Unified Computation Circuit for Time Series ML Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning devices face challenges in optimizing circuit size and power consumption during machine learning inference, particularly when handling time series data, as they often require separate computation resources for data conversion and inference processes.
Innovation Solution
A machine learning device is designed with a unified computation circuit unit that performs both data conversion and inference tasks, utilizing operators like adders, subtractors, and multipliers to process matrices and vectors, and employing transforms like Hadamard, discrete Fourier, or discrete cosine transforms to convert time series data into frequency features, thereby reducing circuit size and enhancing computation speed while minimizing power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If separate computation resources are used for data conversion and inference processes, then functional completeness is ensured, but circuit size increases and power consumption rises
Solution Approach 1:
The patent merges the data conversion function and machine learning inference function into a single computation circuit unit. The computation circuit unit includes operators (adders, subtractors, multipliers) that can be dynamically configured to perform both data conversion operations (such as Hadamard transform, discrete Fourier transform, or discrete cosine transform) and machine learning inference operations (such as matrix multiplication and activation functions). This consolidation eliminates the need for separate computation resources, thereby reducing circuit size and power consumption while maintaining functional completeness.
Solution Approach 2:
The computation circuit unit is designed as a universal resource that can perform multiple functions. By configuring the operators within the computation circuit unit according to different computational formulas, the same hardware can execute diverse operations including data conversion transforms and neural network inference computations. This multi-functionality allows a single circuit unit to replace what would traditionally require separate dedicated circuits, thus reducing overall circuit area.
2Reliability
If separate computation resources are used for data conversion and inference processes, then operational reliability is ensured, but power consumption increases
Solution Approach 1:
The patent merges the data conversion function and machine learning inference function into a single computation circuit unit. The computation circuit unit includes operators (adders, subtractors, multipliers) that can be dynamically configured to perform both data conversion operations (such as Hadamard transform, discrete Fourier transform, or discrete cosine transform) and machine learning inference operations (such as matrix multiplication and activation functions). This consolidation eliminates the need for separate computation resources, thereby reducing circuit size and power consumption while maintaining functional completeness.
Solution Approach 2:
The computation circuit unit is designed as a universal resource that can perform multiple functions. By configuring the operators within the computation circuit unit according to different computational formulas, the same hardware can execute diverse operations including data conversion transforms and neural network inference computations. This multi-functionality allows a single circuit unit to replace what would traditionally require separate dedicated circuits, thus reducing overall circuit area.
3Area of stationary object
If unified computation resources are used for both data conversion and inference, then circuit size is reduced and power consumption decreases, but computation speed may be affected
Solution Approach 1:
The patent employs dynamic configuration of the computation circuit unit. The operators within the computation circuit unit can be dynamically reconfigured based on the computational task at hand. For data conversion operations, the operators are configured according to transform-specific computational formulas, and for machine learning inference, they are configured according to neural network computational formulas. This dynamic adaptability allows the unified computation resources to achieve optimal performance for each specific operation type, mitigating potential speed losses from resource sharing.
4Use of energy by stationary object
If unified computation resources are used for both data conversion and inference, then power consumption is reduced, but operational complexity increases
Solution Approach 1:
The patent utilizes parameter changes to manage the complexity of the unified computation circuit unit. By changing the operational parameters and configuration of the operators within the computation circuit unit, the same hardware can adapt to different computational requirements. For example, the computational formulas applied by the operators change depending on whether the task is data conversion or machine learning inference. This parameter-based flexibility allows the system to maintain low power consumption through resource consolidation while managing operational complexity through configurable parameters rather than requiring entirely different hardware for each function.
Data Source
AI summary
A machine learning device includes a data conversion unit configured to convert time series data inputted thereto into frequency feature quantity data, a machine learning inference unit configured to perform machine learning inference based on the frequency feature quantity data, and a computation circuit unit configured to be commonly used by the data conversion unit and the machine learning inference unit.