Glosssary and References
This glossary provides a standardized reference for the multi-disciplinary concepts—spanning data science, silicon architecture, and digital signal processing—integrated within the eFabric™ ecosystem.
A. Architectural & Hardware Terms
-
Always-On: A continuous monitoring state where a device remains perceptive to specific environmental triggers (acoustic, kinetic, or visual) while the high-power application processor remains in deep sleep.
-
At-Memory Compute: A hardware paradigm that integrates memory and logic gates into a single physical fabric, eliminating the "energy tax" of data movement found in traditional von Neumann architectures.
-
Deterministic Inference: A system property ensuring that neural network processing occurs within a fixed, predictable time window, which is critical for real-time safety and industrial triggers. -eFabric™: A unified machine learning factory and software-to-silicon platform designed to manage the entire ML lifecycle—from raw data ingestion to hardware-native deployment.
-
Neural Decision Processor (NDP): A class of ultra-low-power semiconductors specifically architected to run deep neural networks natively at the microwatt scale.
-
Silicon-Native Execution: The process of mapping a neural network’s computational graph directly onto hardware logic gates rather than running it through a software interpreter or CPU.
-
TML120: A production-ready System-on-Module (SOM) that integrates an NDP with on-board flash memory and digital interfaces, providing a validated "Edge Brain" for rapid prototyping and deployment.
B. Mathematical & Machine Learning Terms
- Activation Function: A mathematical function (such as ReLU or Sigmoid) applied to a neuron's output to determine if—and with what strength—a signal is passed forward through the network.
- Backpropagation: The fundamental algorithm used during the training phase to adjust weights and biases based on the calculated error (loss), enabling the model to "learn" patterns.
- Confidence Threshold (T): The software-defined gate (typically between 0 and 1) that an inference probability score must meet or exceed to trigger a system action.
- Convolutional Neural Network (CNN): A deep learning architecture optimized for spatial and temporal pattern recognition, such as identifying the specific "shape" of a sound event or a physical vibration.
- Gated Recurrent Unit (GRU): A specialized type of Recurrent Neural Network (RNN) that uses internal gates to remember or forget information over time, ideal for sequential events like gestures or baby cries. -Inference: The "live" execution phase where a trained and quantized model processes new, unseen data locally on the chip to produce a classification or trigger. -Quantization-Aware Training (QAT): A training methodology that simulates the effects of low-bit precision during the learning phase, allowing the model to adapt its weights for maximum accuracy on 8-bit hardware. -Weights and Biases: The internal parameters of a neural network; weights determine the strength of connections between neurons, while biases allow the model to shift detection thresholds.
C. Data Engineering & Signal Processing Terms
- Aliasing: A signal distortion occurring when the sampling rate is too low to capture high-frequency components, resulting in "phantom" data that can cause false triggers.
- Fast Fourier Transform (FFT): A mathematical algorithm that decomposes a time-domain signal (a waveform) into its constituent frequencies, enabling spectral analysis.
- Hamming Window: A function applied to raw data frames to smooth transitions and minimize "Spectral Leakage" during frequency analysis.
- Log-Mel Spectrogram: A 2-D visual representation of sound that maps frequencies to the human-centric Mel scale and applies logarithmic scaling to intensity, serving as the "vision of sound" for audio models. -Nyquist Rate: The minimum sampling frequency required to accurately represent a signal, defined as twice the highest frequency present in the input.
- Signal-to-Noise Ratio (SNR): The measurement of the strength of the target signal (e.g., a keyword) relative to background interference; 15dB is the professional benchmark for clean training data.
- Weight Pruning: An optimization technique where unnecessary or "near-zero" neural connections are removed to reduce the model's memory footprint without sacrificing performance.
D. Validation & Reliability Metrics
- Equal Error Rate (EER): The specific balance point where the False Acceptance Rate and False Rejection Rate are identical, used as a universal benchmark for model robustness.
- False Acceptance Rate (FAR): A measure of "over-sensitivity," quantifying how often the system incorrectly identifies noise or non-target events as a match (False Positives).
- False Rejection Rate (FRR): A measure of "under-sensitivity," quantifying how often the system fails to detect a genuine target event (False Negatives).
- F1-Score: The harmonic mean of Precision and Recall, providing a single metric to evaluate models deployed in "imbalanced" real-world environments.