Domain-Aware Model Training as a Service for Use-Inspired Models

Use-inspired artificial intelligence (AI) tailors deep-learning models for image processing tasks in targeted scientific domains. These use-inspired models meet domain requirements for accuracy while parsimoniously using compact and efficient model architectures needed for inference in the field. However, before settling upon a model, domain experts repeatedly train and test models over a wide range of hyperparameters, contextual settings, and data configurations, making model training bespoke, time-consuming, and costly. Model-training-as-a-Service (MTaaS), i.e., cloud services designed for generic training workloads, can reduce training costs, but domain-aware designs and runtime adaptations could yield further reductions. This paper characterizes the potential for domain-aware design and runtime adaptation for MTaaS in digital agriculture. First, we studied the time to train models for 10 use-inspired agricultural datasets using pre-trained model weights derived from other agricultural datasets versus pre-trained weights derived from ImageNet, a widely used benchmark. Using agricultural datasets sped training time by up to 2X for some datasets, but provided modest speedups (<1.07) in the common case; Choosing the right dataset is critical. Next, we present an approach to predict training time given domain-aware pre-trained weights. Our predictions are strongly correlated with training time (r=0.93). Finally, we studied the use of domain-aware pre-trained weights in a MTaaS under Poisson and bursty arrival patterns for training tasks. Under bursty arrivals and tight memory constraints, domain-aware MTaaS reduced training time by 2.8 X and 12.2 X compared to model training using pre-trained ImageNet weights and from scratch, respectively.