CS6825: Computer Vision word cloud

Deep Learning

 

deep learning architectures

  • feed-forward networks

  • auto-encoders (output want to recover input image, middle layer smaller - use results of middle layer for compression)

  • recurrent neural networks (RNNs) (backward feeding at run time as part of input into middle layer)

  • convolutional neural networks (CNNs)

 

topics in Deep Learning

  • Neural Networks: perceptrons, activation functions, and basic neural networks. recall from discussion of NN - you train with input and desired outputs and use error to adjust weights in some kind of "learning algorithm".

  • Logistic Classifier: this is a NN that the output layer has nodes that sum up to 1.0 --- meaning 100% probability. So each output node represents probability of that "thing it respresents" occurring.

 

  • Optimization: techniques for optimizing classifier performance, including validation and test sets, gradient descent, momentum, and learning rates. You can control this in the setup of our "deep learning" neural network.
  • Regularization: techniques, including dropout, to avoid overfitting a network to the training data.
  • Convolutional Neural Networks:building blocks of convolutional neural networks, including filters, stride, and pooling. Image Processing:


    • Input layer = present as input every pixel in the image (or sub-image). Could be rgb or grey or lidar or ?

Object Recongition with Deep Learning

 

typically little image processing --here example where don't pass entire image but, subimages (regions)

 

© Lynne Grewe