Neural Network Processor Weight Supply Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural networks require significant configuration changes when adjusting the number of stages or layers, leading to inefficiencies in processing and increased complexity, especially when monitoring intermediate result values results in a large number of signals.
Innovation Solution
A neural network processor is designed with a weight-value-supply control unit and a control-processor unit that allows for flexible configuration adjustments by using registers to store and select weight values and data, enabling serial or parallel operations without the need for multiple neuron cores, thereby streamlining the processing without massive configuration changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple neuron cores and storage devices are used to adjust the number of parallel operations, then the processing capability is improved, but the configuration complexity increases massively
Solution Approach 1:
A single neuron core is designed to perform multiple functions by sequentially executing different operations with different weight values. The neuron core can process multiple input signals and produce multiple output signals through time-division multiplexing, eliminating the need for multiple dedicated neuron cores for parallel operations.
Solution Approach 2:
The neuron core operates in periodic cycles, where each cycle processes one input signal with a specific weight value. By controlling the operation timing and weight value selection periodically, the system achieves parallel processing capabilities through sequential execution, reducing hardware complexity while maintaining productivity.
2Speed
If the number of neuron cores is increased to perform more parallel operations, then the processing speed is improved, but the device complexity increases
Solution Approach 1:
Weight values are pre-loaded into storage devices before processing operations. The control device selects and supplies appropriate weight values to the neuron core in advance based on the required operation, enabling fast switching between different processing modes without requiring physical reconfiguration of the hardware architecture.
Solution Approach 2:
The system dynamically adjusts the weight values supplied to the neuron core based on the processing requirements. The control device can change weight values during operation, allowing the same hardware configuration to adapt to different processing speeds and operational modes without physical modification.
3Measurement precision
If intermediate result values are monitored in a cascade connection, then the processing accuracy is improved, but the number of signals becomes huge
Solution Approach 1:
The patent extracts and stores intermediate result values in storage devices during the processing sequence. By taking out intermediate results from the signal flow and storing them separately, the system enables monitoring and retrieval of intermediate values without adding them to the active signal count, thus maintaining measurement precision while avoiding signal proliferation.
Data Source
AI summary
A neural network not requiring massive changes in configuration when changing the number of stages (number of levels) of the neural network. This neural network is provided with at least one neuron core 10 performing an analog multiply-accumulate operation and a weight-value-supply control unit 30 supplying the weight value to the neuron core 10. This neural network is subjected to control processing by a control-processor unit 40 controlling the supply of the weight value from said weight-value-supply control unit 30 in synchronization with the timing of the analog multiply-accumulate operation of the neuron core 10, and processing the data output from the neuron core 10 at every analog multiply-accumulate operation performed by said neuron core as serial data and/or parallel data.