一本道无码

一本道无码

Alt Text for Image

December 20, 2024

Learning to Survive

By Jim Blakley, Living Edge Lab Associate Director

Contact Name

A generation or more of American children (including my own) grew up with the computer role playing game, . In this, mostly text-based, game, “pioneer” players attempt to complete a covered wagon journey across the western United States frontier. Staying alive long enough to finish the journey was the primary objective of a game filled with many dangerous obstacles – human and animal threats, disease, weather, and accidents. Players needed to learn what the threats were and how to survive them to reach the lush Willamette Valley in the Oregon Territory. This learning happened mostly through repeated fatal attempts until the last successful try. Of course, learning could and, ideally would, take place during a journey – but only if you survived long enough to gain the knowledge to detect and neutralize the threat and to apply that knowledge to subsequent threats.

This paradigm – what doesn’t kill you makes you smarter – is the key concept behind our work on Survival Critical Machine Learning (SCML). It uses Live Learning from the Hawk project in the context of a threat laden mission environment. Live Learning takes data collected from the environment during a mission to train new ML models for deployment during that mission. Our recent paper, Beyond Federated Learning: Survival-Critical Machine Learning, by Eric Sturzinger and Mahadev Satyanarayanan, introduces a theoretical SCML model centered around autonomous scouts attempting to complete a mission in a hostile environment. The scout has sufficient capacity to train new machine learning models but starts the mission with limited accuracy in detecting the specific threats it will encounter.  It has a limited pool of countermeasures to neutralize threats and a constrained connection to a remote human labeler and cloud AI capability. The SCML model balances the tradeoffs between countermeasure deployment and effectiveness, machine learning accuracy, live learning at the edge, network bandwidth, human labeling speed, threat arrival and lethality, and other factors to maximize the scout’s survival time to end-of-mission.  The figure below shows the use of Hawk with Live Learning for SCML missions. One clear difference between this model and our solitary Oregon Trail players is the existence of a feedback loop with “civilization”. Scouts are able to send small numbers of new candidate target examples for correct classification by a human labeler and cloud AI resources – providing a new ground truth for the scout’s next training round. This process is called semi-supervised learning and relies on a narrowband connection between the scout and the cloud.

Hawk Live Learning for SCML

Hawk Live Learning for SCML

We evaluated the SCML model using the Hawk framework and the publicly available drone surveillance dataset. Our primary question was whether live learning could improve expected mission completion survival over a pre-trained model. We tested different configurations of mission definition, numbers of drones, live learning, threats, and countermeasures. In general, our results show that live learning does show the potential for improved survivability. The SCML model provides an effective tool for evaluating mission characteristics and objectives in specific situations.

Future work will focus on broader validation across other datasets including biological immunity datasets and on more nuanced mission definitions. For example, we plan to test the framework on multi-drone missions with staggered launches where later drones benefit from the “pioneering” drones’ learning.

As with those who made the long journey across the Oregon Trail, it is paramount for an SCML system to learn early.  There is always a risk that the system will perish early in the mission.  But if threats are rare or purposely avoided, the system will never sufficiently learn to detect future threats.  This is the core tradeoff in an SCML system that performs Live Learning. For more information on this ground-breaking approach, please see the paper.

Eric Sturzinger, Mahadev Satyanarayanan, Beyond Federated Learning: Survival-Critical Machine Learning, in 2024 IEEE/ACM Symposium on Edge Computing (SEC), December 2024

Read the Paper