A large dataset is required for successful machine learning training. A recommended minimum starting place is thousands of quality examples for each classification class. The quality and diversity of the data is critical. TI offers reference designs with automatic class labeling, and software to analyze dataset quality to assist engineers in evaluating edge AI applications.

High-quality data:

• Sample rate: Sampled well above the Nyquist frequency of the signals of interest
• Steady state: Labeled frames can't be transitioning between labeled classes highest resolution available

Data collection: The final AI model can be trained on data that is representative of the final circuit design and real-world application. Lab collected data can be used as a starting place to tune the full signal chain including the model and analog circuit design. In this example we collected three different classes of periodic signal data: sine, square and sawtooth. They are saved into three separate folders