Binarized Neural Networks on Microcontrollers

Explore the potential of Binarized Neural Networks (BNNs) on microcontrollers in this tinyML Talk by Lukas Geiger from Plumerai. Discover how BNNs enable real-time deep learning for complex tasks on microcontrollers by encoding weights and activations using only 1 bit instead of 32 or 8 bits. Learn about the advantages of BNNs, including lower memory requirements and efficient execution, as well as the challenges in training algorithms and custom inference software. Gain insights into Plumerai's integrated approach, built on Keras and TFLite, and their open-source libraries for building, training, and benchmarking BNNs on ARMv8-A architectures. Witness the world's first BNN running live on an ARM Cortex-M4 microcontroller and understand how this technology brings unmatched efficiency to TinyML. Delve into topics such as efficient machine learning stacks, going below 8-bit precision, binarized convolution, training neural networks, and the open-source BNN ecosystem. Explore real-world applications like person detection and visual wake words, and learn about model benchmarking, accuracy, and performance on Cortex-M4 processors. Gain valuable insights into the future of TinyML and the role of Arm in providing the software and hardware foundation for this emerging field.

Intro
Why is TinyML not already everywhere
The road ahead
Efficient ML covers the entire stack
Going below 8-bit precision
Memory reduction for different precisic
Binarized Convolution
Training Neural Networks
Training Binarized Neural Networks
Open Source BNN Ecosystem
Larq Compute Engine
Person Detection / Visual Wake Words
Person Detection on Cortex-M4
Model Benchmark on Cortex-M4
Model Accuracy + Real World Performa
Unit tests for Deep Learning Applicatio
Person Detection Networks
Person Detection using BNNS
What's next?
Arm: The Software and Hardware Foundation for tin
Next tinyML Talks