Stochastic Optimization System Using Segmented Neuron Groups
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As the scale of optimization problems increases, existing optimization apparatuses using Ising-type energy functions and neural networks require significant hardware and resources, leading to increased complexity and power consumption, limiting their ability to efficiently handle large-scale operations.
Innovation Solution
An optimization system that calculates energy changes by selecting neuron groups, reducing the need for extensive hardware by separating the calculation of energy changes into local and non-local fields, allowing for efficient updates using a stochastic search method, thereby reducing hardware requirements and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the scale of optimization problems increases, then the number of bits increases, but the amount of hardware increases
Solution Approach 1:
The patent divides all bits into multiple groups and processes them in sequential stages. In each stage, only a subset of bits (neuron group) is actively updated while others are maintained, allowing large-scale problems to be solved by repeatedly applying the same hardware configuration to different bit subsets rather than requiring all bits to be processed simultaneously.
Solution Approach 2:
The patent pre-calculates and stores the initial values of local fields for each bit based on the current states of other bits before the update process begins. This preliminary calculation allows the hardware to directly use these pre-computed values during the stochastic search without performing complex real-time calculations, reducing the computational burden on the hardware during the actual optimization process.
2Quantity of substance
If the scale of optimization problems increases, then the number of bits increases, but power consumption increases
Solution Approach 1:
By dividing the optimization process into stages where only a portion of bits are updated in each stage, the patent reduces the number of active computational operations performed simultaneously. This staged approach allows the same hardware to handle larger problems by distributing the computational workload across multiple cycles rather than requiring peak power for all bits at once.
Solution Approach 2:
The patent uses a storage unit to store connection information (weighting coefficients) that defines the neural network structure. This stored information is reused across multiple update cycles and stages, allowing the hardware to process larger problems without proportionally increasing the physical connectivity hardware, thereby reducing power consumption.
3Quantity of substance
If the scale of optimization problems increases, then the number of bits increases, but the hardware scale increases
Solution Approach 1:
The patent divides all bits into multiple groups and processes them in sequential stages. In each stage, only a subset of bits (neuron group) is actively updated while others are maintained, allowing large-scale problems to be solved by repeatedly applying the same hardware configuration to different bit subsets rather than requiring all bits to be processed simultaneously.
Solution Approach 2:
The same hardware configuration and update rules are reused across multiple stages and different bit groups. The neural network circuit performs the same stochastic search operations on different subsets of bits in different stages, making the hardware multi-functional and capable of handling large-scale problems without requiring proportionally more physical hardware.
Data Source
AI summary
An optimization apparatus calculates a first portion, among energy change caused by change in value of a neuron of a neuron group, caused by influence of another neuron of the neuron group, determines whether to allow updating the value, based on a sum of the first and second portions of the energy change, and repeats a process of updating or maintaining the value according to the determination. An arithmetic processing apparatus calculates the second portion caused by influence of a neuron not belonging to the neuron group and an initial value of the sum. A control apparatus transmits data for calculating the second portion and the initial value to the arithmetic processing apparatus, and the initial value and data for calculating the first portion to the optimization apparatus, and receives the initial value from the arithmetic processing apparatus, and a value of the neuron group from the optimization apparatus.