1.3. Terminology

This glossary defines key terms and abbreviations used throughout the Tiny ML Tensorlab documentation.

1.3.1. General ML Terms

Autoencoder: A neural network architecture that learns to compress input data into a lower-dimensional representation (encoding) and then reconstruct it (decoding). Used for anomaly detection in Tiny ML Tensorlab.
Feature Extraction: The process of transforming raw input data (e.g., time-series signals) into a more meaningful representation for the neural network. Examples include FFT, wavelets, and binning.
Inference: The process of using a trained model to make predictions on new data.
MACs: Multiply-Accumulate Operations. A measure of computational complexity. One MAC = one multiplication followed by one addition.
ONNX: Open Neural Network Exchange. A standard format for representing machine learning models, used as the intermediate format before compilation.
PTQ: Post-Training Quantization. Quantization applied after training is complete. Faster than QAT but may result in lower accuracy.
QAT: Quantization-Aware Training. A technique where quantization effects are simulated during training, resulting in better accuracy compared to post-training quantization.
Quantization: The process of reducing the precision of model weights and activations from floating-point (32-bit) to lower bit-widths (2/4/8-bit). This reduces model size and improves inference speed on MCUs.

GoF Test: Goodness of Fit Test. A visualization tool in Tiny ML Tensorlab that helps evaluate whether a dataset is suitable for classification by plotting class separability using PCA and t-SNE.
ModelMaker: The tinyml-modelmaker repository that orchestrates the training and compilation workflow.
ModelZoo: The tinyml-modelzoo repository containing pre-defined model architectures and example configurations. The primary entry point for users.
NAS: Neural Architecture Search. An automated technique for discovering optimal neural network architectures. Available in Tiny ML Tensorlab for time series classification.
NNC: Neural Network Compiler. TI’s compiler that converts ONNX models into optimized code for TI MCUs.
NPU: Neural Processing Unit. General term for hardware acceleration of neural network operations. TINPU is TI’s implementation.
TINPU: TI Neural Processing Unit. Hardware accelerator for neural network inference present in select TI MCUs (F28P55, AM13E2, MSPM0G5187).
TinyVerse: The tinyml-tinyverse repository containing core training scripts, dataset loaders, and utilities.

C2000: A family of TI 32-bit real-time microcontrollers optimized for control applications. Includes devices like F28P55, F28P65, F2837.
C2000Ware: TI’s software development kit (SDK) for C2000 microcontrollers.
CCS: Code Composer Studio. TI’s integrated development environment (IDE) for programming TI microcontrollers.
LaunchPad: TI’s low-cost development kit for evaluating MCUs.
MSPM0: A family of TI ultra-low-power Arm Cortex-M0+ microcontrollers.
MSPM33C: A family of TI Arm Cortex-M33 microcontrollers with TrustZone security.

feature_extraction_name: Configuration parameter specifying a preset feature extraction pipeline. Example: Generic_1024Input_FFTBIN_64Feature_8Frame.
model_name: Configuration parameter specifying which model architecture to use. Example: CLS_1k_NPU, REGR_500_NPU.
target_device: Configuration parameter specifying the deployment MCU. Examples: F28P55, MSPM0G3507, AM263.
task_type: Configuration parameter specifying the ML task. Options include: generic_timeseries_classification, generic_timeseries_regression, generic_timeseries_forecasting, generic_timeseries_anomalydetection, image_classification.

BYOD: Bring Your Own Data. The practice of using your own dataset with Tiny ML Tensorlab rather than a built-in example dataset.
BYOM: Bring Your Own Model. The practice of adding custom model architectures to the ModelZoo or compiling external ONNX models.
frame_size: The number of samples in each data frame (window) used for training.
stride_size: The overlap between consecutive frames, expressed as a fraction. stride_size: 0.5 means 50% overlap.
variables: The number of input channels or features in the dataset. For multi-axis sensor data, this equals the number of axes.

Model names in Tiny ML Tensorlab often include size indicators:

Suffix	Meaning	Parameter Count
`_100`	~100 parameters	Ultra-minimal
`_1k` or `_1K`	~1,000 parameters	Small
`_4k` or `_4K`	~4,000 parameters	Medium
`_13k` or `_13K`	~13,000 parameters	Large
`_NPU`	NPU-optimized	Architecture follows NPU constraints

Abbreviation	Full Form
AI	Artificial Intelligence
CNN	Convolutional Neural Network
DSP	Digital Signal Processor
FFT	Fast Fourier Transform
GPU	Graphics Processing Unit
LSTM	Long Short-Term Memory (neural network)
MCU	Microcontroller Unit
ML	Machine Learning
MSE	Mean Squared Error
NN	Neural Network
PMSM	Permanent Magnet Synchronous Motor
SMAPE	Symmetric Mean Absolute Percentage Error
SRAM	Static Random Access Memory