Continuous learning visual perception system based on kolmogorov-arnold network

Abstract

The present application relates to the technical field of computer vision and deep learning, and discloses a continuous learning visual perception system based on Kolmogorov-Arnold network. The present application aims to solve the technical problem that traditional Transformer is prone to serious catastrophic forgetting when dealing with sequential incremental tasks. The system mainly comprises an image embedding module, a hybrid encoder composed of KAN-Transformer layers and a continuous learning control module. The core point is to introduce KAN representation into the feedforward network, and replace the fixed activation function with a learnable B-spline basis function. When processing incremental learning tasks, the system uses the local support characteristics of the B-spline basis function, selectively freezes the learned spline coefficients, and absorbs new task characteristics by expanding the node grid, realizing parameter-level task decoupling at the bottom. The present application is mainly used in complex and variable continuous visual perception classification scenarios, greatly reducing the forgetting rate and saving the parameter increment overhead.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning technology, and in particular to a continuous learning visual perception system based on the Kolmogorov-Arnold network. Background Technology

[0002] Currently, visual perception models based on the Transformer architecture have demonstrated excellent feature extraction capabilities in various image tasks. In practical industrial applications and edge computing scenarios, visual perception systems often face the need for continuous learning. That is, new visual tasks arrive sequentially, and the system needs to gradually absorb knowledge of new tasks and maintain accurate recognition capabilities of already learned tasks without revisiting historical task data.

[0003] However, traditional visual Transformer networks suffer from a severe catastrophic forgetting problem when faced with continuously arriving incremental tasks. The core reason for this technical problem lies in the fact that each neuron in the feedforward network of a traditional Transformer shares the same fixed-form global activation function (e.g., GELU or ReLU). When the gradient signal of a new visual task propagates back through the network to these shared parameters, it inevitably causes destructive interference and overwriting of the weight distribution shaped and solidified by previous tasks. Therefore, how to achieve task-level decoupling of the parameter space from the feature representation structure within the network without destroying existing knowledge representation, thereby overcoming catastrophic forgetting during incremental learning, is a key technical problem that urgently needs to be solved in the field of visual perception. Summary of the Invention

[0004] The purpose of this invention is to provide a continuous learning visual perception system based on the Kolmogorov-Arnold network, in order to solve the technical problem that when traditional visual perception models face continuously arriving incremental tasks, their fixed global activation function is easily disturbed by the gradient signal of the new task, resulting in a serious and catastrophic forgetting of the representations generated on previous tasks.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A continuous learning visual perception system based on a Kolmogorov-Arnold network includes: an image segmentation and embedding module for dividing an input image into multiple image blocks and mapping them to an embedding space to obtain an embedding sequence; a hybrid encoder containing at least one KAN-Transformer layer, wherein the KAN-Transformer layer has a multi-head self-attention module and a KAN feedforward network, wherein the KAN feedforward network is configured with a B-spline activation function containing learnable spline coefficients for nonlinear transformation of feature components; a continuous learning control module for selectively fixing the spline coefficients corresponding to the learned task when incrementally learning a new visual task, and introducing new spline coefficients by adjusting the node sequence of the B-spline activation function to fit the feature distribution of the new visual task; and a classification prediction module for outputting a category prediction result based on the global representation vector output by the hybrid encoder.

[0007] In a preferred embodiment of the present invention, the B-spline activation function is constructed by a linear combination of multiple locally supported B-spline basis functions defined on a uniform node sequence. When processing a new visual task, the continuous learning control module freezes the spline coefficients optimized in the previous task and expands the grid size of the uniform node sequence. New B-spline basis functions and additional spline coefficients are generated by inserting new nodes into the existing node sequence. This structure ensures that the gradient of the new task applies only to the newly added and locally activated coefficients, protecting the original coefficients from interference.

[0008] As a preferred technical solution of the present invention, the initialization of the newly added spline coefficients is based on spline refinement interpolation of the frozen existing spline coefficients, so that the new task can inherit the knowledge starting point of the existing function form to participate in the forward calculation, thereby improving the convergence speed and adaptability of the new task.

