Neural Processing Device Noise Adaptive Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural processing devices face reduced inference accuracy when input data contains noise, and increasing resource allocation to improve accuracy leads to increased operation time and power consumption.
Innovation Solution
A neural processing device that includes a noise classifier to determine noise characteristics in input data using Fourier transform values, selecting an appropriate neural network based on these characteristics, and performing inference with weights specific to that network to maintain accuracy while reducing computational resources and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If more resources are allocated to increase the accuracy of the inference result, then the accuracy is improved, but the operation time is increased
Solution Approach 1:
The patent applies dynamics by making the neural network configuration adaptive rather than static. The system dynamically selects between different neural network models (full model and simplified model) based on the noise characteristics of the input data. This dynamic adaptation allows the system to optimize between accuracy and operation time in real-time, choosing the appropriate model complexity according to the actual noise level detected in the input signal.
Solution Approach 2:
The patent changes the parameter of neural network complexity based on noise characteristics. By detecting noise level and type in the input data, the system adjusts which neural network model to use - switching between a full model for low noise conditions and a simplified model for high noise conditions. This parameter change resolves the contradiction by matching model complexity to actual input quality requirements.
2Measurement precision
If more resources are allocated to increase the accuracy of the inference result, then the accuracy is improved, but the power consumption is increased
Solution Approach 1:
The system dynamically adjusts power consumption by selecting different neural network models based on detected noise characteristics. When input data has low noise, the full neural network model is used for maximum accuracy. When noise level is high, the simplified model is selected to reduce computational load and power consumption. This dynamic resource allocation resolves the contradiction between accuracy and power usage.
Solution Approach 2:
The patent changes the computational parameter of the neural network based on noise detection results. By adjusting which model (full or simplified) is executed according to the noise characteristics of input data, the system optimizes the balance between inference accuracy and power consumption, avoiding unnecessary power expenditure when high accuracy is not required due to noisy input.
3Productivity
If a simplified neural network is used to reduce the amount of operation, then the operation time is reduced, but the accuracy of the inference result is reduced
Solution Approach 1:
The patent implements a dynamic model selection mechanism that switches between simplified and full neural networks based on noise characteristics. The system detects the noise level in input data and dynamically chooses the appropriate model complexity - using the simplified model when noise is high (where speed is prioritized) and the full model when noise is low (where accuracy is prioritized). This dynamic approach resolves the contradiction between operation speed and accuracy.
Solution Approach 2:
The system changes the neural network configuration parameter based on detected noise characteristics. By adjusting which model is deployed (simplified or full) according to the actual noise level in input data, the system optimizes the trade-off between operation speed and inference accuracy, ensuring the right model complexity is used for each specific input condition.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The solution enables neural processing devices to achieve high inference accuracy with reduced power consumption and computational resources by dynamically selecting neural networks based on noise types and degrees in input data, improving performance and efficiency.
Implementation Method 1
the noise classifier may determine the noise type and the noise degree based on Fourier transform values of the input data
Data Source
AI summary
Provided is a neural processing device including a noise classifier configured to perform a preprocessing on input data to determine a noise characteristic of the input data, a network selector configured to select one of a plurality of neural networks based on the noise characteristic, and an operator configured to perform inference on the input data based on selected weights corresponding to the selected neural network.