Learn more with our Technical Whitepaper

Explore the technical details behind the Reality AI approach to machine learning with signals

  • Why signals require a different approach than other machine learning problems

  • The importance of "features" to effective machine learning

  • Why the FFT probably isn't good enough, and what other options are better

  • The difference between Reality AI and Deep Learning

whitepaper-e-book

Reality AI technology combines machine learning with advanced signal processing math to deliver extremely accurate and computationally efficient classifiers, predictors and anomaly detectors.

Feature Discovery with AI Explore™

Use the AI Explore option in Reality AI Tools™ to create optimized feature sets for machine learning on signal inputs.

> An automatic, data-driven process for dynamically generating features that maximize separation between training classes.

> Generates features of varying computational complexity, and allows you to select the results best suited to your application. Especially useful when creating embedded software for constrained environments.

> 12 patents awarded, 6 patents pending.

Time frequency plot showing features based on FFT

Time-Frequency plot showing features based on FFT

Time frequency plot showing features based on Reality AI

Time-frequency plot showing features based on Reality AI

Our approach uses advanced mathematics, coupled with machine learning, to get better results.

Automated feature discovery that uses specifics of your data, not a “pre-built library”

Automated-feature-discovery

Explainable AI: See which parts of the time-frequency space actually matter

Time-frequency-heatmap2

Heatmap highlights regions that are most important for classification or correlation.  Brighter colors indicate greater importance. Freq structure, as well as both periodic and non-periodic time structure, are clearly visible.

Feature spaces used by Reality AI software include:

> Common transforms on raw data, including logs, powers, derivatives, sign, and more

> Parametric statistical features and peak analysis

> A large variety of spectral features, including power, phase, spectral shape, cepstral, etc.

> Linear and non-linear dimensional reduction operations

> Time-Frequency sparse coding and time pattern analysis

> Binary pattern and texture analysis

The Reality AI feature discovery algorithm searches 10,000s of possible feature spaces automatically, and tunes the results to your data.

What’s different about Reality AI technology?

1. Learns optimized features directly from the data

Features are not pre-determined but are generated from the data using advanced signal processing mathematics guided by machine learning.

2. Links features to optimized machine learning models 

Automatically tunes parameters of machine learning algorithm based on discovered features. This leads to better accuracy and much greater computational efficiency.

3. Fully automated feature and model generation, with explainability

Our AI Explore™ process automatically converges on combinations of features and classifier types with guidance from complexity cost functions to produce models optimized for small MCUs.

4. Direct export of efficient binaries for inexpensive MCUs

Automated generation of embedded code targeted to a range of MCU targets, including Cortex M-series.  Furnishes a compact binary that the user can link and include in their build, containing only functions and data required for model execution - no bulky libraries.