A brain-inspired adaptive learning method and system based on the human brain's learning mechanism

Abstract

This invention discloses a skill adaptive learning method and system based on the subconscious learning mechanism of the human brain, belonging to the fields of artificial intelligence and brain-like computing. This invention aims to solve the problems in existing technologies where skill learning struggles to form long-term, persistent memory and lacks offline consolidation and a closed loop of forgetting and relearning. Key technical points include: acquiring skill data and extracting features; during system idle periods, using prime numbers to divide the backtracking time interval, counting the number of skill learning sessions within each interval, and accumulating cutoff values ​​to calculate the subconscious memory threshold; when the threshold is greater than or equal to 1, moving the skill into the subconscious permanent memory area; within each idle period, performing memory deepening and forgetting calculation operations on the skills in the permanent memory area based on a Sigmoid curve with increasing step size depending on the idle time; when the skill strength falls below the exit threshold, automatically triggering the relearning closed loop. This invention can be deployed on edge computing devices or brain-like chips, significantly improving the long-term retention rate of skills and achieving adaptive allocation of computing resources.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and brain-like computing technology, specifically to a skill adaptive learning method and system that simulates the subconscious learning and memory consolidation mechanism of the human brain, and is particularly suitable for intelligent agents or edge computing devices that require long-term skill maintenance, offline consolidation, and adaptive forgetting and relearning. Background Technology

[0002] Traditional machine learning methods (such as deep neural networks and reinforcement learning) typically employ online batch training, lacking simulation of key mechanisms in the human brain such as "sleep consolidation," "subconscious memory areas," and the "forgetting and relearning loop." The main technical problems are as follows: 1. Catastrophic forgetting: After learning a new skill, the performance of the old skill declines sharply, making it impossible to achieve long-term and persistent memory of the skill; 2. Lack of offline reinforcement: The model cannot automatically reinforce learned knowledge during learning gaps (idle periods), resulting in low learning efficiency; 3. Vague memory threshold: The conversion conditions between working memory and long-term memory are not distinguished, and all skills are stored and updated using a uniform strategy, resulting in a waste of computing resources; 4. Inability to automatically relearn after forgetting: When skills deteriorate, traditional methods require manual restart of training, lacking closed-loop regulation capabilities. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention provides a skill adaptive learning method and system based on the subconscious learning mechanism of the human brain. By constructing a four-stage architecture of "skill feature extraction → subconscious threshold calculation → persistent memory storage → offline forgetting / deepening closed loop", it achieves efficient long-term skill retention and adaptive allocation of computing resources.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: An adaptive skill learning method based on the subconscious learning mechanism of the human brain includes the following steps: Step 1: Obtain the skill data to be learned as input data, and extract skill features from the input data by simulating the perception and encoding method when the human brain learns skills, and generate a preliminary skill feature expression; Step 2: Based on the memory consolidation mechanism of the human brain's subconscious layer, the preliminary skill feature expression is input into a memory depth enhancement processing area; according to the preset human brain working rhythm model, the skill learning feature table is iteratively calculated during the system's idle period to dynamically determine the subconscious memory threshold required for the current skill to enter subconscious memory; Step 3: Determine whether the consolidation strength of the preliminary skill feature expression reaches or exceeds the subconscious memory threshold. If so, move the corresponding skill feature information into the subconscious memory area and mark it as persistent memory. Step 4: Simulate the forgetting and memory enhancement mechanisms of the human brain's subconscious layer. Following the described human brain work rhythm model, perform batch calculations on the various skill learning characteristic tables accumulated that day during each off-peak work period: For skills already located in the subconscious memory area, memory strengthening or forgetting decay operations are performed according to their frequency of use and the most recent time of use. When any skill feature is removed from the subconscious memory area due to forgetting and decay, the skill feature is reintegrated into the subconscious learning algorithm formula described in step 2, and the subconscious learning process of steps 1 to 3 is executed again.

[0005] In step 1, the perception and encoding method that simulates the human brain learning skills specifically includes: using spatiotemporal pulse coding or sparse autoencoder networks to map skill data into a time-dependent neuronal pulse sequence.

[0006] The subconscious memory threshold θsub is calculated according to the following formula: Where θsub is the subconscious memory threshold, p iterates through all prime numbers less than 30, and a(p) represents the number of skill learning attempts associated with the prime number of days. This formula simulates the sparse encoding characteristics of spatiotemporal patterns by the hippocampal position cell group in the human brain. Each prime number corresponds to a specific encoding unit, and the number of learning attempts is truncated to an upper limit of 3 to simulate the saturation effect of synaptic consolidation. When θsub>=1, the algorithm migrates into the subconscious memory area.

[0007] The skill learning feature table includes at least: skill identifier, feature vector, current memory strength value, recent learning timestamp, and forgetting rate coefficient.

[0008] The simulated mechanism of forgetting and memory enhancement in the subconscious layer of the human brain: Among them, Mmax is the maximum memory strength (usually set to 1), Tidle is the system idle duration, T0 is the semi-activated idle duration benchmark, and k is the sensitivity coefficient; δ represents the proportion of the memory deepening step length that can be obtained in the current idle period, and its value range is 0 to 1; when the idle duration is extremely short, δ→0 and there is no deepening; when the idle duration is long enough, δ→1 and the memory strength approaches the maximum value; when Mmax < Mold and the calculated Mnew < 1, move out of the subconscious area and perform re-learning of a certain memory skill.

