Energy-Efficient Machine Learning Model Training via Threshold Monitoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing complexity of AI/ML models leads to significant energy demands, which contradicts the goal of achieving 'net zero' carbon emissions, as training and inference processes consume substantial energy, posing a challenge for sustainable development and compliance with environmental commitments.
Innovation Solution
A computer-implemented method for training and optimizing ML models that monitors and controls energy consumption by setting target thresholds, using real-time sensor data to adjust training processes, and applying techniques like quantization and weight pruning to reduce energy usage during training and inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If AI/ML models become more complex to meet end-user demands, then solution capability and accuracy are improved, but energy consumption increases exponentially
Solution Approach 1:
The patent applies parameter changes by systematically varying training hyperparameters (learning rate, batch size, optimizer settings) and model architecture parameters (layer depth, neuron count, activation functions) to find the optimal configuration that achieves required accuracy with minimum energy consumption. This resolves the contradiction by transforming the model development process from盲目 increasing complexity to targeted parameter optimization.
Solution Approach 2:
The patent implements partial action by training models for a predetermined number of epochs rather than until complete convergence, and by using partial datasets for training. This approach achieves sufficient model performance without the excessive energy consumption that would result from exhaustive training, directly addressing the energy-accuracy tradeoff.
2Manufacturing precision
If training data is increased and training cycles are extended to improve model performance, then model capability is improved, but training time and energy demand increase
Solution Approach 1:
The patent applies preliminary action through extensive pre-experimentation and hyperparameter tuning before the main training process. By determining optimal learning rates, batch sizes, and architecture configurations in advance through smaller pilot experiments, the system avoids wasting time during full-scale training, thus improving model performance efficiency without proportionally increasing total training time.
Solution Approach 2:
The patent implements feedback mechanisms by continuously monitoring training metrics (loss, accuracy) and using this information to adjust training parameters dynamically. The system uses validation set performance to guide hyperparameter adjustments and early stopping decisions, ensuring model performance improves efficiently without unnecessary extended training cycles.
3Adaptability or versatility
If model complexity increases to deliver sophisticated AI solutions, then solution capability is improved, but carbon emissions increase making net zero commitments difficult to meet
Solution Approach 1:
The patent applies mechanics substitution by replacing brute-force training approaches (increasing model size and training duration) with optimized algorithms and techniques. Specifically, it uses advanced optimizers, efficient data sampling strategies, and intelligent hyperparameter selection to achieve the same model performance with significantly reduced computational workload and associated carbon emissions.
Solution Approach 2:
The patent systematically changes key training parameters including learning rate schedules, batch sizes, and model architecture configurations to find the optimal balance between model capability and energy consumption. This allows sophisticated AI solutions to be delivered with minimized carbon footprint by avoiding unnecessarily complex models and excessive training.
Data Source
Figure 1a
Figure 1b
Figure 1c
AI summary
The present invention relates to energy efficient machine learning (ML) models, in particular to energy efficient ML model generation for energy efficient inferencing. According to an aspect of the invention, there is provided a computer implemented method of operating an ML training system for training a ML model. The method comprising: receiving an energy consumption (EC) target threshold for training; training an ML model based on a training dataset for an Epoch; estimating an EC of the ML training system whilst training said ML model over said Epoch; determining whether an accumulated EC estimate is below the EC target threshold; in response to determining the accumulated EC estimate is less than the EC target threshold, updating the training dataset for a new Epoch and repeating the training, estimating, and generating steps using the updated training dataset with said new Epoch until the ML model is validly trained; terminating training of the ML model in response to determining the accumulated EC meets or exceeds said EC target threshold; outputting said validly trained ML model and associated EC metrics. Further aspects of the invention includes a computer-implemented method of operating a ML training/inference system for performing EC benchmarking of the trained ML model for inferencing and a ML training and inferencing system for implementing said methods.