Computational methods based on artificial intelligence algorithms

Abstract

This invention relates to the fields of computer architecture and artificial intelligence technology, and discloses a computational method based on artificial intelligence algorithms, comprising: obtaining the feature tensor of the first computational level in a cascaded computational model and determining its channel activation distribution vector; calculating the information entropy value of the channel activation distribution vector and the corresponding feature dimension index value; writing the feature dimension index value into a storage area to form an index sequence of consecutive computational levels; calculating the entropy reduction gradient between levels and determining the decision stability based on the highest occurrence frequency of the same index value in the index sequence; stopping the subsequent computation steps of the cascaded computational model when the entropy reduction gradient is lower than a threshold and the model is in a stable state. This invention achieves synchronous perception of feature evolution momentum and decision path stability, effectively identifies the prediction pointing to an oscillating state and blocks computational waste, thereby improving the efficiency of computational resource allocation while ensuring the model's prediction accuracy.

Description

Technical Field

[0001] This invention relates to a computational method based on artificial intelligence algorithms, belonging to the fields of computer architecture and artificial intelligence technology. Background Technology

[0002] Current cascaded AI models represent the features of complex data through multi-level nonlinear mapping and extract semantic information from intermediate tensors through deeply stacked computational layers. This has become the mainstream architecture in the field of computer systems based on specific computational models. However, in edge computing nodes with limited computing power or high-concurrency cloud inference scenarios, the complexity of input data exhibits a non-uniform distribution. Using static computational graphs to process differentiated data leads to simple samples with clear features penetrating the complete deep network structure, generating ineffective floating-point operation overhead. At the same time, for some edge samples at the decision boundary or out-of-domain distributed data, intermediate features enter a semantic evolution plateau during the flow process. The model prediction points to frequently flip between multiple highly probable category indices, causing scalar determinism to fail to reach the preset shutdown threshold. This results in the system processing the entire computation process but failing to obtain substantial decision gains, forming a zero-marginal-return computational trap.

[0003] To address the challenge of asymmetrical resource allocation, existing technologies employ confidence truncation mechanisms based on single numerical comparisons or static thresholds that decay with network depth. Analysis reveals the following main shortcomings of these existing solutions: 1. The singularity of scalar metrics prevents the system from capturing the evolutionary momentum of features near the classification hyperplane, neglecting the stagnation of marginal utility during feature extraction; 2. The oscillatory properties of discrete decision-making are systematically ignored, causing samples at classification saddle points to induce invalid feature reconstruction in deep networks; 3. A mismatch exists between static shutdown logic and dynamic data distribution, making it impossible to approach the theoretical lower limit of computational power consumption while maintaining the stability of the classification boundary. Due to a lack of awareness of the feature evolution trajectory properties in cascaded deduction, existing solutions cannot identify the oscillations generated by the computational model at decision saddle points, leading to early misjudgments of simple samples due to over-prediction, or causing ill-conditioned samples to continuously consume redundant computational power during the saturation period, making it difficult to fundamentally resolve the performance paradox between computational depth and feature quality.

[0004] Therefore, how to simultaneously perceive the decay of momentum in the evolution of intermediate features and the stability of discrete decision-making topological trajectories, so as to block invalid forward inference in the cascaded computation model and improve the dynamic resource scheduling efficiency of the system in processing data with full complexity distribution, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A calculation method based on artificial intelligence algorithms, comprising the following steps: Step S101: Obtain the feature tensor of the i-th computational level output for the sample to be processed in the cascaded computation model, and perform channel dimension mean pooling on the feature tensor to obtain the channel activation distribution vector. Step S102: Calculate the information entropy value of the channel activation distribution vector and determine the feature dimension index value corresponding to the element with the largest value in the channel activation distribution vector; Step S103: Write the feature dimension index value into the index storage area. The index storage area records the feature dimension index sequence of K consecutive computation levels, including the i-th computation level, where K is an integer greater than 1. Step S104: Obtain the information entropy value of the (i-1)th calculation level, and calculate the entropy reduction gradient between the information entropy value and the information entropy value of the (i-1)th calculation level; Step S105: Calculate the highest frequency of occurrence of the same index value in the feature dimension index sequence, and determine the stability state of the cascaded computing model at the i-th computing level based on the highest frequency of occurrence. Step S106: When the entropy reduction gradient is lower than the preset gradient threshold and the stability state is stable, stop the operation logic of the cascaded calculation model at the (i+1)th calculation level and subsequent calculation levels, and output the classification judgment data generated at the i-th calculation level.