[0009] As a preferred technical solution of the present invention, the continuous learning control module is configured with a parameter isolation strategy, which builds a task coefficient library inside the system, allocates and stores an independent set of spline coefficients for each different visual task, so that the nonlinear activation parameters between different tasks are completely isolated from each other, and completely eliminates parameter interference between tasks.

[0010] As a preferred embodiment of the present invention, the system further includes a task identifier module, which is used to automatically determine the target task domain to which the input image belongs when the task label is unknown; based on the determination identifier of the target task domain, the system dynamically retrieves and loads the corresponding spline coefficient set from the task coefficient library into the KAN feedforward network to complete the inference calculation, thereby realizing task adaptation in the inference stage.

[0011] As a preferred technical solution of the present invention, the continuous learning control module, during the training process of a new visual task, uses an elastic weight consolidation method or a synaptic intelligence method to apply additional secondary regularization constraints to the spline coefficients based on the evaluation results of the importance of the parameters in the learned tasks, thereby implementing key protection for the critical intervals that affect the decision-making of the old tasks.

[0012] As a preferred embodiment of the present invention, the hybrid encoder further includes a convolutional KAN feature extraction module connected in series between the image segmentation and embedding module and the KAN-Transformer layer; the convolutional KAN feature extraction module is configured to perform local spatial filtering on the spatial feature map reconstructed from the embedding sequence through depthwise separable convolution, apply B-spline activation transformation with shared mathematical form but independent coefficients along the channel dimension, and finally add the nonlinear output of the convolutional KAN feature extraction module to the original input through skip connections.

[0013] As a preferred embodiment of the present invention, the skip connection is configured with a learnable scaling factor, which is used to adaptively adjust the gradient transfer and data contribution ratio between the local feature map and the direct path during the training phase, so as to avoid the degradation problem that may occur in the local module.

[0014] As a preferred technical solution of the present invention, the classification prediction module includes a KAN classification head, which is configured to project the global representation vector onto a preset category dimension after being transformed by an independent set of B-spline activation functions, and then output the probability distribution of each category after Softmax normalization, so as to flexibly fit the nonlinear decision boundary of each category using a learnable spline function.

[0015] As a preferred technical solution of the present invention, the system is configured with a lookup table quantization strategy for edge-deployed computing. By pre-calculating and storing the B-spline basis function values ​​in each node interval within the effective range of the input values ​​at a fixed step size, a lookup table in fixed-point format is generated. During inference, the corresponding basis function value is obtained by looking up the table according to the given input feature value, and then multiply-add it directly with the spline coefficient, thereby significantly reducing the inference computing latency of the system deployed at the edge.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] First, leveraging the inherent local support characteristics of B-spline basis functions, the network activates only spline parameters within local intervals during forward feature mapping. This ensures that gradient updates during new task learning focus solely on the coefficients of node intervals covered by the current data distribution, with minimal impact on coefficients of other node intervals representing prior knowledge. This structure achieves task-level decoupling of the parameter space at the underlying mechanism, thereby significantly reducing the catastrophic forgetting rate during continuous learning without needing to access historical data.

[0018] Secondly, compared to traditional multilayer perceptrons which must store knowledge for different tasks through a large fully connected weight matrix, this invention uses nonlinear B-spline coefficients for fitting local features and archiving knowledge. While maintaining the same level of complexity in representing nonlinear decision boundaries, it avoids the excessive stacking of network parameters, significantly reducing the growth rate of the total number of model parameters when dealing with sequential incremental tasks, and exhibits extremely high parameter utilization efficiency. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall structure of a continuous learning visual perception system based on the Kolmogorov-Arnold network in a preferred embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0021] The visual perception system proposed in this invention, such as Figure 1 As shown, the core idea lies in introducing the mathematical property of the Kolmogorov-Arnold representation theorem—that a multivariate continuous function can be precisely expressed by the superposition of a finite number of univariate functions—into the feedforward network structure design of the visual Transformer. By replacing the fixed activation function in the traditional multilayer perceptron with learnable B-spline basis functions, the network can selectively freeze, expand, or isolate spline coefficients when faced with continuously arriving incremental visual tasks, absorbing the feature distribution of new tasks without disrupting existing knowledge representation. This design motivation stems from the observation that in the traditional visual Transformer, each neuron in the feedforward network shares the same activation function shape. When the gradient signal of a new task propagates back to these shared parameters, it inevitably interferes with the weight distribution shaped by previous tasks. The local support characteristics of the B-spline basis functions naturally provide a structural basis for task-level decoupling of the parameter space.

