Machine Learning at the Extreme Edge of Always-on Intelligent Sensor Networks

Explore a conference talk on implementing machine learning at the extreme edge of always-on intelligent sensor networks. Delve into the advantages of detecting events at end nodes rather than gateways or cloud systems, including reduced latency, enhanced privacy, and lower bandwidth requirements. Examine the challenges of implementing Deep Neural Networks (DNNs) on resource-constrained end nodes and discover optimization techniques across system, algorithm, architecture, circuit, and process technology levels. Learn about various sensing applications, such as voice activity detection, object recognition, and anomaly detection, utilizing different sensing modalities. Gain insights into quantization, model augmentation, custom data paths, and in-memory compute strategies to achieve significant reductions in area-power figures of merit for IoT sensor nodes.

Introduction
IoT Sensors
Alwayson Detection
System Level Optimization
Constraints
Network Complexity
Weight Quantization
Model Augmentation
Hardware Options
Quantized Networks
Custom Data Paths
Binary Convolution
Summary
InMemory Compute
Quantization and Training
Analog Computation
Sponsors