[0006] Preferably, step S105, which determines the stability state based on the highest frequency of occurrence, includes the following steps: step S201, comparing the highest frequency of occurrence with a preset frequency determination threshold; step S202, if the highest frequency of occurrence is not lower than the frequency determination threshold, then the stability state is determined to be a stable state; step S203, if the highest frequency of occurrence is lower than the frequency determination threshold, then the stability state is determined to be a prediction flip state, and a feature reconstruction instruction is generated to drive the cascaded computing model to perform feature extraction at the (i+1)th computing level.

[0007] Preferably, after outputting the classification judgment data, the method further includes: obtaining the current inference depth value of the cascaded computing model; calculating the correction coefficient for the preset gradient threshold based on the current inference depth value; and adjusting the judgment benchmark of subsequent input samples in step S106 using the correction coefficient so that the judgment benchmark decreases as the inference depth increases.

[0008] Preferably, the information entropy value of the channel activation distribution vector in step S102 is calculated using the following method: ,in, Let C be the total number of feature channels, and calculate the information entropy value for the i-th level. This represents the proportion of the activation value of the j-th channel in the channel activation distribution vector to the total activation value.

[0009] Preferably, step S101 performs channel dimension mean pooling on the feature tensor, including: calculating the average value of all spatial location elements in each feature channel of the feature tensor; arranging the obtained average values ​​of each feature channel in the order of the original channel index to construct a channel activation distribution vector.

[0010] Preferably, after determining the stability state as the predicted flip state, the method further includes: a statistical cascade calculation model in continuous... The rate of change of the highest frequency of occurrence within each computational level, where M is an integer between 2 and 5; if the highest frequency of occurrence does not increase within M consecutive computational levels, and the information entropy value is not lower than the preset saturation upper limit threshold, then a computation termination instruction is output, and the remaining computation steps corresponding to that sample in the cascaded computation model are skipped.

[0011] Preferably, in step S103, the storage depth K of the index storage area is set to an integer between 3 and 7, and the index storage area updates the feature dimension index sequence according to the first-in-first-out logic.

[0012] Preferably, the output of classification determination data in step S106 includes: determining the feature dimension index value corresponding to the highest activation value output by the i-th calculation level as the predicted category label; encapsulating the predicted category label and the confidence score calculated by the i-th calculation level to generate an inference result data package.

[0013] Preferably, the cascaded computing model consists of multiple cascaded residual blocks or attention mechanism layers, each of which integrates a logic control node for executing steps S101 to S106.

[0014] Preferably, step S106, which stops the operation logic of the (i+1)th computation level and subsequent computation levels, includes: generating a skip signal for all computation levels after the i-th computation level, so as to release the processor's computation thread and route it to the next set of data streams to be processed.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In the computational methods of artificial intelligence algorithms, a composite decision mechanism consisting of entropy reduction gradient and dimension index transition rate between computational levels is used to identify the plateau period and decision oscillation state of feature evolution, thereby blocking the forward inference process without gain. When processing out-of-domain distributed data or marginal samples, if the feature experiences high-frequency flipping of dimension index near the classification hyperplane, the mechanism determines that the current feature has entered a non-convergent oscillation trap and forcibly triggers a truncation action, avoiding the system from spending huge floating-point computing resources on invalid prediction corrections, and establishing an objective mapping relationship between the allocation of computing resources and the intrinsic complexity of the samples.

[0016] 2. By introducing discrete-dimensional decision trajectory monitoring into the uncertainty measure of the scalar dimension, the control stability of the cascaded computing model under dynamic conditions is improved. An index state register with a window depth of K is used to record the transition path of the predicted category and is logically coupled with a dynamic threshold that decays with depth. This enables the system to effectively distinguish between hard-to-classify samples in normal evolution and simple samples in a decision-locked state. Under the premise of ensuring classification accuracy, the hidden redundant computational overhead in the deep network is eliminated.

[0017] 3. The constructed two-dimensional topology awareness mechanism relies only on basic operators such as spatial global average pooling and maximum value index extraction. It has high engineering feasibility in edge computing nodes with limited computing power. Compared with the scheme of building auxiliary indexes by adding extra storage space, the logic flow scheduling is performed by using the intermediate feature tensors generated by the model. The extra computing power consumption generated by its control logic has an asymmetric advantage of two to three orders of magnitude over the overhead saved by deep network inference, which alleviates the power consumption bottleneck in the deployment of large-scale computing models. Attached Figure Description

[0018] Figure 1 This is a flowchart of the steady-state determination and cascaded calculation process of the artificial intelligence algorithm of this invention; Figure 2 This is a diagram of the computing power dynamic truncation and logical scheduling architecture of the artificial intelligence algorithm of this invention.

