Day #198 - Semantic deep learning: segmentation and regression - Jorge Cardoso - DeepA2Z

Great Talk and Good Summary

Key Lessons

• Map Image from one space to another space
• Vector mapped to output space
• Regression - Continuous Value
• All are types of Regression Problems

Deep Learning VS ML

• Most work in ML is creating features
• Mapping function feature space to classification

DL

• DL learns features from Data
• Key DL Components
• Architecture - Loss Function - Optimizer (part of learning process)

Building Blocks

• Activation Functions (Relu, Elu, SRelu, PreLU
• Convolutional Layers (Dimensionality vs complexity)
• Aggregation Layers (Pooling / Convolutions)

• Convolution Example
• Input X Kernel = Output
• Kernels are matrices

• Backprop apply gradient descent
• Weighted matrix multiplication is convolution
• Standard Convolution 3x3 kernel
• Dialted Convolution - Spacing between elements of kernel (Bring Context and Relationships)
• Strided Convolution (Apply convolution every X number of Pixels)
• Unpooling - Upsampling Image
• Residual Connections (apply a little + plus)
• Drop Out, Batch Norm (Form of Regularization)
• Classification (K Classes)
• Scale -> Depth 
• Conv + Relu + Pooling
• Extract Features that scale
• Segmentation (Input -> Output Same cardinality)

• Unet Based Approach 
• Vnet (Residual Model)
• DeepMedic (Downsample / Crop) + Merge Later
• HighResNet (Right Features / Right Scale / Learn Relationships)

• Segmentation Task
• Tips and Tricks
• Data Augmentation to avoid overfitting

• Hyper Parameter Tuning (Grid Search, Random Search, Bayesian Optimization)

• Abstraction Layers for missing inputs
• Uncertainty of networks

Happy Mastering DL!!!