Dynamic resonance topology processing-based body, memory and calculation integrated calculation architecture and method

Abstract

The invention discloses a dynamic resonance topology processing-based calculation architecture and method with integrated body, memory and calculation functions, and relates to the field of computer architecture, artificial intelligence and neuromorphic calculation. The architecture simulates a brain to encode information through a dynamic resonance topology of a neural node cluster, and the core comprises a sensing-action layer, a topology calculation layer adopting a layered heterogeneous dynamic network, and a global modulation system composed of a rhythm generator and a value regulator. The working method is based on predictive coding, self-organizing evolution of an internal topological structure is driven through a logic closed loop of perception-prediction-resonance-action-feedback-hierarchical learning, and therefore information processing, decision generation and lifelong learning are achieved. According to the method, the fundamental transformation of the calculation normal form from'instruction driving 'to'topological evolution' is realized, and the method has remarkable advantages in the aspects of energy efficiency, autonomous learning, generalization ability and real-time processing, and is suitable for edge intelligent scenes such as automatic driving and next-generation robots.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer architecture, artificial intelligence and neuromorphic computing, and in particular to a somatic intelligent computing architecture and method inspired by the brain neural network information processing mechanism and adopting dynamic resonance topology processing and storage-computing integrated technology. BACKGROUND

[0002] The current mainstream computing architecture is based on the Von Neumann architecture, which faces fundamental bottlenecks in processing unstructured data and achieving low-power real-time intelligence due to its "memory wall" problem and instruction-driven serial processing mode. Existing neuromorphic computing schemes mostly focus on simulating the pulse behavior of individual neurons or synapses, but fail to simulate the core characteristics of brain information processing from a system level: information is not stored in fixed locations, but encoded, stored and computed through dynamic, synchronous / asynchronous resonance activities in the transient topology structure formed by distributed neuron node clusters. At the same time, existing architectures generally lack deep integration design with physical bodies and environment (somaticity), making the emergence and learning of intelligence heavily dependent on offline massive data training, rather than online, interactive and autonomous evolution with the environment. SUMMARY

[0003] The present application aims to overcome the shortcomings of the prior art and provide a somatic storage-computing integrated computing architecture and method based on dynamic resonance topology processing. The architecture aims to simulate the core logic of the brain encoding information and driving behavior through dynamic resonance topology of neural node clusters, realize the paradigm shift from "instruction-driven computing" to "topology self-organizing evolution computing", and thus achieve breakthroughs in energy efficiency, real-time performance, autonomous learning and generalization ability.

[0004] To achieve the above purpose, the present application provides the following technical solutions:

[0005] The somatic storage-computing integrated computing architecture based on dynamic resonance topology processing comprises:

[0006] a sensor-action layer for event-driven interaction with the physical environment, which receives event signals of the physical environment through multi-modal sensors and generates sparse pulse signals only when changes in the environment are perceived;

[0007] a topology computing layer as a core information processing unit, which encodes information and performs computation through dynamically changing connection patterns between internal node clusters, the computation being the evolution process of the topology structure, and the topology computing layer comprising a topology computing core composed of a plurality of storage-computing integrated micro-columns;

[0008] a global modulation system for providing timing coordination and learning guidance signals for the self-organizing process of the topology computing layer;

[0009] An internal generative model integrated in the architecture and operating based on predictive coding principles, guides the dynamic evolution of the topological computing layer by comparing the error between the prediction signal output by the internal generative model and the actual sensory signal input by the sensor-actuator layer;

[0010] Wherein, the picosecond-level topological connection reorganization between the storage-computing integrated microcolumns in the topological computing core is realized through a reconfigurable interconnection network of silicon photon waveguide.

[0011] As a further technical solution of the present application, the topological computing layer adopts a hierarchical heterogeneous dynamic network structure, including:

[0012] An instinctive reflex sub-layer composed of hardwired digital / analog circuits, used for processing high real-time and high reliability reflex behaviors requiring millisecond-level response, the high reliability reflex behaviors including obstacle avoidance;

[0013] The topological computing core, each storage-computing integrated microcolumn is internally provided with a relatively fixed or slow plasticity memristor cross array, the memristor cross array is used for realizing weight storage, state memory and vector matrix multiplication calculation, and constitutes a basic functional unit of the architecture.

