MLSUS02-BP01 Consider AI services and pre-trained models
Using AI services and pre-trained models can significantly reduce the resources needed for machine learning workloads, enabling you to quickly implement AI capabilities without developing custom models from scratch.
Desired outcome: You identify opportunities to use managed AI services or pre-trained models instead of building custom models, reducing the environmental impact of training and deploying ML solutions while accelerating time to market. By using existing capabilities through APIs or fine-tuning pre-trained models, you minimize computational resources required for data preparation, model training, and deployment.
Common anti-patterns:
-
Building custom models from scratch when suitable managed services already exist.
-
Collecting and processing large datasets unnecessarily when pre-trained models could be utilized.
-
Ignoring transfer learning opportunities that could reduce training time and computational resources.
-
Overlooking model optimization techniques that could reduce inference resource requirements.
Benefits of establishing this best practice:
-
Reduced environmental impact through decreased computational resource usage.
-
Lower operational costs for ML development and deployment.
-
Faster time-to-market for ML-powered solutions.
-
Access to state-of-the-art models without specialized ML expertise.
Level of risk exposed if this best practice is not established: Medium
Implementation guidance
When developing machine learning solutions, evaluate whether you truly need to build a custom model or if existing services and pre-trained models can meet your requirements. Many common use cases like image recognition, natural language processing, or recommendation systems can use pre-built capabilities available through APIs. These managed services handle the underlying infrastructure, data processing, and model maintenance, reducing both environmental impact and operational overhead.
If fully managed services don't meet your specific requirements, consider starting with pre-trained models that you can fine-tune with your data. This approach, known as transfer learning, allows you to benefit from models that have already undergone resource-intensive training on large datasets. By fine-tuning, you can achieve similar or better performance than training from scratch while using significantly fewer computational resources.
For example, instead of training a computer vision model from scratch, you could use a pre-trained image recognition model from Amazon Rekognition or fine-tune a model available through SageMaker AI JumpStart with your specific images. This approach reduces the carbon footprint associated with extensive model training while delivering high-quality results.
Implementation steps
-
Assess your ML use case requirements. Before developing an ML solution, clearly define your requirements and success criteria. Understand the specific problem you're trying to solve, the data you have available, and the performance metrics that matter for your application.
-
Explore AWS AI services. AWS AI services
provide ready-to-use capabilities through APIs for common ML tasks. These services include Amazon Rekognition for image and video analysis, Amazon Comprehend for natural language processing, Amazon Forecast for time-series forecasting, and Amazon Personalize for personalization and recommendations. -
Evaluate foundation models in Amazon Bedrock. Amazon Bedrock
provides serverless access to leading foundation models from AI companies like Anthropic, Cohere, AI21 Labs, and Amazon's own Nova models through a single API. These models can handle tasks like text generation, summarization, chatbots, and content creation without requiring model training. -
Explore pre-trained models from AWS Marketplace. AWS Marketplace
offers over 1,400 ML-related assets that you can subscribe to, including pre-trained models for various industry-specific use cases that can be deployed with minimal configuration. -
Leverage SageMaker AI JumpStart. SageMaker AI JumpStart provides pre-trained, open-source models for a wide range of problem types to get started with machine learning. You can incrementally train and fine-tune these models before deployment, reducing the computational resources needed compared to training from scratch.
-
Consider Hugging Face models. Hugging Face with Amazon SageMaker AI enables you to use thousands of pre-trained transformer models for NLP, computer vision, and audio tasks. These models can be fine-tuned with your specific data to achieve high-quality results with minimal training.
-
Implement efficient fine-tuning techniques. When customizing pre-trained models, use efficient fine-tuning methods like parameter-efficient fine-tuning (PEFT), low-rank adaptation (LoRA), or quantization to minimize computational resources while maintaining performance.
-
Monitor resource usage and optimize. After deploying your solution, continuously monitor its resource consumption and performance. Look for opportunities to optimize through model compression, quantization, or pruning to further reduce computational requirements.
-
Leverage expanded pre-trained model libraries. Use the expanded SageMaker AI JumpStart catalog which now includes broader selection of pre-trained models and industry-specific solutions, reducing the need for custom model development and associated computational resources.
-
For generative AI workloads, implement retrieval-augmented generation (RAG). For generative AI applications requiring domain-specific knowledge, consider using retrieval-augmented generation
with SageMaker AI JumpStart foundation models instead of fine-tuning large models, which can significantly reduce computational resources while improving accuracy.
Resources
Related documents:
