Abstract
In this study, we introduce a novel platform, Resource-Aware AutoML (RA-AutoML), which enables flexible and generalized algorithms to build machine learning models subjected to multiple objectives, as well as resource and hardware constraints. RA-AutoML intelligently conducts Hyper-Parameter Search (HPS) and Neural Architecture Search (NAS) to build models optimizing predefined objectives. At its core, RA-AutoML relies on the MOBOGA search engine, which combines modified constraint-aware Bayesian Optimization and Genetic Algorithms to construct Pareto optimal candidates. Experiments on the CIFAR-10 dataset demonstrate high accuracy while strictly adhering to resource constraints such as model size.
Introduction
Deep Learning models are increasingly deployed on resource-constrained hardware like mobile phones, FPGAs, and IoT devices. While traditional Automated Machine Learning (AutoML) focuses primarily on maximizing accuracy, real-world engineering often requires balancing competing goals—such as maintaining a specific memory footprint or inference latency. RA-AutoML addresses this gap by providing a framework that treats hardware limitations as first-class citizens during the architecture search process.
Methods: The RA-AutoML Framework
The platform utilizes a multi-phase workflow powered by the MOBOGA (Multi-Objective Bayesian Optimization and Genetic Algorithm) search engine.
1. Problem Setup
The process begins with the user defining the search space $\Omega$, which includes architecture parameters (e.g., layer sizes) and hyperparameters (e.g., learning rate). Crucially, the user also specifies multiple objectives $Q$ and hard resource constraints $C$.
2. Exploration and Exploitation
The core search process is divided into two main stages:
- Exploration: This phase uses Constraint-Aware Bayesian Optimization (BO) to suggest new candidates. A modified "Expected Improvement" function penalizes regions that violate hard constraints. These candidates are then processed by the NSGA-II Genetic Algorithm to identify "non-dominated" models that offer the best trade-offs across all objectives.
- Exploitation: Once the search concludes, the system presents a Pareto front of optimal models. Users can then select the "best" model using the TOPSIS algorithm or manual inspection, ensuring the final choice aligns with specific deployment requirements.
Experiments and Results
The researchers evaluated RA-AutoML on the CIFAR-10 dataset.
- Transfer Learning: Using VGG-based models, the system identified a configuration with a 7.33% Top-5 error while strictly staying under an 80MB memory limit.
- Neural Architecture Search: When building from scratch with ResNet blocks, the platform produced a model with an 8.73% Top-5 error using only 1.155 MB of memory and 4.034 million FLOPs.
Compared to standard ResNet18 (10.76% error) and ResNet34 (8.74% error), RA-AutoML models achieved competitive or superior accuracy with significantly reduced resource footprints.
Key Insights
- Pareto Front Transparency: Providing a map of trade-offs allows for informed business decisions, such as trading a small amount of accuracy for significant power savings.
- Hybrid Optimization: Combining the global search capabilities of BO with the multi-objective efficiency of GA creates a robust search engine for complex hardware-constrained environments.
Future Work
Future developments will focus on scaling to larger datasets like ImageNet and integrating hardware-in-the-loop profiling to measure real-time latency and power consumption directly on target devices, moving beyond theoretical proxy metrics.
Summary updated