Building efficient and scalable end-to-end AI applications is complex and often comes with a steep learning curve due to the many tools, libraries, and optimization methods required.

This session introduces a solution: five turnkey, downloadable AI reference kits tailor-made to solve business problems across a variety of industries, delivering higher accuracy and better performance while decreasing development cycles. Each is built with Intel-designed AI workflows and optimized tools, frameworks, and libraries.

This video shows:

• An overview of the use cases: predictive asset maintenance, credit card fraud detection, disease prediction, correspondence indexing, and anomaly detection.
  
• How to use the kits to jumpstart development of your AI applications, including customizing them for your specific needs.
• How to run them with Docker* containers, bare metal, or Argo Workflows on Kubernetes* using the Helm* package manager.

Skill level: Novice

[00:13] Introduction of the speakers.

[02:00] Introduction to AI software development.

[07:07] Multimodal disease prediction

• [07:24] The AI reference kit builds a tool to implement transfer learning on vision models and to support disease prediction and anomaly detection.
• [09:54] The multimodal approach combines the results of image and natural language processing (NLP) models to give better accuracy.
• [10:36] Demo of multimodal disease prediction for breast cancer.

[15: 44] Credit card fraud detection

• [15:51] Distributed inference and training is required to detect credit card fraud quickly and to minimize losses.
• [16:17] The reference kit uses custom graph neural networks (GNN) to analyze user behavior patterns and significantly increase fraud classification accuracy.
• [17:17] How to define the input data and customize the configuration files to make the use case best fit a workload.

[21:30] Anomaly detection: visual quality inspection

• [21:31] Address manufacturing defects that can cause significant liability to companies and product consumers.
• [22:53] The process starts with a model that was pretrained on ImageNet*.
• [24:23] For inference, test images are used to extract the most important features.
• [24:44] The expected output includes the accuracy and the area under the receiver operator characteristic curve (AUROC).
• [24:56] An overview of software tools used.
  
  

[26:28] Document Automation

• [26:32] Enterprises accumulate large volumes of documents scanned in image formats. It is a time-consuming challenge to index, search, and gain insight into the documents.
• [27:34] Queries against the document images require multimodal AI that understands images and languages, and that multimodal process requires deep expertise in machine learning.
• [28:07] This reference kit implements and demonstrates an end-to-end solution to build an AI-augmented multimodal semantic search system for document images.
• [28:41] This architecture of the reference kit consists of three pipelines.
  

[33:58] Predictive asset maintenance

• [34:11] This kit implements a time-series analysis using the BigDL Time-Series Toolkit.
• [35:35] This example shows how the TS Dataset (an API for processing historical data) is used to assess elevator maintenance.
• [36:36] Learn how to detect pattern and trend anomalies.
• [38:43] Demo of document automation.

[45:05] Five key takeaways.

[49:37] Q&A[00:13] Introduction of the speakers.

[02:00] Introduction to AI software development.

[07:07] Multimodal disease prediction

• [07:24] The AI reference kit builds a tool to implement transfer learning on vision models and to support disease prediction and anomaly detection.
• [09:54] The multimodal approach combines the results of image and natural language processing (NLP) models to give better accuracy.
• [10:36] Demo of multimodal disease prediction for breast cancer.

[15: 44] Credit card fraud detection

• [15:51] Distributed inference and training is required to detect credit card fraud quickly and to minimize losses.
• [16:17] The reference kit uses custom graph neural networks (GNN) to analyze user behavior patterns and significantly increase fraud classification accuracy.
• [17:17] How to define the input data and customize the configuration files to make the use case best fit a workload.

[21:30] Anomaly detection: visual quality inspection

• [21:31] Address manufacturing defects that can cause significant liability to companies and product consumers.
• [22:53] The process starts with a model that was pretrained on ImageNet*.
• [24:23] For inference, test images are used to extract the most important features.
• [24:44] The expected output includes the accuracy and the area under the receiver operator characteristic curve (AUROC).
• [24:56] An overview of software tools used.
  
  

[26:28] Document Automation

• [26:32] Enterprises accumulate large volumes of documents scanned in image formats. It is a time-consuming challenge to index, search, and gain insight into the documents.
• [27:34] Queries against the document images require multimodal AI that understands images and languages, and that multimodal process requires deep expertise in machine learning.
• [28:07] This reference kit implements and demonstrates an end-to-end solution to build an AI-augmented multimodal semantic search system for document images.
• [28:41] This architecture of the reference kit consists of three pipelines.
  

[33:58] Predictive asset maintenance

• [34:11] This kit implements a time-series analysis using the BigDL Time-Series Toolkit.
• [35:35] This example shows how the TS Dataset (an API for processing historical data) is used to assess elevator maintenance.
• [36:36] Learn how to detect pattern and trend anomalies.
• [38:43] Demo of document automation.

[45:05] Five key takeaways.

[49:37] Q&A
 

Many Intel®-optimized AI libraries and frameworks showcased in this session are downloadable as part of the AI Tools. They are also available as stand-alone products:

• Intel® Neural Compressor
• PyTorch* Optimizations from Intel
• TensorFlow* Optimizations from Intel
• Intel® Extension for Scikit-learn*
• Modin*

• The AI Reference Kits Library offers overviews of and access to all 34 kits.
• Review and download an extensive collection of ready-to-use code samples to develop, optimize, and offload multiarchitecture applications.