[0022] This implementation focuses on three aspects: first, the specific network structure and data flow of the KAN-ViT hybrid encoder; second, the continuous learning strategy for incremental task scenarios; and third, the task adaptation mechanism during the inference phase. Furthermore, a lightweight convolutional KAN variant scheme for edge deployment is also presented.

[0023] In a typical implementation scenario, the input image First, the image undergoes a patch segmentation operation. Specifically, the image is divided along the spatial dimensions into... There are three non-overlapping image blocks, each with a spatial size of [missing information]. In this embodiment, take For a resolution of The input image yields 196 image patches. Each image patch is flattened to a length of... After a one-dimensional vector is generated, it is mapped to a linear projection layer. 3D embedding space, to obtain the embedding sequence ,in In this embodiment, 768 is used. At the same time, a learnable classification label vector is concatenated at the beginning of the embedded sequence, and a learnable positional code is superimposed to preserve the spatial positional information of each image patch.

[0024] Following the patch embedding layer and before entering the Transformer encoding layer, this invention introduces a convolutional KAN feature extraction module to capture local spatial relationships within image patches and between adjacent image patches. The design rationale behind this module is that while standard self-attention mechanisms excel at modeling global dependencies, their ability to perceive low-level features such as local textures and edges is relatively weak. In continuous learning scenarios, different visual tasks often exhibit significant differences at the low-level local feature level. If these local features can be fully and tunably extracted in the early stages of encoding, it will facilitate the subsequent KAN layer's refined expression of task-specific features.

[0025] The specific structure of this convolutional KAN module is as follows: The input features first pass through a layer... The depthwise separable convolution, in this implementation, performs a two-dimensional morphological rearrangement of the embedded sequence, that is... Remodeling The spatial feature map is used, and depthwise separable convolutions perform spatial filtering independently on each channel, extracting local spatial patterns with extremely low parameter cost. The output of the depthwise separable convolutions is then applied along the channel dimension to a KAN activation layer, which performs a nonlinear transformation on the scalar value at each channel location through a set of B-spline basis functions. The KAN activation layer here shares the same mathematical form as the spline activation used in the KAN feedforward network described later, but has an independent set of learnable coefficients. Finally, the output of the convolutional KAN module is added to the module input through a one-hop connection. The entire process can be described as follows:

[0026]

[0027] in This indicates a depthwise separable convolution operation. This represents the spline activation transformation along the channel dimension. This residual structure ensures that even if the contribution of the convolutional KAN module is small in some tasks, information can still be losslessly transferred through the skip path, avoiding the gradient vanishing problem caused by module degradation. In practical deployments, this skip connection can also introduce a learnable scaling factor. This allows for adaptive adjustment of the contribution ratio between local features and the direct path during training.

[0028] The feature sequence processed by the convolutional KAN module is then fed into... The encoder body is composed of stacked KAN-Transformer layers. In this embodiment, we take... This is consistent with the standard ViT-Base. The computation process for each KAN-Transformer layer is as follows.

[0029] Input features (No. Layer input, First, a layer normalization operation is performed to obtain normalized features. The layer normalization adopts the Pre-Norm paradigm, which applies normalization before the self-attention and feedforward networks. This design has been widely verified to have better training stability in deep Transformers.

