1.3. Terminology

Key terms and abbreviations used in the Tiny ML ModelZoo documentation.

1.3.1. General ML Terms

Autoencoder: A network that compresses input data and reconstructs it. Used for anomaly detection.
Early stopping: Ending training when validation accuracy has not improved for a number of epochs (early_stopping_patience).
Feature Extraction: Transforming raw input data (for example time-series signals) into a representation for the neural network. Examples: FFT, wavelets, binning, MFCC.
Golden vectors: Real input samples with the outputs the model should produce, written to test_vector.c so a device implementation can be checked against the PC.
Inference: Using a trained model to make predictions on new data.
MACs: Multiply-Accumulate Operations. A measure of computational complexity.
Mixed-precision quantization: Using different bit-widths for different layers, chosen automatically (Hessian-aware, HAWQ).
ONNX: Open Neural Network Exchange. The model format used between training and compilation.
PTQ: Post-Training Quantization. Quantization applied after training. Faster than QAT but may lose more accuracy.
QAT: Quantization-Aware Training. Quantization effects are simulated during training, which usually gives better accuracy than post-training quantization.
Quantization: Reducing the precision of model weights and activations from 32-bit floating point to lower bit-widths. This reduces model size and speeds up inference on MCUs.

BYOD: Bring Your Own Data. Using your own dataset instead of a built-in example dataset.
BYOM: Bring Your Own Model. Adding your own model architecture to the model zoo or compiling an external ONNX model.
ModelMaker: The tinyml_modelmaker package that orchestrates training and compilation.
ModelZoo: The tinyml-modelzoo repository and the tinyml_modelzoo package: model definitions, example configurations and run scripts. The entry point for users.
NAS: Neural Architecture Search. Automated search for network architectures.
NNC: Neural Network Compiler (ti_mcu_nnc). TI’s compiler that turns ONNX models into optimized code for TI MCUs.
NPU: Neural Processing Unit. General term; TINPU is TI’s implementation.
Tensorlab: tinyml-tensorlab, the name of the toolchain up to release 1.4.0. It is deprecated as the entry point and replaced by tinyml-modelzoo.
TINPU: TI Neural Processing Unit (TinyEngine NPU). Hardware accelerator for neural network inference in select TI MCUs (F28P55, MSPM0G5187, AM13E2).
TinyVerse: The tinyml_tinyverse package with training scripts, dataset loaders and utilities.

AM13: TI’s Arm Cortex-M33 MCUs. AM13E2 has an NPU.
AM26x: TI’s Arm Cortex-R5 MCUs (AM263, AM263P, AM261).
C2000: TI’s family of 32-bit real-time MCUs for control applications (C28 DSP: F28P55, F28P65, F2837, … and C29: F29H85, F29P58, F29P32).
C2000Ware: TI’s software development kit for C2000 MCUs.
CCS: Code Composer Studio, TI’s IDE for MCU development.
LaunchPad: TI’s low-cost development kit for evaluating MCUs.
MSPM0: TI’s ultra-low-power Arm Cortex-M0+ MCUs.
MSPM33C: TI’s Arm Cortex-M33 MCUs.

feature_extraction_name: A preset feature extraction pipeline, for example Generic_1024Input_FFTBIN_64Feature_8Frame.
frame_size: The number of samples in each data frame (window).
model_name: The model architecture, for example CLS_1k_NPU.
projects_path: common.projects_path: where run output is written (default ./data/projects).
stride_size: The step between consecutive frames, as a fraction of the frame size.
target_device: The deployment MCU, for example F28P55, MSPM0G3507 or AM263.
task_type: The ML task, for example generic_timeseries_classification, generic_timeseries_regression, generic_timeseries_forecasting, generic_timeseries_anomalydetection, arc_fault, audio_classification, image_classification or radar_classification.
variables: The number of input channels in the dataset.

Model names often include a size indicator, for example CLS_1k_NPU or CLS_4k_NPU.

Name part	Meaning
`_100`, `_1k`, `_4k`, `_13k`	Approximate parameter count (100, 1,000, 4,000, 13,000)
`_NPU`	Architecture follows the NPU constraints

See Model Zoo Reference for the exact models in 1.5.0.

Abbreviation	Full form
CNN	Convolutional Neural Network
DSP	Digital Signal Processor
FFT	Fast Fourier Transform
LSTM	Long Short-Term Memory
MCU	Microcontroller Unit
MFCC	Mel-Frequency Cepstral Coefficients
ODL	On-Device Learning
PMSM	Permanent Magnet Synchronous Motor
SMAPE	Symmetric Mean Absolute Percentage Error
SRAM	Static Random Access Memory