Intel releases open source AI reference kits
The reference kits include AI model code, end-to-end machine learning pipeline instructions, libraries and Intel oneAPI components for cross-architecture performance. These kits, says the company, enable data scientists and developers to learn how to deploy AI faster and more easily across healthcare, manufacturing, retail and other industries with higher accuracy, better performance and lower total cost of implementation.
“Innovation thrives in an open, democratized environment,” says Wei Li, Ph.D., Intel vice president and general manager of AI and Analytics. “The Intel accelerated open AI software ecosystem including optimized popular frameworks and Intel’s AI tools are built on the foundation of an open, standards-based, unified oneAPI programming model. These reference kits, built with components of Intel’s end-to-end AI software portfolio, will enable millions of developers and data scientists to introduce AI quickly and easily into their applications or boost their existing intelligent solutions.”
The AI reference kits, built in collaboration with Accenture, are designed to accelerate the adoption of AI across industries. They are open source, pre-built AI with meaningful enterprise contexts for both greenfield AI introduction and strategic changes to existing AI solutions.
Four kits are available for download today:
- Utility asset health: As energy consumption continues to grow worldwide, power distribution assets in the field are expected to grow. This predictive analytics model was trained to help utilities deliver higher service reliability. It uses Intel-optimized XGBoost through the Intel oneAPI Data Analytics Library to model the health of utility poles with 34 attributes and more than 10 million data points. Data includes asset age, mechanical properties, geospatial data, inspections, manufacturer, prior repair and maintenance history, and outage records. The predictive asset maintenance model continuously learns as new data, like new pole manufacturer, outages and other changes in condition, are provided.
- Visual quality control: Quality control (QC) is essential in any manufacturing operation. The challenge with computer vision techniques is that they often require heavy graphics compute power during training and frequent retraining as new products are introduced. The AI Visual QC model was trained using Intel AI Analytics Toolkit, including Intel Optimization for PyTorch and Intel Distribution of OpenVINO™ toolkit, both powered by oneAPI to optimize training and inferencing to be 20% and 55% faster, respectively, compared to stock implementation of Accenture visual quality control kit without Intel optimizations for computer vision workloads across CPU, GPU and other accelerator-based architectures. Using computer vision and SqueezeNet classification, the AI Visual QC model used hyperparameter tuning and optimization to detect pharmaceutical pill defects with 95% accuracy.
- Customer chatbot: Conversational chatbots have become a critical service to support initiatives across the enterprise. AI models that support conversational chatbot interactions are massive and highly complex. This reference kit includes deep learning natural language processing models for intent classification and named-entity recognition using BERT and PyTorch. Intel Extension for PyTorch and Intel Distribution of OpenVINO toolkit optimize the model for better performance – 45% faster inferencing compared to stock implementation of Accenture customer chatbot kit without Intel optimizations – across heterogeneous architectures, and allow developers to reuse model development code with minimal code changes for training and inferencing.
- Intelligent document indexing: Enterprises process and analyze millions of documents every year, and many of the semi-structured and unstructured documents are routed manually. AI can automate the processing and categorizing of these documents for faster routing and lower manual labor costs. Using a support vector classification (SVC) model, this kit was optimized with Intel Distribution of Modin and Intel Extension for Scikit-learn powered by oneAPI. These tools improve data pre-processing, training and inferencing times to be 46%, 96% and 60% faster, respectively, compared to stock implementation of Accenture Intelligent document indexing kit without Intel optimizations for reviewing and sorting the documents at 65% accuracy.
The kits can be downloaded for free on the Intel.com AI Reference Kits website. The kits are also available on Github.