[0030] The normalized features are fed into the multi-head self-attention module. Let the number of attention heads be . Each head has a dimension of Then the query, key, and value matrices are obtained by linear projection: , , ,in Attention weights are calculated using a scaled dot product and then normalized using Softmax. This result is multiplied by the value matrix to obtain the attention output, which is then merged into a multi-head output via a linear projection layer. The output of the attention module is added to the input through a residual connection to obtain... .

[0031] Subsequently, After undergoing layer normalization again, the material enters the KAN feedforward network, which is the key structure that distinguishes this invention from existing technologies. Traditional Transformer feedforward networks typically consist of two linear transformation layers and a fixed activation function (such as GELU or ReLU), and their expressive power is limited by the preset shape of the activation function. This invention replaces this with a KAN feedforward network, and its specific calculation process is as follows:

[0032] First, the input features are projected through an up-dimensional projection matrix. Mapping to higher-dimensional space:

[0033]

[0034] at this time The scaling factor is set to 4 times to maintain consistency with the standard Transformer design, but it can also be adjusted to 2 times or 8 times depending on the complexity of the specific task.

[0035] right A B-spline activation function is applied independently to each component along the feature dimension. For the The first token Each feature dimension has an activation output as follows:

[0036]

[0037] in Let be the order of the B-spline, which is taken as in this embodiment. That is, using third-order (fourth-order) B-splines; Let B-spline basis functions be defined on a uniform sequence of nodes; These are learnable spline coefficients. B-spline basis functions are defined by a recursive relation; the zeroth-order basis function is a piecewise constant function, and higher-order basis functions are constructed stepwise by linear combinations of lower-order basis functions. The node sequence is uniformly distributed according to the range of the input data at the start of training, with a grid size of [missing information]. Control the number of nodes, A larger value indicates finer segmentation of the spline and a stronger fitting ability. In the initial training phase of this embodiment, the value is taken as... .

[0038] It should be noted that an important mathematical property of the B-spline activation function mentioned above is its local support (each basis function...). Only in adjacent (taking non-zero values ​​within a range of nodes), which means that when the input value... When the data falls within a specific interval, only a few basis functions participate in the calculation, while the contributions of the remaining basis functions are zero. This property is particularly crucial in continuous learning scenarios: when the data distribution of a new task is mainly concentrated in a sub-region of the input space, it is only necessary to update the spline coefficients corresponding to that sub-region, without affecting the entire parameter set, thus naturally mitigating the risk of catastrophic forgetting at the structural level.

[0039] Features after B-spline activation transformation are projected through a reduced-dimensional projection matrix. Compress back to the original dimension:

[0040]

[0041] Finally, the output of this layer is obtained through residual connections:

[0042]

[0043] go through After KAN-Transformer encoding, the output vector corresponding to the classification label position is taken, or a global average pooling operation is performed on the output of all tokens to obtain the image-level global representation vector. The vector is then fed into the classification head to complete the final category prediction. In this invention, the classification head itself also adopts a KAN structure instead of a simple linear layer. That is, the global representation vector is transformed by a set of B-spline activation functions, then projected onto the category dimension and subjected to Softmax normalization to output the probability distribution of each category. The motivation for using a KAN classification head is that the category boundaries of different visual tasks often have non-linear characteristics. Learnable spline functions can more flexibly fit these decision boundaries, and their coefficients can also be isolated and managed according to the task dimension.

[0044] The continuous learning process is described in detail below. The continuous learning framework of this invention supports a task-incremental learning paradigm, that is, new visual tasks arrive sequentially, and the system needs to gradually absorb the knowledge of new tasks and maintain the ability to recognize already learned tasks without revisiting historical task data.

[0045] Let's take a three-stage incremental learning process as an example. In the first stage, the system receives training data for the first visual classification task (e.g., a general object classification dataset containing 10 categories). At this point, the spline coefficients of all KAN layers in the KAN-ViT hybrid encoder... Starting with random initialization, optimization is performed using standard backpropagation and gradient descent, employing the cross-entropy loss function. The initial grid size is set to... spline order After training is complete, a snapshot of the coefficients of all KAN layers is saved as the knowledge archive for the first task.

