Feature learning for image classification Kai Yu and Andrew Ng Andrew Ng Computer vision is hard Andrew Ng Machine learning and feature representations pixel 1 Learning algorithm Input pixel 2 pixel 2 Input space Motorbikes “Non”-Motorbikes pixel 1 Andrew Ng Machine learning and feature representations handle wheel Feature representation Learning algorithm Input Feature space pixel 2 “handle” Input space Motorbikes “Non”-Motorbikes pixel 1 “wheel” Andrew Ng How is computer perception done? Object detection Image Low-level vision features Recognition Audio classification Audio Low-level audio features Speaker identification Helicopter control Helicopter Low-level state features Action Andrew Ng Learning representations Sensor Feature Representation Learning algorithm Andrew Ng Computer vision features SIFT HoG Textons Spin image RIFT GLOH Andrew Ng Audio features Spectrogram MFCC Problems of hand-tuned features 1. Needs expert knowledge 2. Time-consuming and expensive Flux Rolloff ZCR to other domains 3. Does not generalize Andrew Ng Computer vision is more than pictures Images Video Thermal Infrared 3d range scans (flash lidar) Camera array Audio Can we automatically learn good feature representations? Andrew Ng Learning representations Sensor Feature Representation Learning algorithm Andrew Ng Sensor representation in the brain Seeing with your tongue Human echolocation (sonar) Auditory cortex learns to see. Auditory Cortex [BrainPort; Martinez et al; RoeAndrew et al.]Ng Unsupervised feature learning Find a better way to represent images than pixels. Andrew Ng The goal of Unsupervised Feature Learning Unlabeled images Learning algorithm Feature representation Andrew Ng Tutorial outline 1. Current methods. 2. Sparse coding for feature learning. — Break — 3. Advanced classification. 4. Advanced concepts & deep learning. Andrew Ng