Neural network model for common CIFAR10 benchmark problem in Machine Learning
About CIFAR10 (University of Toronto - Computer Science)
Benchmarked Results of Other Models
The obtained model achieves 91.49% accuracy, which can be verified by cifar_verify.py script
The resulting model is a Deep Convolutional Neural Network. It has a total of 12 convolution layers and 2 regular layers. Convolutional layers are split in 4 groups by 3 layers in each group, with batch normalization after each layer. At the end of each group there's a max pooling layer and a dropout layer.
To improve training the dataset was slightly augmented. A horizontally flipped copy of each image was added, as well as two rotated (-3 and +3 degrees) copies of each image were added. Resulting training dataset reached 200 000 images.
Training of this model was performed on two Nvidia GeForce GTX 1070 cards, with Intel Core i5 8600K, and 16GB RAM.