore, the first decoding module outputs the corresponding address as the access address for the target node's identifier memory. Querying through this access address reveals that the target node is X. After querying the CAM using neuron B, two entries are matched. Therefore, the decoding module outputs two corresponding addresses as access addresses for the target node's identifier memory. Querying through these access addresses reveals that the target nodes are Y and Z.

[0047] In the embodiments, the CAM-based dendritic structure is shared by all neurons within a single neuron node, such as... Figure 4As shown, it includes modules such as a backward connection CAM, a forward connection CAM, a second decoding module, and a connection feature memory. The dendritic structure supports the expression of connection relationships under two different working modes: inference and learning.

[0048] In inference mode, the backward connectivity relationship is obtained using the dendritic structure representation method, that is, the connection relationship and connection strength of one or more neurons in the current node of the source neuron that triggers the pulse are obtained: When the pulse data packet emitted by the source neuron arrives at the target node, the source node identifier and source neuron identifier information in the data packet are used to query the backward connectivity CAM. By comparing it with the stored source node identifier and source neuron identifier, one or more hit / missing information can be obtained; The second decoding module decodes the hit / missing information output by the backward connectivity CAM to obtain the query address of one or more connectivity feature memories; The connectivity feature memory outputs the corresponding target neuron identifier and connection strength and other connectivity features according to the query address output by the decoding module.

[0049] Figure 4 In the first case, neuron A from node W is connected to neuron D in the current node, indicating that the source neuron is connected to only one neuron in the current node. Neuron B is connected to neurons E and F in the current node, indicating that the source neuron is connected to multiple neurons in the current node. When pulse data originating from A reaches the current node, a query of the backward connection CAM finds one entry. Therefore, the second decoding module outputs the corresponding address as the destination connection characteristic memory. A query using this address reveals that the neuron connected to it is neuron D in the current node, with a connection strength of W0. Similarly, when pulse data originating from B reaches the current node, a query of the backward connection CAM finds two entries. Therefore, the second decoding module outputs two corresponding addresses as the destination connection characteristic memory. A query using these addresses reveals that the neurons connected to it are neurons E and F in the current node, with connection strengths of W1 and W2, respectively.

[0050] In learning mode, the forward connectivity is obtained using a dendritic structure representation method. This involves acquiring the connections and connection strengths between neurons in the current node that received or generated pulses during the current time period and one or more neurons in the current node or other nodes. If a neuron in the current node received or generated a pulse during the current time period, its forward connections and their characteristics need to be acquired, and the relevant connection strengths updated. The forward connectivity CAM is queried using the neuron's identifier. By comparing this with the stored current neuron identifier, one or more hit / missing information can be obtained. The second decoding module decodes the hit / missing information output by the forward connectivity CAM to obtain query addresses for one or more connectivity characteristic memories. The connectivity characteristic memory outputs the corresponding target neuron identifier and connection strength based on the query addresses output by the decoding module.

[0051] Figure 4 A neuron connected to neuron D indicates a single connection with another neuron; two neurons connected to neuron E indicate multiple connections with other neurons. In the diagram, when neuron D in the current node receives or is excited by a pulse during the current time period, and its connection strength needs updating, the neuron identifier D is used to query the forward connection CAM. One entry is found, so the second decoding module outputs the corresponding address as the destination connection characteristic memory. The connection strength W0 can be obtained by querying this address. Similarly, when neuron E in the current node receives or is excited by a pulse during the current time period, and its connection strength needs updating, the neuron identifier E is used to query the forward connection CAM. Two entries are found, so the second decoding module outputs two corresponding addresses as the destination connection characteristic memory. The connection strengths W1 and W2 can be obtained by querying these addresses.

[0052] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A neuromorphic chip based on content-addressable memory, characterized in that, It includes multiple neuron nodes and their corresponding multiple on-chip routing units; Each neuron node includes multiple neurons, an axonal structure based on content-addressable memory, and a dendritic structure based on content-addressable memory. The axonal structure of the source neuron contains the connection relationship between each neuron and the node where the target neuron is located, and the dendritic structure of the target neuron contains the connection relationship between the source neuron and the neurons within the target neuron node. The axon structure includes a content-addressable memory, a first decoding module, and a target node identifier memory. The content-addressable memory stores the neuron identifiers of the current source neuron node. Different hit or missing information is obtained by querying the content-addressable memory using different neuron identifiers. The first decoding module is used to decode the hit or missing information output by the content-addressable memory to obtain the access addresses of one or more target node identifier memories. The target node identifier memory outputs the stored target node identifiers according to the access addresses output by the first decoding module. The on-chip routing unit is used for the transmission and forwarding of pulse data between neuron nodes.

