Altamira Data Scientists Present at AI/ML For Multi-Domain Operations Conference
Altamira scientists and engineers are presenting at the SPIE Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications IV Conference in Orlando, FL on 4-7 April 2022! Both talks are on 5 April. Read the abstracts below and come see us at the event!
The Prediction Management Framework: ethical, governable, and interpretable deployment of machine learning
Author: Dan Grahn, Lead Research Engineer, Altamira Technologies Corporation; Dr. Mel Richey, Vice President of Technology
As the military integrates machine learning into evermore critical multi-domain operations and mission sets, especially those at the tactical edge with near real-time decision making, the necessity of a standardized, robust framework for deployment and management is increasing. In this paper we propose a Prediction Management Framework (PMF) to provide comprehensive visibility into the deployment of machine learning models. We begin by exploring different requirements for the framework paired with example use cases. The requirements include aspects such as: deployment-to-retirement model governance, model and prediction explainability, end-user prediction interpretability, prediction integrity, model & prediction revocability, and more. Next, we offer a novel solution for communicating model information and safety based on the well-known FDA Nutrition Facts Label and OSHA Hazard Communication Standards. We extend this solution to individual predictions and provide a method for notifying decision-makers of the bottom-line/up-front information regarding AI/ML processed data. Further, we recommend security measures to ensure data and/or predictions are not modified after processing. Finally, we provide a reference architecture for the Prediction Management Framework. We implement the basic functionality of this system and make recommendations for extending it to a production-ready system.
CNN+LSTM Image Segmentation of Solar Dynamics Observatory Multi-spectral Imagery: encoding known physics into AI/ML inputs to improve space weather forecasting
Author: Dr. James Stockton, Lead Data Scientist, Altamira Technologies Corporation
We present multi-spectral solar image segmentation results for a combined CNN+LSTM that matches the accuracy of existing methods in the Solar Dynamics Observatory (SDO) image processing pipeline, but provides significant computational acceleration during inference. This enables processing of 8+ Pb of previously unlabeled data, a thousand fold increase in available inputs for space weather models. When training the segmentation model we encoded solar phase-dependent prior probabilities for sunspot locations into the input tensors and matched them to SDO multi-spectral imagery. Packaging prior physics knowledge into an ingestible format for AI/ML algorithms is a useful cousin to physics-informed objective functions.See Event