On-Device Continuous Event-Driven Deep Learning to Avoid Model Drift

Explore on-device continuous event-driven deep learning techniques to prevent model drift in a 21-minute conference talk from tinyML EMEA 2022. Delve into the advantages of event-driven solutions over intent-driven approaches for maintaining up-to-date models in embedded systems. Discover Davinsy, an autonomous machine learning system that continuously learns from real-time data, eliminating the need for user intervention in industrial applications. Examine the impact of changing environments on model performance and the importance of adapting to current conditions. Learn about Virtual Models and the DALE AI engine, which enable polymorphism in local models. Gain insights into voice-control applications using Davinsy Voice, implemented on various embedded platforms. Understand the benefits of on-device learning in reducing data bias and maintaining model accuracy in dynamic environments.

Intro
The law of Big Data.. on the ground!
Ex: speech to intent Al workflow
Learning ground reality
Data centric rather than parametric model generator Event-driven rather than intent-driven
On-device Comprehensive Al Software Product
How to adapt Voice CMD Acceptance Rate to reality?
Competitive Landscape