[0014] As a further technical solution of the present application, the global modulation system includes:

[0015] A rhythm generator for listening to the neural oscillation patterns spontaneously emerging in the topological computing layer, detecting and amplifying the neural oscillation patterns to form a global synchronous rhythm, and locking and stabilizing a specific node cluster resonance group through the global synchronous rhythm;

[0016] A value regulator for receiving a prediction error signal, quantizing the prediction error signal based on the amplitude of the prediction error signal, and releasing modulation signals of different intensities and ranges according to the quantization results to grade the plasticity of the topological computing layer.

[0017] As a further technical solution of the present application, the architecture performs information retrieval through a content addressing mechanism: broadcasts partial clue information to the topological computing layer, and activates the complete node cluster topology related to the clue information in the topological computing layer in parallel through a sparse diffusion activation mechanism.

[0018] As a further technical solution of the present application, the architecture supports an offline consolidation mechanism: during the idle period of the system without external task processing or the period simulating biological sleep, the topological activity patterns of important tasks previously marked are spontaneously replayed to strengthen the connections of the node clusters corresponding to the important tasks, and prevent catastrophic forgetting in the lifelong learning process.

[0019] A somatic storage and calculation integrated computing method based on dynamic resonance topology processing, comprising the following steps:

[0020] Step 1, the sensing-action layer perceives the event signal of the physical environment and generates a sparse pulse signal input topology computing layer, and activates the relevant node microcolumn cluster of the topology computing layer; at the same time, the internal generation model in the architecture predicts the next step of sensory input and action to be executed based on the activation state of the current topology computing layer, and forms an internal prediction;

[0021] Step 2, after the rhythm generator of the global modulation system detects the synchronous oscillation sketch of the node cluster related to the current task, a same-frequency global rhythm is generated and broadcasted, and a stable node cluster resonance group is formed in the topology computing layer;

[0022] Step 3, the mode of the node cluster resonance group is translated into specific actuator driving commands through the instinctive reflex sublayer or the motion interface, and an external action is generated;

[0023] Step 4, the sensing-action layer receives the feedback signal of the physical environment to the external action, compares the feedback signal with the internal prediction in step 1, and calculates the prediction error;

[0024] Step 5, the value regulator of the global modulation system adjusts the connection weight of the topology computing layer or the physical connection structure between the storage and calculation integrated microcolumns according to the amplitude of the prediction error, and completes a learning cycle; wherein small errors only fine-tune the synaptic weight within the resonance group, large errors guide the recombination of the physical connection between the silicon photon waveguide microcolumns, and large errors with punishment drive large-scale topology reconstruction.

[0025] The present application has the following advantages and beneficial effects:

[0026] 1. Paradigm-level innovation: from "instruction-driven" to "topology evolution-driven", which more essentially simulates the working mechanism of biological intelligence.

[0027] 2. Extremely high energy efficiency: event-driven, sparse computing, storage and calculation integration and other multiple technologies make it consume energy only when processing relevant information.

[0028] 3. Autonomous lifelong learning: through somatic interaction with the environment and prediction error driving, the system can continuously optimize itself without the need for a large amount of offline data labeling.

[0029] 4. Strong generalization and robustness: dynamic topology structure enables it to adapt to unknown scenes through recombination, and hierarchical and redundant design provides inherent fault tolerance capability.

[0030] 5. Real-time parallel processing: the parallel evolution of three-dimensional topology structure is naturally suitable for processing multi-modal and unstructured real-time information. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is the overall hierarchical block diagram of DyReTo-CPU architecture;

[0032] Figure 2 is the detailed diagram of hierarchical heterogeneous network structure of topology computation layer;

[0033] Figure 3 is the workflow diagram of “perception-resonance-action-learning” closed loop based on predictive coding;

[0034] Figure 4 is the schematic diagram of hierarchical plasticity mechanism. DETAILED DESCRIPTION

[0035] The application is further described below in conjunction with the embodiments. It should be noted that the embodiments are only examples and illustrations of the inventive concept, and those skilled in the art can make various modifications or supplements to the described embodiments or replace them with similar ways, as long as they do not deviate from the inventive concept or exceed the scope defined by the claims.

[0036] Referring to Figures 1-4 , the application discloses a body-stored-computing integrated computing architecture based on dynamic resonance topology processing, which is composed of four core layers to form a logical closed loop system:

[0037] 1. Sensing-Acting Layer:

[0038] Function: As the interface for the architecture to interact with the physical world, it contains multi-modal sensors (such as event-driven vision, Memristor tactile skin) and actuators.

