Processing Core Metadata Conditional Graph Execution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine intelligence systems, particularly artificial neural networks (ANNs), are computationally and energy-intensive due to large data structures and immense discrete computations, making them resource-hungry and difficult to parallelize, with modern deep learning approaches exacerbating these issues.
Innovation Solution
A processing core and computer-implemented method for conditional execution of directed graphs, utilizing metadata to conditionally execute instructions and operations, reducing computational complexity and memory bandwidth requirements by dynamically modifying data tiles and generating metadata for efficient execution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional ANNs with fully connected layers are used, then comprehensive computation is achieved, but computational complexity increases with power law relationship to layer size
Solution Approach 1:
The patent segments the fully connected layer into multiple sparse sub-layers, each handling a subset of connections. This segmentation reduces the computational complexity from O(n²) to O(kn) where k is the sparsity factor, while maintaining the ability to perform comprehensive computation across all sub-layers.
Solution Approach 2:
The patent applies partial action by computing only the necessary subset of operations in each sparse sub-layer rather than all possible connections. This partial computation approach reduces overall computational complexity while the aggregation of results from multiple sub-layers ensures comprehensive computation is achieved.
2Measurement precision
If deep learning approaches with substantial layers and recursive relationships are used, then model accuracy is improved, but computational resources and energy consumption increase
Solution Approach 1:
The patent segments deep neural networks into multiple sparse sub-layers within each computational layer. This segmentation maintains the depth and accuracy of deep learning models while reducing the energy consumption of each layer by eliminating redundant computations in fully connected structures.
Solution Approach 2:
The patent changes the density parameter of the network from fully connected (density=1) to sparsely connected (density<1). This parameter change reduces energy consumption proportionally to the sparsity factor while maintaining model accuracy through the aggregation of results from multiple sparse sub-layers.
3Adaptability or versatility
If large data structures are used in machine intelligence systems, then model capacity is increased, but memory bandwidth requirements and computational resources increase
Solution Approach 1:
The patent segments large data structures into smaller subsets that are processed in parallel across multiple sparse sub-layers. This segmentation improves memory bandwidth efficiency by reducing the amount of data that needs to be transferred at any given time, while the parallel processing across sub-layers maintains overall model capacity.
Solution Approach 2:
The patent introduces a new dimension of sparsity into the data structure organization, transforming the computation from a dense 2D matrix operation to a sparse 3D tensor operation across multiple sub-layers. This dimensional change improves memory bandwidth efficiency by enabling more efficient memory access patterns and reducing redundant data transfers.
Data Source
AI summary
A processing core for the efficient execution of a directed graph is disclosed. The processing core includes a memory and a first and a second data tile stored in the memory. The first and second data tiles include a first and a second set of data elements stored contiguously in the memory. The processing core also includes metadata relationally stored with the first data tile in the memory. The processing core also includes an execution engine, a control unit, and an instruction. Execution of the instruction uses the execution engine, a first data element in the first set of data elements, and a second data element in the second set of data elements. The control unit conditions execution of the instruction using the metadata. A standard execution of the instruction generates a standard output. A conditional execution of the instruction operation generates a conditionally executed output.