Spatially Concatenated Matrix Operations for Neural Network Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computing the output of neural networks is resource-intensive, requiring significant time, memory, and computing resources, which can be improved by optimizing matrix operations.
Innovation Solution
Implementing spatially concatenated matrix operations that combine multiple matrix operations into a single operation, reducing the need for padding and minimizing computational overhead by leveraging existing dimensions and storing information for efficient separation of outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple matrix operations are performed separately in neural network computation, then each operation can be executed with standard procedures, but the total computational time and resource consumption increases significantly
Solution Approach 1:
The patent combines multiple separate matrix operations into a single unified matrix operation by spatially concatenating input matrices along a spatial dimension. This merging allows the system to perform what would otherwise require multiple sequential operations in a single operation, thereby improving computational throughput and reducing computation time.
Solution Approach 2:
The patent introduces spatial concatenation along a new spatial dimension to combine multiple input matrices. This dimensional approach allows multiple operations to be merged without increasing complexity in existing dimensions, enabling efficient batch processing while maintaining the ability to separate outputs afterward using stored dimension information.
2Device complexity
If input matrices are padded to match dimensions for unified processing, then combined matrix operations can be performed, but memory usage and computational overhead increase
Solution Approach 1:
The patent applies local quality by allowing different input matrices to maintain their original dimensions and characteristics while being processed together. Instead of forcing all matrices to conform to a uniform padded size, the system processes each matrix's local region independently within the unified operation, avoiding unnecessary memory consumption from padding.
3Ease of operation
If multiple separate matrix operations are performed, then each operation maintains its independence and clarity, but the overall processing efficiency decreases
Solution Approach 1:
The patent applies segmentation by storing information about the spatial arrangement and dimensions of individual input matrices within the concatenated structure. This allows the unified operation to be segmented back into individual operation results afterward, maintaining operational clarity and traceability while benefiting from the efficiency of combined processing.
Data Source
AI summary
Apparatuses, systems, and techniques to perform matrix computations associated with computing output of a neural network. In at least one embodiment, one or more circuits cause one or more feature maps of one or more neural networks to be spatially concatenated.