[0019] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0021] A computational method based on artificial intelligence algorithms includes the following steps: Step S101: Obtain the feature tensor of the i-th computational level output for the sample to be processed in the cascaded computation model, and perform channel dimension mean pooling on the feature tensor to obtain the channel activation distribution vector. Step S102: Calculate the information entropy value of the channel activation distribution vector and determine the feature dimension index value corresponding to the element with the largest value in the channel activation distribution vector; Step S103: Write the feature dimension index value into the index storage area. The index storage area records the first... A sequence of feature dimension indices for K consecutive computational levels, including the computational level, where K is an integer greater than 1; Step S104: Obtain the information entropy value of the (i-1)th calculation level, and calculate the entropy reduction gradient between the information entropy value and the information entropy value of the (i-1)th calculation level; Step S105: Calculate the highest frequency of occurrence of the same index value in the feature dimension index sequence, and determine the stability state of the cascaded computing model at the i-th computing level based on the highest frequency of occurrence. Step S106: When the entropy reduction gradient is lower than the preset gradient threshold and the stability state is stable, stop the operation logic of the cascaded calculation model at the (i+1)th calculation level and subsequent calculation levels, and output the classification judgment data generated at the i-th calculation level.

[0022] Preferably, step S105, which determines the stability state based on the highest frequency of occurrence, includes the following steps: step S201, comparing the highest frequency of occurrence with a preset frequency determination threshold; step S202, if the highest frequency of occurrence is not lower than the frequency determination threshold, then the stability state is determined to be a stable state; step S203, if the highest frequency of occurrence is lower than the frequency determination threshold, then the stability state is determined to be a prediction flip state, and a feature reconstruction instruction is generated to drive the cascaded computing model to perform feature extraction at the (i+1)th computing level.

[0023] Preferably, after outputting the classification judgment data, the method further includes: obtaining the current inference depth value of the cascaded computing model; calculating the correction coefficient for the preset gradient threshold based on the current inference depth value; and adjusting the judgment benchmark of subsequent input samples in step S106 using the correction coefficient so that the judgment benchmark decreases as the inference depth increases.

[0024] Preferably, the information entropy value of the channel activation distribution vector in step S102 is calculated using the following method: ,in, Let C be the total number of feature channels, and calculate the information entropy value for the i-th level. This represents the proportion of the activation value of the j-th channel in the channel activation distribution vector to the total activation value.

[0025] Preferably, step S101 performs channel dimension mean pooling on the feature tensor, including: calculating the average value of all spatial location elements in each feature channel of the feature tensor; arranging the obtained average values ​​of each feature channel in the order of the original channel index to construct a channel activation distribution vector.

[0026] Preferably, after determining the stability state as the predicted flip state, the method further includes: calculating the rate of change of the highest frequency of occurrence in the cascaded calculation model within M consecutive calculation levels, where M is an integer between 2 and 5; if the highest frequency of occurrence does not increase within M consecutive calculation levels, and the information entropy value is not lower than the preset saturation upper limit threshold, then a calculation termination instruction is output, and the remaining calculation steps corresponding to the sample in the cascaded calculation model are skipped.

[0027] Preferably, in step S103, the storage depth K of the index storage area is set to an integer between 3 and 7, and the index storage area updates the feature dimension index sequence according to the first-in-first-out logic.

[0028] Preferably, the output of classification determination data in step S106 includes: determining the feature dimension index value corresponding to the highest activation value output by the i-th calculation level as the predicted category label; encapsulating the predicted category label and the confidence score calculated by the i-th calculation level to generate an inference result data package.

[0029] Preferably, the cascaded computing model consists of multiple cascaded residual blocks or attention mechanism layers, each of which integrates a logic control node for executing steps S101 to S106.

[0030] Preferably, step S106, which stops the operation logic of the (i+1)th computation level and subsequent computation levels, includes: generating a skip signal for all computation levels after the i-th computation level, so as to release the processor's computation thread and route it to the next set of data streams to be processed.

