AT Lecture 5

• Lecture 5 discusses advanced techniques in ML and Deep Learning, specifically focusing on Deep Learning for Computer Vision

Image Classification and Vision Tasks

• Image classification is the task of assigning a label to an entire image
• Computer vision tasks include:
• Classification
• Classification + Localization
• Object detection

Pixels and Color Spaces

• Pixel values range from 0 (black) to 255 (white), representing intensity
• RGB Colorspace is multidimensional color space, that represents all colors made of Red, Green and Blue light

Limitations of Fully-Connected Neural Networks for Image Recognition

• Fully-Connected Neural Networks (FCNNs) are inefficient for image recognition due to dimension and size problems and overfitting
• Dimension Problem: Reshaping images to 1D loses spatial structure
• Size Problem: High-resolution images lead to extremely large models (e.g., a 1000x1000 RGB image with 100 layers would result in a 2.4GB model)
• Overfitting Problem: One weight per pixel leads to overfitting

Convolutional Neural Networks (CNNs)

• Convolutional Neural Networks (CNNs) are inspired by visual cortices and have special structures
• CNNs gained prominence after success in ImageNet competitions in 2012
• CNNs have layers with volume (length, width, depth)
• The depth of a CNN is typically 3 for color images, 1 for grayscale images, and arbitrary for hidden layers
• Basic operations in CNNs: convolution, pooling, and ReLU
• Convolution is analogous to matrix multiplication in a conventional network

Convolution Operation

• Convolution preserves the spatial structure of the image
• Filters are convolved with the image by sliding the filter spatially and computing dot products
• Filters extend the full depth of the input volume
• Activation maps are generated by sliding filters over the image spatially
• Multiple filters result in multiple activation maps, which can be stacked to produce a new image
• A Convolution Network is a sequence of Convolutional Layers, interspersed with activation functions

Convolution Layer Parameters

• When a stride is used in a layer, the convolution will be performed at set locations
• The stride determines the locations where convolution is performed in a layer
• Sparse Connectivity: A feature is creating from a region smaller than the input size (filter)
• Shared Weights: The same filter is used across the entire spatial volume
• Features in hidden layers capture properties of a region of the input image

Padding

• Padding is used in CNNs to extend the input image, adding a 'frame
• In CNN, padding addresses the issues where edges are used less and output images shrink
• Without padding: n’ = n – f + 1
• With padding of p: , where n is the filter height or width, and p is the amount of padding
• Padding can be of any size up to Fp-1
• Padding by a size P increases both the length and width of the input by 2P

Stride

• Strides involve the filter jumping a few pixels at each step
• Using a stride can reduce compute time, diminish the output size
• With padding of p and stride s, the output size is equal to

Pooling

• Pooling operates on square regions of size P x P and finds the max of the values in those regions
• Pooling preserves depth, is a subsampling process, and great reduces parameters for subsequent layers

Convolution as Matrix Operation

• Convolution can also be expressed as a matrix operation

Convolutional Layer Summary

• Accepts volumes of W1 x H1 x D1, Requires number of filters K, Their spatial extent F, their stride S, the amount of zero padding P $$ W_2 =(W_1 - F + 2P) / S + 1 $$$$ H_2 = (H_1 - F + 2P) / S + 1 $$$$ D_2 = K $$

Feature Recognition and Convolution Example

• Feature Recognition Example: The feature determines whether a picture is of an X or an O
• The process is as follows:
  • Line up the feature and the image patch, multiply each image pixel by the corresponding feature pixel, add them up, divide by the total number of pixels in the feature
• Convolution: Search every possible match
• Using Multiple Filters: One image becomes a stack of features (filtered images)

Max Pooling

• Max-pooling is a way of shrinking the input stack
• Window size (usually 2 or 3)
• Pick a stride (usually the sane window size, to prevent overlapping)
• Walk you window across filtered images, taking max
• A stack of images becomes a stack of smaller images

ReLU

• ReLU: Set zero for negative numbers; $$ f(x) = max(0, x) $$
• Stacking multiple layers is done by convolution, ReLU, and pooling

LeNet-5

• LeNet-5 is an example of a convolutional network for hand-written digit classification
• LeNet-5 has 7 layers, made of 3 convolutional layers (C1, C3, C5), 2 sub-sampling (pooling) layers (S2 and S4), 1 fully connected layer (F6), and an output (O) layer
• C5 is basically a fully connected layer as each unit is mapped to all the 16 input activation maps and filter size is same as input maps
• Pattern: Conv -> Pool Conv -> Pool -> FullyConn -> FullConn -> Output

CNN Applications

• CNNs have been used in "Classifications of multispectral colorectal cancer tissues using convolution neural network" J Pathol Inform 2017, 8:1