Deep learning cifar10

Esse trabalho é resultado do uso de técnicas de deep learning utilizando a base de dados CIFAR10 para apresentação ao Curso de Ciência de Dados e Analytcs da Universidade Politécnica do Pernambuco

---

This work is the result of the use of deep learning techniques using the CIFAR10 database for presentation to the Data Science and Analytcs Course at the Polytechnic University of Pernambuco.

Techniques

• Simple VGG
• Data Augmentation (random flip, random rotation)
• Otimizer Adam and Adamax
• Growing Dropout
• Batch Normalization

Models and results

Model	acc	acc_validation
M1	30.52	36.04
M2E3	27.14	31.54
M3E3Max	97.09	80.07
M3E3Max + Drop V1	91.88	82.58
M3E3Max + Drop V2	85.71	84.18
M3E3Max + augmentation V1	79.62	76.80
M3E3Max + augmentation V2	85.65	81.30
M3E3Max + Drop + augmentation 50 épocas	76.21	79.24
M3E3Max + Drop + augmentation 100 épocas	80.35	82.74
M3E3Max + Drop + augmentation 100 épocas V2	83.17	84.18
M4E4Max + Drop + augmentation 100 épocas	84.77	83.49
M4E3Adam - Drop e augmentation 100 épocas	77.79	79.91
M4E4Max + Drop + augmentation + BatchNorm 100 épocas	89.07	87.42