[0031] Example 1: In a high-concurrency video stream defect detection gateway scenario deployed at the edge of a production line, the system integrates a cascaded computing model for performing multi-class classification on real-time acquired workstation images. Regarding the attention mechanism layer in the cascaded computing model, the specific generation method of the channel activation distribution vector is as follows: global mean compression is performed on the embedding vector dimension containing 512 components. The resulting 512 sets of means are arranged according to the original vector index order to construct a 512-dimensional channel activation distribution vector reflecting the current feature energy distribution. Since the vast majority of samples during production line operation are well-defined good products, and such samples generate deterministic semantic expressions in the shallow computation steps of the cascaded computing model, if further... Performing full-depth forward propagation computation will result in ineffective floating-point operations at edge computing nodes and induce frequency drops due to processor heat accumulation, leading to significant fluctuations in end-to-end latency of the detection system. To address this technical challenge of asymmetrical computing power allocation, this method reconstructs the topological flow path of the computation graph through the collaborative perception of feature evolution momentum and decision trajectory stability. Specifically, the edge processor obtains the feature tensor output from the i-th computational level in the cascaded computation model, performs channel-dimensional mean pooling on this feature tensor, calculates the average value of all spatial elements within each feature channel in the feature tensor to obtain the channel activation distribution vector, and calculates the information entropy value of this level according to the following formula. : ,in, Let C be the total number of feature channels, and calculate the information entropy value for the i-th level. The proportion of the activation value of the j-th channel in the channel activation distribution vector to the total activation values; simultaneously extract the feature dimension index value corresponding to the element with the largest value in the channel activation distribution vector. It is then written to the index storage area according to the first-in-first-out logic to update the feature dimension index sequence of K consecutive computation levels, where K is set to 5.

[0032] During the feature extraction process at each level, the system backtracks to obtain the information entropy value stored at the (i-1)th computational level. And calculate the entropy reduction gradient of the current level. Simultaneously, the system calculates the highest frequency of identical index values ​​in the feature dimension index sequence and compares it with a preset frequency threshold to determine stability. When the system processes typical, well-defined, high-quality samples, the computational level... As the number of features increases, the channel activation distribution vector exhibits a unimodal characteristic, causing the calculated entropy reduction gradient ΔH to decay and fall below the preset gradient threshold. Furthermore, the feature dimension index values ​​in the index storage area remain locked at a fixed category number representing good products, resulting in the highest frequency not falling below the frequency judgment threshold. At this point, the system determines the stability state as stable, directly triggering an early termination strategy, stopping the operation logic of all subsequent computational levels, and outputting the classification judgment data for the current level. For edge samples at the decision boundary or out-of-domain distribution data containing abnormal noise, the feature dimension index sequence in the index storage area exhibits high-frequency fluctuations due to frequent flipping of features between multiple category indices. The highest frequency remains below the frequency judgment threshold, and the system thus determines the stability state as a prediction flipping state. At this point, the system initiates discrete monitoring of the feature space trajectory, statistically analyzing the rate of change of the highest frequency within M consecutive computational levels, where M is set to 3. If the highest frequency is not detected to increase within consecutive levels, and the information entropy value... If the value remains above the preset saturation upper limit, it proves that the current feature has fallen into the decision saddle point of the classification hyperplane and lost its feature convergence ability.

[0033] The system outputs a computation termination command, trunculating the subsequent forward inference of the cascaded model and fundamentally blocking the calculation of zero marginal benefit for ill-conditioned samples, when the highest frequency occurs. Below the frequency threshold Furthermore, when the stability state is the predicted flip state, the feature reconstruction instruction modifies the expansion rate of the convolution kernel at the (i+1)th computational level to adjust the receptive field coverage, extracts the oscillation frequency within the index sequence of the feature dimension in the index storage area, increases the activation coefficient of the attention mechanism layer at the (i+1)th computational level when two specific category indices are detected to alternate within M consecutive levels, tilts the model's computational resources toward highly significant local feature regions and breaks the degeneracy of features near the classification boundary, shifts the feature tensor toward the target category steady-state attractor, and completes edge sample depth feature enhancement before outputting the classification decision data at the (i)th computational level, where M is a preset constant for the decision oscillation period. Calculate the highest frequency of occurrence for level i. Using the frequency determination threshold, through the deep collaboration of the aforementioned two-dimensional topology perception mechanism and dynamic routing strategy, this edge-side defect detection gateway achieves a causal mapping between computing resources and intrinsic data complexity. The average inference depth of good samples is reduced from the original 50 levels to 12 levels, and the effective computing power utilization of the processor under full load is increased by more than 3 times. Furthermore, by intercepting decision oscillation states, it avoids early misjudgment interference introduced by local abnormal activation, ensuring the overall balance of system processing performance and response determinism in a high-concurrency inference flow environment.