2. The neuromorphic chip based on content-addressable memory according to claim 1, characterized in that, The axonal structure is shared by all neurons within a neuronal node.

3. The neuromorphic chip based on content-addressable memory according to claim 1, characterized in that, The content-addressable memory compares the input index neuron representation with each bit of the stored neuron identifier using an NAND gate circuit, and outputs the final hit or missing information using an AND gate circuit.

4. The neuromorphic chip based on content-addressable memory according to claim 1, characterized in that, The dendritic structure is shared by multiple neurons within a neuron node, supporting both inference and learning modes. In inference mode, the backward connectivity is obtained using the representation method of the dendritic structure, i.e., the connectivity and connection strength of the source neuron of the excitation pulse in one or more neurons of the current neuron node are obtained. In the learning mode, the forward connectivity is obtained by using the dendritic structure representation method, that is, the connection relationship and connection strength between the neurons in the current node that have received or been excited by pulses in the current time period and one or more neurons in the current neuron node or other neuron nodes.

5. The neuromorphic chip based on content-addressable memory according to claim 4, characterized in that, The dendritic structure includes a backward-connected content-addressable memory, a forward-connected content-addressable memory, a second decoding module, and a connection feature memory; The backward-connected content addressing memory is used in inference mode to store source node identifiers and source neuron identifiers. Different hit or missing information is obtained by querying the backward-connected content addressing memory using the source node representation and source neuron identifier contained in the received pulse information. The forward connection content addressing memory is used in learning mode to store the current neuron identifier. Different hit or missing information can be obtained by querying the content addressing memory through different neuron identifiers. The second decoding module is used to decode the query address of one or more connection feature memories based on the hit or missing information output by the backward join content addressing memory and the forward join content addressing memory; The connectivity feature memory outputs the corresponding target neuron identifier and connectivity strength based on the query address output by the second decoding module.

6. A connection representation method for a neuromorphic chip based on content-addressable memory, characterized in that, include: In neuromorphic chips, an axonal structure based on content-addressable memory and a dendritic structure based on content-addressable memory are constructed for each neuron node. In the axonal structure of the source neuron, the connection relationship between each neuron and the node where the target neuron is located is constructed. In the dendritic structure of the target neuron, the connection relationship between the source neuron and the neurons within the node of the target neuron is constructed. The constructed axonal structure is shared by all neurons within a neuron node, and includes a content-addressable memory, a first decoding module, and a target node identifier memory. The content addressing memory stores the neuron identifier of the current source neuron node. By querying the content addressing memory with different neuron identifiers, and comparing them with the stored neuron identifier of the current neuron node, different hit or missing information is obtained. The first decoding module decodes the hit or miss information output by the content-addressable memory to obtain the access address of one or more target node identifier memories. The target node identifier is output by the target node identifier memory based on the access address output by the first decoding module.

7. The connection representation method for neuromorphic chips based on content-addressable memory according to claim 6, characterized in that, The constructed dendritic structure is shared by multiple neurons within a neuron node, supporting both inference and learning modes. In inference mode, the backward connectivity is obtained using the representation method of the dendritic structure, that is, the connection relationship and connection strength of the source neuron of the excitation pulse in one or more neurons in the current neuron node are obtained. In the learning mode, the forward connectivity is obtained by using the dendritic structure representation method, that is, the connection relationship and connection strength between the neurons in the current node that have received or been excited by pulses in the current time period and one or more neurons in the current neuron node or other neuron nodes.

8. The connection representation method for neuromorphic chips based on content-addressable memory according to claim 7, characterized in that, The constructed dendritic structure includes a backward-connected content-addressable memory, a forward-connected content-addressable memory, a second decoding module, and a connection feature memory; In inference mode, the memory is accessed by querying the backlink content addressing memory through the source node representation and source neuron identifier contained in the received pulse information, and compared with the stored source node identifier and source neuron identifier to obtain different hit or missing information. In learning mode, the content is queried in the memory by different neuron identifiers and compared with the stored current neuron identifiers to obtain different hit or missing information; The second decoding module decodes the hit or missing information output by the backward join content addressing memory and the forward join content addressing memory to obtain the query address of one or more join feature memory. The corresponding target neuron identifier and connection strength are output by connecting the characteristic memory according to the query address output by the second decoding module.

Images