Variable Sensitivity Nodes With On-Board Transfer Function Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks implemented in ASICs lack the ability to efficiently adjust transfer functions of nodes without relying on peripheral processing and memories, limiting the speed and efficiency of signal processing.
Innovation Solution
An ASIC-implemented neural network with variable sensitivity nodes that include an input setting channel, logic element, and on-board parameter-setting element, allowing immediate adjustment of transfer functions through stored parameter settings, eliminating reliance on peripheral processing and enhancing processing speed and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If transfer function settings are stored in peripheral memories, then the neural network can be configured, but the processing speed decreases due to reliance on peripheral processing
Solution Approach 1:
The patent merges the transfer function setting storage capability directly into the node circuitry itself, eliminating the need for separate peripheral memories. Each node includes an on-board memory element that stores its transfer function settings locally, allowing immediate access during signal processing without relying on external memory access operations.
Solution Approach 2:
The patent introduces an on-board parameter-setting element as an intermediary component within each node that enables local configuration of transfer function parameters. This intermediary allows the node to self-configure without requiring peripheral processing interventions, thereby maintaining configurability while improving processing speed.
2Speed
If transfer function settings are stored on-board in each node, then processing speed increases, but device complexity increases
Solution Approach 1:
The patent designs the on-board parameter-setting element to serve multiple functions: it stores transfer function settings, enables local configuration, and facilitates immediate parameter adjustment without requiring separate dedicated memory structures. This multi-functional approach reduces overall device complexity despite adding storage capability to each node.
Solution Approach 2:
The patent utilizes parameter changes in the transfer function settings stored in on-board memory to achieve reconfiguration of node behavior. By allowing the transfer function parameters (such as weights and biases) to be dynamically changed within the node itself, the system achieves high processing speed without requiring complex external reconfiguration mechanisms.
3Adaptability or versatility
If immediate weights or parameters are applied to nodes, then adaptability improves, but the complexity of implementing the adjustment mechanism increases
Solution Approach 1:
The patent implements self-service capability within each node through the on-board parameter-setting element that enables the node to adjust its own transfer function settings independently. This self-service mechanism allows immediate application of weights or parameters without requiring complex external adjustment mechanisms or peripheral processing interventions.
Data Source
AI summary
A variable sensitivity node for a neural network that can be implemented as an information processing device such as an ASIC, and can be adjusted simply by applying immediate weights or parameters that change and/or amplify or de-amplify the output of the nodes. The information processing device has one or more nodes each with an input setting channel, a logic element configured to translate the input signal into the output signal based on a mathematical function that includes a parameter setting, and a parameter-setting element configured to set the parameter setting based on a control input to the input setting channel, thereby altering the mathematical relationship of the logic element without compromising speed and efficiency.