[0034] Example 2: In a visual simulation environment simulating an industrial automated quality inspection process, an edge server with 15 TFLOPS floating-point computing power and 32GB of video memory was used to run a cascaded convolutional neural network containing 50 residual blocks. The experimental dataset consisted of 10,000 sets of real-world production line images from an industrial defect image library. To simulate electromagnetic interference in the industrial environment, Gaussian white noise with a signal-to-noise ratio of 20dB was superimposed on the original input signal, and a 50Hz power frequency interference harmonic was introduced. The value of the number of continuous calculation levels K for the key parameter was chosen to balance the system's tolerance to noise fluctuations with its sensitivity to capturing the steady state of the decision. A value too small for K could lead to errors induced by instantaneous disturbances. Truncation risks exist, but excessively large values ​​of K increase the storage load on the state register and delay the truncation timing. Gradient step tests within the range of 2 to 10 were performed, and it was observed that when K was set to 5, the system exhibited definite discriminative steady-state characteristics in sample streams containing the aforementioned noise interference. Therefore, K was set to 5 in this experiment. This experiment was divided into the present invention sample group and the control sample group. The control sample group adopted a fixed-path deduction mode. The present invention sample group activated dynamic path control logic based on information entropy and feature dimension index sequences. During the experimental run, a set of original images containing typical scratch defects was used as input. When the present invention sample group reached the 12th calculation level, the channel activation distribution vector exhibited a unimodal distribution. At this time, the information entropy value of this level... The value is 0.28, and the entropy reduction gradient ΔH of the 11th level is 0.08. This value is lower than the preset gradient threshold of 0.10. The feature dimension index sequence of five consecutive levels recorded in the index storage area is retrieved simultaneously. The sequence is {3,3,3,3,3}, where the highest frequency of the same index value 3 reaches 100%. This proportion exceeds the preset frequency judgment threshold of 80%. Based on the above data, the system determines that the current cascaded calculation model is in a stable state and directly triggers the early termination strategy, stopping the operation logic of the subsequent 38 levels. The final output classification judgment data identifies the scratch defect category, and the cumulative floating-point operation consumed is 1.42 GFLOPs. In contrast, the comparison sample group needs to perform a complete 50-level inference to complete the same task, and the computing power consumption reaches 6.15 GFLOPs. The data shows that the sample group of the present invention achieves a 76.9% reduction in computing power while ensuring classification accuracy.

[0035] To determine the effective working window for the frequency judgment threshold, a performance inflection point test was introduced using edge samples at the decision boundary. When the frequency judgment threshold was set below 60%, the system exhibited false truncation for samples containing random noise, with the false alarm rate jumping from 0.4% to 3.8%. This indicates that an excessively low threshold caused random jumps in feature dimensions to be identified as stable states. When the frequency judgment threshold was increased to above 95%, the average inference depth increased from 14 layers to 42 layers, exhibiting diminishing marginal returns on computational costs, and the improvement in classification accuracy was less than 0.1%. When processing ill-conditioned samples that predicted high-frequency flips, the sample group using the complete scheme of this invention could determine the predicted flip state by identifying the state with the highest occurrence frequency below the threshold, driving the model to continue reconstructing features deeper until the information entropy value was reached. The convergence requirement was met, confirming the logical synergy between entropy reduction gradient and index frequency. Specifically, the entropy reduction gradient provides a convergence basis for calibrating energy distribution, while the index frequency eliminates spurious regressions through path consistency. This experiment objectively recorded the adaptive response process of the proposed method under different input complexities. When processing industrial data streams containing 20dB background noise, the average inference depth of the proposed sample group remained at 12.4 layers, and the system throughput increased from 45 frames per second to 138 frames per second in the comparison sample group. Furthermore, the consistency between the core detection indicators and the full-depth inference results reached 99.2%. The experimental data confirms that the parameter range defined in this invention covers the balance between improving computational efficiency and maintaining prediction accuracy. By injecting a logical adjudication mechanism into the computation nodes of the cascaded computation model, excessive computation for deterministic features is blocked, providing a definite engineering implementation path for the deployment of artificial intelligence algorithms in resource-constrained environments.

