EmbeddedML-Benchmark

Inference Benchmark Results

Explore the performance metrics of various machine learning models optimized for embedded systems.

Welcome to the EmbeddedML-Benchmark results page. Explore the performance metrics of various machine learning models optimized for embedded systems below. The results are based on real-world benchmarks for different ML tasks: Keyword Spotting, Image Classification, Anomaly Detection, and Emotion Detection.

Benchmark Results

Use Case Dataset TFLite Model Size Quality Target
Image Classification CIFAR 10 (32x32) TinyML (9 KB) 85% (Top-1)
Keyword Spotting Speech Communication (49x10) Dense-CNN (53 KB) 89% (Top-1)
Anomaly Detection ToyADMOS (5x128) FC-AutoEncoder (41 KB) 99% (Top-1)
Emotion Detection Multi-Class Emotional Sentences (298x12) LSTM (161 KB) 89% (Top-1)

Performance Comparison

Anomaly Detection Performance

Anomaly Detection Results

Emotion Detection Performance

Emotion Detection Results

Impact of the Benchmark

This benchmarking framework helps to evaluate the deployment of machine learning models in TinyML applications, offering insight into the trade-offs between accuracy, execution speed, and memory efficiency. It will assist researchers and engineers in choosing the most effective models and optimizations for real-time embedded systems.