DeepLearning 101

Theoretical and Practical Aspects fo Deep Learning towards a 1-Day workshop at DAIICT, May 2018

Transcript

1. DEEP LEARNING Making Machines Smarter, since 1980s
2. Parth Agrawal Peddling stories & movies/tvseries since dial-up internet era
3. Remembering Math • Linear Algebra, Partial Derivatives • Matrix Operations, Probability/Stats • Intuition will help you understand the topic but as the saying goes,
4. Neural Networks Perceptron
5. Feed Forward and BackPropagation
6. Neural Networks Multi Layer Perceptron (2 Layer Neural Network) Deep Neural Network
7. CNN • Only use fully connected layers • Only accept vectors as input Localization • Also use sparsely connected layers • Also accept matrices as input MLP
8. Filters : Edge/Shape/Feature Detectors convolved with image to generate feature maps Stride: To specify jumps done by filter Intuition
9. Subsampling and Flatten Max Pooling Flatten (Looks Familiar) Got It Bouncer
10. AutoEncoders • Tries to learn a function hW,b (x) ≈ x • Lossy Compression • Used sometimes for denoising images
11. Putting The Smart In Smartphones Recipe: Object Classification (Online/Offline) Hardware : Your Mobile (Androids only) Magician Name : Convolutional Neural Net Evil Corp : Google (Tensorflow) Architecture : MobileNet Demo : CropX
12. Smarter, Faster, Deeper Recipe: Real Time Object Detection Hardware : GPU enabled devices Magician’s Tool : You Only Look Once (YOLO), Mask RCNN https://pjreddie.com/darknet/yolo/ Magician Names : Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi Architecture : Deep CNN with hybrid modifications
13. Traffic Surveillance System
14. Recurrent Neural Network Neural networks which learn and also memorize Used for generating sequences, time series analysis. Drawing Demo
15. Drawbacks • Difficult to train (too many parameters maintained over time) • Vanishing/Exploding Gradient
16. LSTMs
17. Intuition Solving Vanishing Gradient
18. Theory Practical
19. Generative Adversarial Networks
20. GANs • Inspired by Game Theory • Generator tries to generate fake images that fool the discriminator into thinking that they’re real. • Discriminator will try to distinguish between a real and a generated image. • Both get stronger together until discriminator can’t distinguish between the real and the generated images anymore