[0036] Example 3: In a computing power allocation application scenario deployed on a high-bandwidth distributed inference cluster, the cluster is used to support a large-scale pre-trained cascaded computing model. Because the processed data stream contains noisy samples at the distribution edge, the system faces the challenge of decision oscillations caused by non-convergence of feature representations. To ensure that the triggering of the computation termination instruction has a definite physical basis, this example provides a calibration procedure for preset gradient thresholds and frequency determination thresholds, along with corresponding underlying storage scheduling methods. In the initial deployment phase of the cascaded computing model, the system executes an initial state definition procedure, selecting a historical sample set containing 5000 groups with typical distribution characteristics as calibration input, and configuring a processor with vector operation acceleration capabilities as the enabling environment. During the calibration phase, the processor performs forward inference for each preset early termination node in the cascaded computing model, synchronously collecting data from each computing level. Output information entropy value and feature dimension index value The startup process involves determining the quantization procedure by calculating the probability density distribution of the entropy reduction gradient ΔH of all good samples in the historical sample set at the i-th calculation level. The value at the point where the cumulative probability reaches 95% is extracted as the initial reference value for the preset gradient threshold. Regarding the determination of the frequency threshold, the system statistically analyzes the offset of sample features within K consecutive calculation levels, calculates the expected value for maintaining consistency of the same index value in the feature dimension index sequence, and sets 1.2 times this expected value as the frequency threshold. This serves as the benchmark value for inhibiting spurious steady-state detection and triggering the calculation termination command. The preset gradient threshold Γ and the frequency threshold are then determined. Based on the N sets of labeled samples in the validation set, the entropy reduction gradient ΔH and the highest frequency of occurrence of the output feature tensor at each computational level i are obtained from the cascaded computational model. The cumulative probability distribution of ΔH of the validation set at each level is calculated, and the 95th percentile value is extracted as the initial baseline for level Γ. The correct classification samples are statistically analyzed within K consecutive calculation levels. The observed sequence was set as follows: the value obtained by subtracting twice the standard deviation from the sequence mean was used as the standard deviation. Calculate the theoretical entropy value under a fully random distribution based on the total number of feature channels C. , with 0.85 The product of these factors serves as the upper saturation threshold for determining whether the feature evolution has entered a stagnant phase. .

[0037] To achieve low-latency maintenance of the feature dimension index sequence in the index storage area, this method defines a circular buffer as the index storage area, whose addressing logic follows a first-in-first-out (FIFO) rule. When the processor generates the current computation level... Feature dimension index value At that time, the write pointer of the buffer is updated by an incrementing counter with a step size of 1, and the write pointer of the buffer is updated. Write to the physical address pointed to by the pointer, synchronously overwriting the old data at the iKth computation level. The vote counting unit performs parallel comparisons on the K elements in the circular buffer and calculates the results with the current index value. The number of identical elements is counted, and the highest frequency of occurrence is calculated using the following formula. : ,in, Calculate the highest frequency of occurrence for level i. For the index storage area and The total number of elements with equal values, where K is the length of the feature dimension index sequence. If the calculated... If the frequency threshold is below the threshold and no positive shift occurs for M consecutive levels, the system determines the stability state as a prediction flip state and sends an interrupt request to the processor, suspending the current inference thread. To address the issue of computing performance drifting with device aging, this system incorporates a dynamic evolution procedure for the judgment benchmark. By obtaining the current inference depth D of the cascaded computing model, a correction coefficient σ is calculated for the preset gradient threshold. The specific calculation method is as follows: σ = exp(-D / N), where σ is the correction coefficient, D is the current inference depth of the cascaded computing model, and N is the total number of computing layers in the cascaded computing model. By multiplying the preset gradient threshold with the correction coefficient σ, the judgment benchmark monotonically decreases with increasing inference depth. Therefore, when the model infers to deeper spaces, the system can use a more stringent convergence criterion to offset feature drift caused by noise accumulation, locking the output classification judgment data within the steady-state attractor of the feature space. This resolves the contradiction between computational real-time performance and classification reliability, reducing the system response latency jitter rate to below 5% under full load operation.

[0038] Example 4: In an application scenario where a cascaded computing model is deployed on a heterogeneous hardware platform, to address the numerical truncation error caused by processor quantization operations, the system determines a physical benchmark for a preset gradient threshold through a pre-calibration program. This program selects a verification set containing S sets of standard distribution samples and initiates forward inference, where S is the total number of verification samples. It records the computational levels at different quantization accuracies. Output information entropy value And calculate the mean of the entropy reduction gradient ΔH on the validation set. with standard deviation Then, the preset gradient threshold is determined according to the following calculation formula. Where Γ is the preset gradient threshold, Let be the mean of the entropy-decreasing gradient. The standard deviation of the entropy reduction gradient is used as the basis for establishing the relationship between computational logic and hardware numerical characteristics through this physical quantization mapping process, and the benchmark automatically shields numerical fluctuations caused by the hardware environment.

