Inverted Convolution Gradient Update for Image Processing Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image processing models face inefficiencies in gradient updating, leading to long training times and high energy consumption due to low gradient updating efficiency using the gradient descent method.
Innovation Solution
The method involves determining convolution kernels and convoluted data for each convolution layer using a DMA controlling module, storing them in cache spaces, and performing inverted convolution calculations to update original feature and weight gradients, thereby improving efficiency and reducing training time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the gradient descent method is used for training image processing models, then the model can be trained with standard procedures, but the gradient updating efficiency is low, training time is long and energy consumption is high
Solution Approach 1:
The patent applies inversion by transforming the traditional forward convolution operation into an inverted convolution operation for gradient calculation. Instead of following the standard forward propagation direction, the method performs convolution in the reverse direction during backpropagation, enabling parallel computation of gradients for both image features and weights simultaneously, thus resolving the time efficiency problem
Solution Approach 2:
The patent merges the separate gradient calculation processes for image features and weights into a single inverted convolution operation. By combining these operations and utilizing cache memory to store intermediate results, the method eliminates redundant computations and achieves simultaneous gradient updates, improving productivity while reducing training time
2Productivity
If the gradient descent method is used for training image processing models, then the model can be trained with standard procedures, but the energy consumption is high
Solution Approach 1:
By inverting the convolution operation direction, the patent enables more efficient gradient computation that reduces the number of computational steps required. This inversion allows the system to calculate gradients for multiple parameters simultaneously through a single operation, thereby reducing energy consumption while maintaining high gradient updating efficiency
Solution Approach 2:
The patent performs preliminary actions by pre-computing and storing convolution kernels and intermediate results in cache memory before the actual gradient update. This preparation work is done once and reused during backpropagation, avoiding redundant energy-consuming operations and significantly reducing overall energy consumption during training
3Ease of operation
If traditional convolution operations are performed during backpropagation, then the gradient calculation follows standard procedures, but the computation requires multiple separate operations increasing time and energy costs
Solution Approach 1:
The patent merges multiple separate gradient calculation operations into a single inverted convolution operation. By combining the gradient calculations for image features, weights, and bias into one unified operation that leverages cache memory, the method maintains computational simplicity while dramatically improving gradient update speed and productivity
Data Source
AI summary
The present disclosure is applied for artificial intelligence (AI) technology field and provided a method for gradient updating of an image processing model and a related apparatus thereof. The method includes: determining a convolution kernel and convoluted data corresponding to each convolution layer by invoking a direct memory access (DMA) controlling module, according to convolution parameters of each convolution layer in the image processing model, and storing the convolution kernel and the convoluted data into a first cache space and a second cache space, respectively, the convolution kernel including a convolution kernel for an original image feature gradient and a convolution kernel for an original weight gradient; and performing an inverted convolution calculation based on the convolution kernel in the first cache space and the convoluted data in the second cache space to update the original image feature gradient and the original weight gradient of each convolution layer.