[0046] When the second visual task arrives—for example, a fine-grained classification dataset containing 100 categories—the system performs the following operation: The KAN layer coefficients corresponding to the first task are... Freeze the KAN layer so that it no longer receives gradient updates during subsequent training. At the same time, expand the grid size of the KAN layer from... Increase to The physical meaning of mesh expansion is to insert new nodes into the original node sequence, thereby introducing new B-spline basis functions and a corresponding new set of coefficients. Due to the local support of B-spline basis functions, the newly added basis functions mainly cover the gaps between the original basis functions, thus the new coefficients... Initialization can be done based on spline refinement interpolation of the existing coefficients. The advantage of this is that the training starting point of the new task inherits the function form learned in the first task, rather than starting from zero. During forward propagation, all basis functions, including the old basis functions corresponding to the frozen coefficients and the newly added basis functions corresponding to the new coefficients, participate in the calculation, ensuring that the new task can utilize existing knowledge while performing incremental adaptation.

[0047] Regarding the constraint strategy for parameter updates, this invention supports two optional regularization methods to further suppress forgetting. The first is the elastic weight consolidation method, the core idea of ​​which is to use the Fisher information matrix to quantify the importance of each parameter in the learned task, and add a regularization term to the loss function of the new task:

[0048]

[0049] in For the classification loss of new tasks, These are the diagonal elements of the Fisher information matrix. These are the parameter values ​​at the end of training for the old task. This is the regularization intensity hyperparameter. The Fisher information matrix is ​​calculated by taking the squared expectation of the log-likelihood gradient of the model parameters with respect to the training data of each task after training for that task. For the KAN layer, coefficients with higher Fisher information content correspond to spline intervals that have a greater impact on decisions made in previous tasks. These coefficients are subject to stronger constraints during training for new tasks, thus effectively protecting existing knowledge.

[0050] The second approach is synaptic intelligence. This method does not require explicit computation of the Fisher information matrix; instead, it accumulates the contribution of each parameter to the decrease in the loss function online during training. Specifically, for each parameter... Maintain a cumulative contribution ,in Iterate through all training steps. After the task is completed, Parameters with larger values ​​are considered more important to the current task, and a secondary regularization constraint is applied during training for the next task. Synaptic intelligence methods have lower computational overhead than elastic weight consolidation, making them suitable for scenarios with a large number of tasks or limited computational resources.

[0051] When a third visual task arrives (e.g., a large-scale classification task with 200 categories), the system can employ a parameter isolation strategy. This strategy maintains a completely independent set of KAN coefficients for each task, ensuring that coefficients from different tasks do not interfere with each other. Specifically, in each KAN layer of the network, a coefficient library is set up, storing the coefficient matrix corresponding to each task. During forward inference, the coefficient set of the target task is retrieved from the coefficient library based on the task identifier and loaded into the KAN layer. This strategy completely eliminates parameter interference between tasks, at the cost of a linear increase in coefficient storage as the number of tasks increases. However, since the number of parameters in B-spline coefficients is much smaller than the number of parameters in the fully connected weight matrix of a traditional MLP (under the same input-output dimensions, the number of parameters in a KAN layer mainly depends on...), this strategy is more efficient. Instead Therefore, even with parameter isolation, the total parameter growth is still significantly lower than that of an equivalent-sized MLP scheme.

[0052] During the inference phase, this invention includes a task identifier module to automatically determine the task domain of an input image when the task label is unknown. This task identifier can be implemented in several ways: one lightweight approach is to synchronously record the distribution statistics (such as mean vector and covariance matrix) of the training data for each task in the global representation space during training; during inference, the Mahalanobis distance between the test image representation and the distribution of each task is calculated, and the task with the smallest distance is selected as the predicted task. Another approach is to train a small auxiliary classification network that takes the global representation as input and outputs the task label. After the task label is determined, the system dynamically loads the corresponding KAN layer coefficient configuration from the coefficient library, completes the inference calculation, and outputs the classification probability. If the task label is known during inference in the application scenario (e.g., the task type is explicitly specified in a specific business process), the task identification step can be skipped, and coefficients can be loaded directly according to the task label.