[0039] Characteristics: Native support for event-driven mode, only when changes are perceived, sparse pulse signals are generated as the trigger source for upper topology evolution.

[0040] 2. Topology Computation Layer:

[0041] Function: As the core information processing unit, it encodes information and performs computation through the dynamic connection mode of its internal node cluster.

[0042] Structure: Realized by hierarchical heterogeneous dynamic network:

[0043] Innate reflex sub-layer: composed of hardwired digital / analog circuits, processing guaranteed reflex behaviors (such as obstacle avoidance) that require millisecond-level response, ensuring basic safety of the system.

[0044] Topology computing core: a three-dimensional network composed of a large number of storage-computing integrated microcolumns. Each microcolumn is internally a cross-array of fixed or slowly plastic memristors, realizing the basic functional unit; the microcolumns are connected through a silicon photonic waveguide rerouting network to realize fast (picosecond level) topology connection reorganization; the clusters are connected through an event-driven packet switching network to realize long-range and sparse communication.

[0045] 3. Global modulation system:

[0046] Function: Provide timing coordination and learning guidance for the self-organization process of the topology computing layer.

[0047] Composition:

[0048] Rhythm generator: listens to the spontaneous emergence of neural oscillations in the topology computing layer, detects and amplifies them, and forms a global synchronous rhythm (such as gamma wave, theta wave), locks and stabilizes the relevant resonance group in the form of "task invitation", and solves the feature binding problem.

[0049] Value regulator: receives feedback signals (such as rewards, punishments, prediction errors) from internal predictions and external environments, multi-threshold quantizes them, and releases global modulation signals (such as analog neurotransmitters) of different intensities and ranges to regulate the plasticity of the topology computing layer.

[0050] 4. (Implicit) prediction and learning engine:

[0051] Function: As the core algorithm logic of the architecture, it runs through the above hardware levels.

[0052] Mechanism: Run an internal generative model within the architecture, whose core is prediction coding. The system continuously predicts future sensory inputs and action results based on the current topology state. Prediction error is the fundamental driving force for the evolution of the entire system topology.

[0053] The present application also discloses an intelligent storage-computing integrated computing method based on dynamic resonance topology processing, which runs according to the following steps to form a complete "perception-thinking-action-learning" closed loop:

[0054] 1. Perception and prediction projection:

[0055] The sensor-action layer converts external events into sparse pulse signals, which are input to the topology computing layer to activate relevant node microcolumn clusters.

[0056] The internal generative model predicts the next sensory input and action to be performed based on the current activated topology pattern, and forms a "ready state" in the topology computing layer in advance.

[0057] 2. Resonance and action generation:

[0058] The rhythm generator detects the emergence of a synchronized oscillation in the node cluster associated with the current task, and generates and broadcasts a coherent rhythm signal, amplifying and locking the resonance group.

[0059] This stable resonance topology pattern, whose pattern itself represents a specific "perception-action concept", is "read" by the instinctive reflex sublayer or direct motor interface, and translated into specific actuator driving commands, producing external behavior.

[0060] 3、Feedback and hierarchical learning:

[0061] The system compares the actual sensory feedback with the prediction in step 1, producing a prediction error signal.

[0062] The value regulator quantifies this error signal:

[0063] Small error: triggers short-term plasticity, fine-tuning the weights of synapses within the resonance group without changing physical connections.

[0064] Large error: triggers long-term plasticity, releasing a modulatory signal to increase the plasticity of the relevant area, and guiding the silicon photonic network to reconfigure the physical connections between microcolumns, correcting the topology.

[0065] Giant error with penalty: drives large-scale topology reconfiguration and significantly increases the threshold for future activation of similar topologies.

[0066] 4、Memory consolidation and retrieval:

[0067] Content-addressable retrieval: when recalling a concept or skill is needed, partial clues (such as a feature of an object) are broadcast in parallel through the sparse diffusion activation mechanism in the topology computation layer, efficiently awakening the entire corresponding resonance topology.

[0068] Prevent catastrophic forgetting: the system spontaneously re-enacts the topology activity patterns of important tasks during idle periods (such as "sleep" state), accompanied by simulated consolidation signals, actively strengthening the connections corresponding to old skills, preventing them from being overwritten by new learning.

[0069] The above is an exemplary description of the invention, and obviously the specific implementation of the invention is not limited by the above method. As long as this non-essential improvement or direct application of the invention's concept and technical solution to other fields is adopted, it is within the scope of protection of the invention.

