Neural Network Unit Deactivation for Accuracy-Energy Balance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial neural networks (ANNs) face challenges in managing computational accuracy, energy consumption, and heat generation, particularly in high-dimensional data processing, where previous pruning methods are static and do not adapt to runtime conditions.
Innovation Solution
A hardware architecture for ANNs allows selective and resettable deactivation of processing units based on predefined conditions, ensuring at least one unit remains active in each layer, with a control unit managing the deactivation and reactivation to balance computational accuracy, energy consumption, and heat generation dynamically.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple processing units are employed to process high-dimensional data in parallel, then computational accuracy and processing capability are improved, but energy consumption and heat generation increase
Solution Approach 1:
The patent implements dynamic deactivation of processing units based on runtime conditions such as input data quality, temperature, and computational workload. The system transitions from static all-unit operation to dynamic selective operation, where processing units are activated or deactivated according to actual needs, resolving the contradiction between maintaining high computational accuracy and reducing energy consumption.
Solution Approach 2:
The system changes the operational state parameter of processing units from fixed (all active) to variable (selectively active/inactive). By monitoring parameters like input quality and temperature, the system adjusts the activation state of processing units, allowing the system to adaptively balance between computational accuracy and energy consumption based on current operating conditions.
2Productivity
If all processing units operate continuously to maintain high computational accuracy, then processing capability is improved, but heat generation increases leading to reduced hardware lifespan
Solution Approach 1:
The system implements periodic or intermittent operation of processing units based on detected conditions. Rather than continuous operation, processing units are activated in periods when high computational accuracy is needed and deactivated when conditions permit, thereby reducing cumulative heat generation while maintaining processing capability when required.
Solution Approach 2:
The system uses self-monitoring mechanisms where processing units or the system as a whole detects its own thermal state and operational needs, automatically adjusting its operation accordingly. This self-regulating behavior reduces unnecessary heat generation while maintaining adequate processing capability, thereby extending hardware lifespan without requiring external intervention.
3Use of energy by stationary object
If static pruning methods are used to reduce hardware complexity, then energy consumption is reduced, but adaptability to runtime conditions is lost
Solution Approach 1:
The patent transforms static pruning into dynamic selective deactivation. Instead of permanently removing or permanently activating processing units, the system dynamically adjusts the operational state of processing units based on runtime conditions. This allows the system to adapt to varying computational needs, input data quality, and thermal conditions while managing energy consumption flexibly rather than through fixed configuration.
Data Source
Figure 1
Figure 2
AI summary
A hardware architecture (10) for an artificial neural network, KNN (1), wherein the KNN (1) comprises a juxtaposition of an input layer (2), multiple processing layers (3a, 3b) and an output layer (4), wherein each layer (2, 3a, 3b, 4) maps a set of input variables onto a set of output variables and wherein output variables of the input layer (2) and of each processing layer (3a, 3b) are input variables of the particular layer (3a, 3b, 4) that follows in the juxtaposition, wherein the hardware architecture (10) comprises a multiplicity of computing units (5a-5h), wherein the realization of each layer (2, 3a, 3b, 4) is distributed over at least two of the computing units (5a-5h) and wherein there is provision for at least one resettable disconnection device (6a-6h) by means of which at least one computing unit (5a-5h) is selectively deactivable independently of the input variables supplied thereto such that at least one other computing unit (5a-5h) remains activated in all layers (2, 3a, 3b, 4) that this computing unit (5a-5h) helps to realize. A corresponding method (100) for operating an artificial neural network, KNN (1), wherein computing units (5a-5h) are selectively deactivated (130) and later reactivated (140). An associated computer program.