Image Classification with Fashion-MNIST and CIFAR-10 Khoi Hoang - PowerPoint PPT Presentation

Image Classification with Fashion-MNIST and CIFAR-10 Khoi Hoang California State University, Sacramento

Problem Statement  Image Classification is one of the most fundamental problem in the field of machine learning.  Has many area of applications: Computer Vision • Self-driving car (real time) • Facial recognition, biometrics •  This project will implement various machine learning models, and examine different features extraction techniques to reduce and training time and improve models’ performance.

Original MNIST dataset  Original MNIST dataset is too easy! (99% accuracy)

Fashion-MNIST dataset  Available: https://github.com/zalandoresearch/fashion-mnist  Also available on Keras:

CIFAR-10 dataset  Available: https://www.cs.toronto.edu/~kriz/cifar.html  Also available on Keras:

Approaches and Methodology 1. Models without features extraction  SVM  KNN  Random Forest  Decision Tree  CNN 2. Features Extraction with PCA 3. Features Extraction with Autoencoder

Principal Component Analysis (PCA)  Choosing the right number of components: https://towardsdatascience.com/an-approach-to-choosing-the-number-of- components-in-a-principal-component-analysis-pca-3b9f3d6e73fe

Autoencoder  “Learn” a representation, or encoding, of the data  Attempt to recreate the original image

Conclusion  Features extraction techniques help reduce the training time and increase the performance of the models.  PCA does not work very well with CNN  Autoencoder with SVM achieved the best performance, this can be improved using pretrained model or deeper autoencoder to extract features  Future work: Experimenting with Deep Residual Network (Resnet) o Use transfer learning o Fine tune and increase depth of autoencoder o

Question?