[0053] To verify the effectiveness of the above method, experiments were conducted on standard incremental learning benchmarks. The experimental setup was as follows: the base datasets were CIFAR-10 (10 classes), CIFAR-100 (100 classes), and TinyImageNet (200 classes), which were uniformly segmented by class to construct incremental task sequences. The image patch size of the KAN-ViT encoder was [not specified]. Embedded Dimensions Transformer layer number Attention count Initial grid size spline order Training was performed using the AdamW optimizer with an initial learning rate of [missing information]. Cosine annealing scheduling was used, with each task trained for 50 epochs and a batch size of 128. For baseline comparison, a ViT model with the same architecture but using a standard MLP (GELU activation function) as the feedforward network was trained under the same incremental task settings.

[0054] Experimental results show that on the Split-CIFAR and Split-TinyImageNet incremental benchmarks, the ViT model using a KAN feedforward network reduces the average forgetting rate from 95.2% to 71.0% compared to the ViT model using an MLP feedforward network, a reduction of 24.2 percentage points. Simultaneously, the growth rate of the total number of model parameters after training for each task is reduced by approximately 40%. The significant reduction in forgetting rate is mainly attributed to the natural parameter isolation effect brought about by the local support of B-spline coefficients—gradient updates during the new task learning process concentrate on the coefficients corresponding to the newly added or currently distributed node intervals, while having minimal impact on the coefficients of other node intervals. The reduction in parameter growth stems from the fact that the KAN layer stores task knowledge in the form of spline coefficients rather than fully connected matrices, requiring fewer parameters to express nonlinear transformations of equivalent complexity.

[0055] The following describes a lightweight convolutional KAN variant designed for edge devices. In resource-constrained edge deployment scenarios, the computational cost of the standard KAN-ViT model may exceed hardware capabilities. Therefore, this invention provides a simplified convolutional KAN network as an alternative. The core idea of ​​this solution is to replace the fixed activation function after the convolutional layer in the standard convolutional neural network with... The KAN activation layer. Specifically, the input feature map is first applied with... Standard convolution (or depthwise separable convolution) completes spatial feature extraction, and then through a... Convolution adjusts the number of channels to the target dimension, and finally applies a B-spline activation transformation along the channel dimension to each spatial location. Because The computational cost of KAN activation is much lower than that of fully connected KAN layers. This scheme significantly reduces the computational cost while retaining the expressive flexibility brought by learnable activation functions.

[0056] To further reduce inference latency, this invention proposes a lookup table quantization strategy for B-spline basis functions. After training, for each KAN activation layer, the values ​​of all B-spline basis functions are pre-calculated within the valid range of the input values ​​at a fixed step size, generating a lookup table. During inference, for a given input value, the corresponding basis function value is obtained by looking up the table, and then multiplying and adding it with the spline coefficients to obtain the activation output, eliminating the need for real-time recursive calculations. This lookup table can be stored in a fixed-point number format to adapt to the computing power of embedded processors. In validation experiments on the MNIST handwritten digit classification dataset, the network using the above-mentioned lightweight convolutional KAN structure improved the classification accuracy by 0.3 percentage points while reducing the number of parameters by 15% compared to a standard CNN of the same size, indicating that the KAN activation function can bring meaningful performance gains even on very small models.