[0009] The invention also provides a skill adaptive learning system based on the human brain subconscious learning mechanism, including: a skill feature extraction module, which is used to obtain input data and perform skill feature extraction on the input data by simulating the way of the human brain learning skills, and generate a preliminary skill feature expression; a subconscious threshold calculation module, connected to the skill feature extraction module, which is used to map the preliminary skill feature expression to the memory depth enhancement processing area based on the memory consolidation mechanism of simulating the human brain subconscious, and calculate the skill learning feature table during the system idle period, and dynamically output the subconscious memory threshold; a persistent memory determination module, connected to the subconscious threshold calculation module, which is used to judge whether the consolidation strength of the skill feature expression reaches or exceeds the subconscious memory threshold, and store the qualified skill feature information in the subconscious memory area; a forgetting and deepening closed-loop adjustment module, respectively connected to the persistent memory determination module and the subconscious threshold calculation module, which is used to batch calculate the daily skill learning feature table in each idle cycle, perform deepening or forgetting attenuation on the skill features in the subconscious memory area, and when the skill features exit the subconscious memory area, trigger the skill feature extraction module and the subconscious threshold calculation module to perform re-learning on it.

[0010] Compared with the prior art, the present invention has the following beneficial effects: 1. Simulating the subconscious consolidation mechanism: By calculating the "subconscious memory threshold" offline during the idle period and performing deepening / forgetting, the long-term retention rate of skills is significantly improved, and catastrophic forgetting is avoided. 2. Efficient use of computing resources: Only the information that reaches the threshold enters the persistent storage area, reducing invalid storage and calculation; batch processing during the idle period does not interfere with online tasks. 3. Closed-loop adaptive re-learning: Degraded skills will be automatically re-integrated into the learning process, realizing the closed-loop adjustment of "forgetting → re-learning → re-consolidation" without manual intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flowchart of a skill adaptive learning method based on the human brain subconscious learning mechanism provided by an embodiment of the present invention;

[0012] Figure 2This is a structural diagram of a skill adaptive learning system based on the subconscious learning mechanism of the human brain, provided in an embodiment of the present invention.

[0013] Figure 3 This is a Sigmoid curve diagram showing the increase in skill memory intensity over time during idle periods, as provided in this embodiment of the invention.

[0014] Figure 4 This is a schematic diagram of the time interval division of the spiral memory elements of the human brain provided in an embodiment of the present invention.

Claims

1. A skill adaptive learning method based on the subconscious learning mechanism of the human brain, characterized in that, The process includes the following steps: Step 1: Preliminary skill feature extraction. The skill data to be learned is obtained as input data. By simulating the perception and encoding method when the human brain learns skills, the input data is used to extract skill features and generate a preliminary skill feature expression. Step 2: Subconscious Memory Consolidation and Threshold Calculation. Based on the memory consolidation mechanism of the human brain's subconscious layer, the preliminary skill feature expression is input into a memory depth enhancement processing area; according to the preset human brain working rhythm model, the skill learning feature table is iteratively calculated during the system's idle period to dynamically determine the subconscious memory threshold required for the current skill to enter subconscious memory; Step 3: Persistent memory storage determines whether the consolidation strength of the preliminary skill feature expression reaches or exceeds the subconscious memory threshold. If so, the corresponding skill feature information is moved into the subconscious memory area and marked as persistent memory. Step 4: Closed-loop regulation of subconscious forgetting and memory enhancement. Simulating the forgetting and memory enhancement mechanisms of the subconscious layer of the human brain, according to the aforementioned human brain work rhythm model, batch calculations are performed on the various skill learning feature tables accumulated that day during each work idle period: 4.

1. For skills and characteristics already located in the subconscious memory area, perform memory strengthening or forgetting decay operations according to their frequency of use and the most recent time of use; 4.

2. When any skill feature is removed from the subconscious memory area due to forgetting and decay, the skill feature is reintroduced into the subconscious learning algorithm formula described in step 2, and the subconscious learning process of steps 1 to 3 is executed again.

2. The method according to claim 1, wherein the subconscious memory threshold in step 2 is: in, θsub is the subconscious memory threshold, p iterates through all prime numbers less than 30, and a(p) represents the number of skill learning attempts associated with the prime number of days. This formula simulates the sparse encoding characteristics of spatiotemporal patterns by the hippocampal position cell group in the human brain. Each prime number corresponds to a specific encoding unit, and the number of learning attempts is truncated to an upper limit of 3 to simulate the saturation effect of synaptic consolidation. When θsub>=1, the algorithm migrates into the subconscious memory area.

3. The method according to claim 1, characterized in that, Step 3 describes the simulation of the forgetting and memory enhancement mechanisms in the subconscious layer of the human brain, including: simulating the physiological characteristic that the human brain can only trigger effective memory consolidation after reaching a certain idle time (i.e., slow-wave sleep time), and that the consolidation gain gradually saturates with increasing idle time; and using the following formula to calculate the total number of days that the updated memory items meet the prime number criteria: in, δ This represents the consolidation gain coefficient during the current idle period. M `max` represents the total number of prime days that a certain memory item meets the requirement (including the current idle value). M `old` represents the total number of prime days originally present in this memory item; when `Mmax <` M old (calculated) M When new < 1, the process involves removing a memory from the subconscious area and relearning that skill.

4. The method according to claim 1, characterized in that, The subconscious memory threshold is dynamically calculated from the following parameters: the number of skill repetitions, the strength of skill association, the cumulative duration of system idle periods, and the modulation factor that simulates human brain sleep time.

5. The method according to claim 1, characterized in that, The skill learning feature table includes at least: skill identifier, feature vector, current memory strength value, recent learning timestamp, and forgetting rate coefficient.

Images