[0039] When the system encounters a situation where the data distribution in the inference flow shifts, in order to maintain the decision-making topology trajectory stability, the system initiates a dynamic recalibration method based on the feature drift rate, by continuously collecting the feature dimension index values ​​output by the continuous computation level i in real time. And calculate the feature dimension index value. The Euclidean distance between the centroid and the historical feature centroid, if the mean distance is observed to increase within a continuous time window, and the highest frequency of occurrence is detected... If the frequency remains below the threshold, it indicates a mismatch between the current model parameters and the data field. The system retrieves the compensation operator from the memory and steps to correct the preset gradient threshold, ensuring that the convergence path of the classification data in the feature space meets the stability criterion. The average inference latency jitter rate of the system in the heterogeneous hardware environment is reduced to below 5%.

[0040] Example 5: In a heterogeneous computing cluster carrying high-concurrency inference requests, there are differences in processor clock speed and memory bandwidth between different nodes. The system faces timing misalignment when synchronizing feature data streams across nodes. To address this physical constraint, this example provides a parameter mapping method for heterogeneous environments. During the initialization phase, the system's main control unit measures the transmission delay of the feature tensor of the i-th computational level in the cascaded computing model to the logic adjudication unit via the physical bus, and determines the synchronization compensation period according to the following formula. : Where ΔT is the synchronization compensation period, and ξ is the system's preset congestion scheduling coefficient. The physical storage size of the feature tensor. The system bus's measured physical bandwidth is used as the basis for the vote counting unit to dynamically adjust the read / write pointer offset of the index storage area based on the synchronization compensation period ΔT. This ensures that the feature dimension index sequences of K consecutive computational levels within the circular buffer achieve phase alignment at the physical clock level, avoiding false prediction flip-state determinations caused by uneven hardware throughput. The system main control unit then adjusts the based on the measured physical bandwidth of the bus. The system dynamically corrects the addressing step size of the index storage area to compensate for timing jitter caused by the execution logic control node of the heterogeneous hardware platform. It monitors in real time the residence time of the feature tensor of the i-th computation level of the cascaded computing model in physical memory. If the transmission delay exceeds 1.2 times the preset synchronization compensation period ΔT, it retrieves the feature state image data from the processor cache and uses a first-order linear regression algorithm to calculate the feature dimension index value of the missing node, ensuring that the vote counting unit maintains the highest occurrence frequency when processing high-concurrency inference requests. Statistical continuity is maintained to avoid false predictions and flip-state determinations caused by uneven hardware throughput. Where ΔT is the bus physical bandwidth, ΔT is the synchronization compensation period, and i is the hierarchy index. This represents the highest frequency of occurrence.

[0041] When the system detects a situation where data packets are being dropped due to rising network water levels, in order to maintain continuous awareness of decision-making path stability, the system initiates a redundant reconfiguration procedure. This involves reserving a fixed-size feature state mirror area in the processor cache and backing up the information entropy values ​​output from previous computational layers in real time. and feature dimension index value If the vote counting unit detects that the current level data packet verification is invalid, it retrieves historical data from the mirror area as the input benchmark and uses a first-order linear regression algorithm to calculate the feature dimension index value of the missing node. To ensure the highest frequency of occurrence The statistical process has logical continuity under non-ideal channel conditions, and the system's decision delay under sudden traffic surge conditions remains within a constant range of 15ms.

[0042] Example 6: In the execution logic construction process of the cascaded computing model, the system dynamically truncates computing resources by setting a logic branch switch after each computing node. When the logic decision unit outputs a stable state determination result for the i-th computing level, the processor sends a computing termination instruction to the subsequent levels of the current computing pipeline and skips all subsequent loading instructions and calculation steps of weight parameters by modifying the control bits of the task descriptor. Since each computing level... The output feature tensor has been latched in the register before the logical decision; at this point, the system directly calls the first... The classification and judgment data at the computational level are used as the output results, thereby avoiding the power consumption of the processor clock signal switching on redundant computation links while maintaining inference accuracy.

[0043] To address the stateless condition at the start of the cascaded computing model, the system initializes the computing path using a pre-configured environment. Before receiving the first frame of data to be processed, the processor clears the space of K consecutive registers in the index storage area and uses the background constant sequence to determine the baseline information entropy value, which serves as the information entropy value for level 0. Stored in the status register, when the system detects that the entropy reduction gradient ΔH exceeds the preset gradient threshold for M consecutive calculation cycles, it determines that it is currently in the cold start adaptation phase and automatically increases the tolerance coefficient of the preset gradient threshold. After the system enters a stable working cycle, the tolerance coefficient is returned to the unit value through an adaptive feedback loop to ensure the logical stability of the computation graph topology flow path under dynamic load impact conditions.

