Tiny ML Face Recognition Models Using Genetic Architecture Search

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for creating face recognition systems in tiny machine learning models are challenging due to the need for skilled resources and the inability of existing techniques to work effectively in resource-limited settings, such as microcontroller-based edge devices, without requiring GPU support and with high search times.

Innovation Solution

A method and system using a genetic algorithm to create a face recognition system in a tiny machine learning model by preprocessing images, creating a master dataset, iteratively training gene structures with reduced datasets, and selecting the best gene structure based on accuracy and model size, ensuring the system focuses on facial features for accurate identification.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If large, full-fledged DNN models are used for face recognition, then feature extraction capability is improved, but device resource requirements and model size increase

Engineering Contradiction:
Improvefeature extraction capabilityVSAvoidmodel size
Core Design Contradiction:
Measurement precisionVSWeight of stationary object

Solution Approach 1:

The patent extracts only the essential components needed for face recognition from complete DNN models. It uses dimensionality reduction to extract dominant features from images and employs a simplified neural network architecture that retains only the critical layers and operations necessary for face recognition, discarding redundant parts of full-fledged DNN models.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies different processing strategies to different parts of the image data. It uses dimensionality reduction specifically for feature extraction from facial regions, applies preprocessing techniques like normalization and augmentation selectively, and uses different neural network configurations for different stages of the recognition pipeline.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If existing NAS methods are used to create customized models, then model accuracy is improved, but computing resource requirements and search time increase

Engineering Contradiction:
Improvemodel accuracyVSAvoidcomputing structure complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the model creation process into distinct phases: dimensionality reduction method selection, parameter optimization through genetic algorithms, and iterative training with reduced datasets. This segmentation allows each phase to be optimized independently, reducing the overall computational complexity compared to end-to-end NAS approaches.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system uses genetic algorithms that automatically evolve and optimize model parameters without requiring manual intervention or complex external optimization infrastructure. The genetic algorithm self-manages the search process, evaluating and selecting parameters based on performance metrics, thereby reducing the need for complex computing structures.

Inventive Principle:
Principle #25Self-service

3Adaptability or versatility

If tiny models with limited layers are used for resource-limited devices, then device compatibility is improved, but feature extraction capability deteriorates

Engineering Contradiction:
Improvedevice compatibilityVSAvoidfeature extraction capability
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent performs dimensionality reduction and feature extraction preprocessing before the data enters the tiny neural network. By pre-processing images to extract dominant features and reduce dimensionality beforehand, the tiny model receives already-processed feature vectors rather than raw images, enabling it to achieve good performance despite having very limited layers and operations.

Inventive Principle:
Principle #10Preliminary action

4Weight of stationary object

If handcrafted tiny models are created for image classification, then model size is reduced, but creation difficulty and required expertise increase

Engineering Contradiction:
Improvemodel sizeVSAvoidmodel creation ease
Core Design Contradiction:
Weight of stationary objectVSEase of manufacture

Solution Approach 1:

The system uses genetic algorithms to automatically optimize model parameters and architecture without requiring manual handcrafting by experts. The genetic algorithm self-manages the model creation process by automatically selecting dimensionality reduction methods, optimizing neural network parameters, and iterating through training cycles, thereby eliminating the need for specialized expertise in model architecture design.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4617920A1Framework for creating a face recognition system in a tiny machine learning model
Publication Date: 2025.09.17 TATA CONSULTANCY SERVICES LTD
  • EP4617920A1 patent drawingFigure 1~2
  • EP4617920A1 patent drawingFigure 3A
  • EP4617920A1 patent drawingFigure 3B

AI summary

Nowadays, intelligent analysis of image data becomes increasingly pivotal in various domains such as robotics, autonomous vehicles, security, and first level disease detection. Existing techniques available for image processing requires large computing structure and suffers from high search time for searching networks connected in complex manner. Present disclosure provides method and system for creating face recognition system in tiny machine leaning (ML) model using genetic algorithm. The system first receives training images which are then pre-processed to obtain facial images. Then, system creates master dataset and reduced datasets. Further, system creates plurality of gene structures using tiny neural network. Each gene structures is then trained with created reduced dataset. Furthermore, performance of each trained gene structure is evaluated based on accuracy score and model size which helps in determining best gene structure among the plurality of gene structures. Finally, tiny new ML model is created based on best gene structure.