Claims

1. A somatic computing architecture based on dynamic resonance topology processing, characterized in that, Comprise: a sensor-action layer for event-driven interaction with the physical environment, which receives event signals of the physical environment through multi-modal sensors, and generates sparse pulse signals only when changes in the environment are perceived; a topology computing layer as the core information processing unit, which encodes information and performs calculations through dynamically changing connection patterns between internal node clusters, the calculations being the evolution process of the topology structure, and the topology computing layer comprising a topology computing core composed of a plurality of memory-computing micro-columns; a global modulation system for providing timing coordination and learning guidance signals for the self-organizing process of the topology computing layer; an internal generative model integrated in the architecture and operating based on predictive coding principles, which guides the dynamic evolution of the topology computing layer by comparing the error between the predicted signals output by the internal generative model and the actual sensory signals input by the sensor-action layer; wherein the topology computing core realizes picosecond-level topology connection reorganization through a silicon photonic waveguide reconfigurable interconnection network between memory-computing micro-columns.

2. The somatic computing integrated computing architecture based on dynamic resonance topology processing according to claim 1, wherein, The topology computing layer adopts a hierarchical and heterogeneous dynamic network structure, comprising: an instinctive reflex sub-layer composed of hard-wired digital / analog circuits, for processing high real-time and high reliability reflex behaviors that require millisecond-level response, including obstacle avoidance; The topology computing core is provided with a relatively fixed or slowly plastic memristor cross array inside each memory-computing micro-column, which is used to realize weight storage, state memory and vector matrix multiplication calculation, and constitutes the basic functional unit of the architecture.

3. The somatic computing integrated computing architecture based on dynamic resonance topology processing according to claim 1, wherein, The global modulation system comprises: a rhythm generator for listening to the neural oscillation patterns spontaneously emerging in the topology computing layer, detecting and amplifying the neural oscillation patterns to form a global synchronous rhythm, and locking and stabilizing specific node cluster resonant groups through the global synchronous rhythm; a value regulator for receiving prediction error signals, quantizing the amplitudes of the prediction error signals by multiple thresholds, and releasing modulation signals of different intensities and ranges according to the quantization results to regulate the plasticity of the topology computing layer in stages.

4. The somatic computing integrated computing architecture based on dynamic resonance topology processing according to claim 1, wherein, The architecture retrieves information through a content addressing mechanism: it broadcasts partial clue information to the topology computing layer, and activates the complete node cluster topology related to the clue information in the topology computing layer in parallel through a sparse diffusion activation mechanism.

5. The somatic computing integrated computing architecture based on dynamic resonance topology processing according to claim 1, wherein, The architecture supports an offline consolidation mechanism: during the idle period of the system without external task processing or the period simulating biological sleep, the topology activity patterns of important tasks previously marked are spontaneously replayed to strengthen the connections of the node clusters corresponding to the important tasks, preventing catastrophic forgetting in the lifelong learning process.

6. A somatic computing integrated computing method based on dynamic resonance topology processing implemented on the architecture of any one of claims 1-5. The steps comprise: Step 1, the sensor-action layer perceives the event signals of the physical environment and generates sparse pulse signals to input the topology computing layer, activating the relevant node micro-column clusters of the topology computing layer; at the same time, the internal generative model in the architecture predicts the next step of sensory input and the action to be executed based on the activation state of the current topology computing layer, forming an internal prediction; Step 2, after detecting the synchronous oscillation prototype of the node cluster related to the current task, the rhythm generator of the global modulation system generates and broadcasts the same frequency global rhythm, forming a stable node cluster resonance group in the topology calculation layer; Step 3, the mode of the node cluster resonance group is translated into specific actuator driving commands through the instinctive reflex sub-layer or the motion interface, generating external actions; Step 4, the sensor-action layer receives the feedback signal of the physical environment to the external action, compares the feedback signal with the internal prediction in step 1, and calculates the prediction error; Step 5, the value regulator of the global modulation system adjusts the connection weight of the topology calculation layer or the physical connection structure between the memory-computing integrated microcolumns according to the amplitude of the prediction error, completing a learning cycle; wherein small errors only fine-tune the synaptic weights within the resonance group, large errors guide the recombination of the physical connection between the microcolumns, and large errors with punishment drive large-scale topology reconstruction.

Images