[0057] The above embodiments are merely specific examples of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent substitutions or modifications made within the spirit and principles of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A continuous learning visual perception system based on a Kolmogorov-Arnold network, characterized in that, include: An image segmentation and embedding module is used to divide the input image into multiple image blocks and map them to an embedding space to obtain an embedding sequence; a hybrid encoder includes at least one KAN-Transformer layer, which has a multi-head self-attention module and a KAN feedforward network, which is configured with a B-spline activation function containing learnable spline coefficients for nonlinear transformation of feature components. The continuous learning control module is used to selectively fix the spline coefficients corresponding to the learned tasks when incrementally learning new visual tasks, and to introduce new spline coefficients by adjusting the node sequence of the B-spline activation function to fit the feature distribution of the new visual task. The classification prediction module is used to output category prediction results based on the global representation vector output by the hybrid encoder.

2. The continuous learning visual perception system based on the Kolmogorov-Arnold network according to claim 1, characterized in that: The B-spline activation function is constructed by a linear combination of multiple locally supported B-spline basis functions defined on a uniform node sequence. When processing a new visual task, the continuous learning control module freezes the spline coefficients optimized in the previous task and expands the grid size of the uniform node sequence. It generates new B-spline basis functions and new spline coefficients by inserting new nodes into the original node sequence.

3. The continuous learning visual perception system based on the Kolmogorov-Arnold network according to claim 2, characterized in that: The initialization of the newly added spline coefficients is based on spline refinement interpolation of the frozen existing spline coefficients, so that the new task can inherit the knowledge starting point of the existing function form to participate in the forward calculation.

4. The continuous learning visual perception system based on the Kolmogorov-Arnold network according to claim 1, characterized in that: The continuous learning control module is configured with a parameter isolation strategy, which builds a task coefficient library inside the system, allocates and stores an independent set of spline coefficients for each different visual task, so that the nonlinear activation parameters between different tasks are completely isolated from each other.

5. The continuous learning visual perception system based on the Kolmogorov-Arnold network according to claim 4, characterized in that: The system also includes a task identifier module, which is used to automatically determine the target task domain to which the input image belongs when the task label is unknown; based on the determination identifier of the target task domain, the system dynamically retrieves and loads the corresponding spline coefficient set from the task coefficient library into the KAN feedforward network to complete the inference calculation.

6. The continuous learning visual perception system based on the Kolmogorov-Arnold network according to claim 1, characterized in that: During the training process of a new visual task, the continuous learning control module applies additional quadratic regularization constraints to the spline coefficients based on the evaluation results of the importance of the parameters in the learned tasks, using an elastic weight consolidation method or a synaptic intelligence method.

7. The continuous learning visual perception system based on the Kolmogorov-Arnold network according to claim 1, characterized in that: The hybrid encoder also includes a convolutional KAN feature extraction module connected in series between the image segmentation and embedding module and the KAN-Transformer layer. The convolutional KAN feature extraction module is configured to perform local spatial filtering on the spatial feature map reconstructed from the embedding sequence through depthwise separable convolution, apply B-spline activation transformation with shared mathematical form but independent coefficients along the channel dimension, and finally add the nonlinear output of the convolutional KAN feature extraction module to the original input through skip connections.

8. The continuous learning visual perception system based on the Kolmogorov-Arnold network according to claim 7, characterized in that: The skip connections are configured with learnable scaling factors to adaptively adjust the gradient transfer and data contribution ratio between local feature maps and direct paths during the training phase.

9. The continuous learning visual perception system based on the Kolmogorov-Arnold network according to claim 1, characterized in that: The classification prediction module includes a KAN classification head, which is configured to project the global representation vector onto a preset number of categories after being transformed by an independent set of B-spline activation functions, and then output the probability distribution of each category after Softmax normalization.

10. The continuous learning visual perception system based on the Kolmogorov-Arnold network according to claim 1, characterized in that: The system is configured with a lookup table quantization strategy for edge-oriented computation. It generates a fixed-point number format lookup table by pre-compiling and storing the B-spline basis function values ​​in each node interval within the effective range of the input values ​​at a fixed step size. During inference, it retrieves the corresponding basis function value from the table based on the given input feature value and then directly performs a multiplication and addition operation with the spline coefficient.

Images