[0044] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A computational method based on artificial intelligence algorithms, characterized in that, Includes the following steps: Step S101: Obtain the feature tensor of the i-th computational level output for the sample to be processed in the cascaded computation model, and perform channel dimension mean pooling on the feature tensor to obtain the channel activation distribution vector. Step S102: Calculate the information entropy value of the channel activation distribution vector and determine the feature dimension index value corresponding to the element with the largest value in the channel activation distribution vector; Step S103: Write the feature dimension index value into the index storage area. The index storage area records the feature dimension index sequence of K consecutive computation levels, including the i-th computation level, where K is an integer greater than 1. Step S104: Obtain the information entropy value of the (i-1)th computational level, and calculate the entropy reduction gradient between the information entropy value and the information entropy value of the (i-1)th computational level; Step S105: Calculate the highest frequency of occurrence of the same index value in the feature dimension index sequence, and determine the stability state of the cascaded computing model at the i-th computing level based on the highest frequency of occurrence. Step S106: When the entropy reduction gradient is lower than the preset gradient threshold and the stability state is stable, stop the operation logic of the cascaded calculation model at the (i+1)th calculation level and subsequent calculation levels, and output the classification judgment data generated at the i-th calculation level.

2. The calculation method based on artificial intelligence algorithm according to claim 1, characterized in that, Step S105 determines the stability state based on the highest occurrence frequency, including the following steps: Step S201, compare the highest occurrence frequency with a preset frequency determination threshold; Step S202, if the highest occurrence frequency is not lower than the frequency determination threshold, then determine the stability state as a stable state; Step S203, if the highest occurrence frequency is lower than the frequency determination threshold, then determine the stability state as a prediction flip state, and generate a feature reconstruction instruction to drive the cascaded computing model to perform feature extraction at the (i+1)th computing level.

3. The calculation method based on artificial intelligence algorithm according to claim 1, characterized in that, After outputting the classification judgment data, the method also includes: obtaining the current inference depth value of the cascaded computing model; calculating the correction coefficient for the preset gradient threshold based on the current inference depth value; and adjusting the judgment benchmark of subsequent input samples in step S106 using the correction coefficient so that the judgment benchmark decreases as the inference depth increases.

4. The calculation method based on artificial intelligence algorithm according to claim 1, characterized in that, In step S102, the information entropy value of the channel activation distribution vector is calculated using the following method: ,in, Let C be the total number of feature channels, and calculate the information entropy value for the i-th level. This represents the proportion of the activation value of the j-th channel in the channel activation distribution vector to the total activation value.

5. The calculation method based on artificial intelligence algorithm according to claim 1, characterized in that, In step S101, channel dimension mean pooling is performed on the feature tensor, including: calculating the average value of all spatial location elements in each feature channel of the feature tensor; arranging the obtained average values ​​of each feature channel in the order of the original channel index to construct the channel activation distribution vector.

6. The calculation method based on artificial intelligence algorithm according to claim 2, characterized in that, After determining the stability state as the predicted flip state, the process also includes: the rate of change of the highest frequency of occurrence in the statistical cascaded computation model within M consecutive computational levels, where M is an integer between 2 and 5; if the highest frequency of occurrence does not increase within M consecutive computational levels, and the information entropy value is not lower than the preset saturation upper limit threshold, then a computation termination instruction is output, and the remaining computational steps corresponding to that sample in the cascaded computation model are skipped.

7. The calculation method based on artificial intelligence algorithm according to claim 1, characterized in that, In step S103, the storage depth K of the index storage area is set to an integer between 3 and 7, and the index storage area updates the feature dimension index sequence according to the first-in-first-out logic.

8. The calculation method based on artificial intelligence algorithm according to claim 1, characterized in that, Step S106 outputs classification determination data, including: determining the feature dimension index value corresponding to the highest activation value output by the i-th calculation level as the predicted category label; encapsulating the predicted category label and the confidence score calculated by the i-th calculation level to generate an inference result data package.

9. The calculation method based on artificial intelligence algorithm according to claim 1, characterized in that, The cascaded computing model consists of multiple cascaded residual blocks or attention mechanism layers, each of which integrates a logic control node for executing steps S101 to S106.

10. The calculation method based on artificial intelligence algorithm according to claim 1, characterized in that, Step S106 stops the operation logic of the (i+1)th calculation level and subsequent calculation levels, including: generating a skip signal for all calculation levels after the i-th calculation level, so as to release the processor's calculation thread and route it to the next set of data streams to be processed.

Images