Deep Learning Attention Control for Partial Image Decisions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning networks face high computational costs due to the need to process entire images, which can be inefficient in determining optimal actions, especially in tasks like video games where only partial information is necessary.
Innovation Solution
Implementing a continuous control of attention mechanism that alternates between selecting observation-actions to observe partial images and control-actions to obtain rewards, updating respective value functions, thereby reducing computational complexity by focusing on sub-areas of the environment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the entire image is processed to make decisions, then the accuracy of decision-making is improved, but the computational cost increases
Solution Approach 1:
The patent divides the image processing task into two segments: (1) an observation network that selects and processes only relevant sub-areas of the image, and (2) a control network that makes decisions based on the processed sub-areas. This segmentation allows the system to avoid processing the entire image while maintaining decision-making accuracy for tasks where only partial information is necessary.
Solution Approach 2:
The patent applies local quality by using attention mechanisms that selectively focus computational resources on specific sub-areas of the image that are most relevant to the current task. The observation network learns to identify and process only the locally relevant portions of the image, allocating computational quality unevenly across different regions based on their importance.
2Loss of information
If the entire image is processed, then complete information is obtained, but the processing time increases
Solution Approach 1:
The observation network performs preliminary action by pre-processing the image to identify and extract only the relevant sub-areas before the control network makes decisions. This preliminary selection of information reduces the amount of data that needs to be processed subsequently, thereby reducing processing time while maintaining the necessary information completeness for decision-making.
Solution Approach 2:
The patent applies partial action by processing only a subset of the image (relevant sub-areas) rather than the entire image. The observation network determines which partial portions of the image are sufficient for making optimal decisions, avoiding the excessive processing of irrelevant regions and thus reducing processing time.
3Measurement precision
If deep learning networks are used to extract features, then the feature extraction capability is improved, but the computational complexity increases
Solution Approach 1:
The patent segments the deep learning architecture into two separate networks: an observation network that handles feature extraction from image sub-areas, and a control network that handles decision-making. This segmentation reduces the computational complexity of each individual network compared to a single full-image processing network, while maintaining overall feature extraction capability.
Solution Approach 2:
The observation network performs partial feature extraction by processing only relevant sub-areas of the image rather than the entire image. This partial action reduces the computational complexity of feature extraction while maintaining the capability to extract necessary features for decision-making.
Data Source
AI summary
A computer-implemented method for reducing computation cost associated with a machine learning task performed by a computer system by implementing continuous control of attention for a deep learning network includes initializing a control-value function, an observation-value function and a sequence of states associated with a current episode. If a current epoch associated with the current episode is odd, an observation-action is selected, the observation-action is executed to observe a partial image, and the observation-value function is updated based on the partial image and the control-value function. If the current epoch is even, a control-action is selected, the control-action is executed to obtain a reward corresponding to the control-action, and the control-value function is updated based on the reward and the observation-value function.