mountains-high / retra-cnn

PoC implementation of NN-like classification system. System combines genetic algorithm and convolution neural network. Genetic algorithm automatically generates the best NN-model. And afterwards the generated NN-model solves classification task.

retra-cnn

The NN-like system which succesfully solves classification task. Classification performance and quality was tested using CIFAR-10 image database, which contains 100 thousand of images devided by 10 classes.

Major results:

• ability to automate NN-models creation using genetic algorithm;
• classification quality is in range between 73 - 75 %;
• model capable to solve localization task using a CAM-model (class activation maps model). Quality is about 50%.

Live result:

Classification and localization in the single model:

Evolution process of the genetic algorithm: