Host: Chitta Baral and Yezhou Yang
Time: Every Thursday at Noon
Location: BYENG 210
CCC Post-Doc best practice program;
NVIDIA DLI program.
Intro to Deep Learning: Yezhou Yang
Activation Functions: Basic perceptron, Sigmoid, ReLU, (leaky) ReLU, identity, tanh, softmax (Yezhou, Vyom) Neural network modules: feed-forward, Word2Vec, CNN (Convolutional Neural Networks), Recurrent Neural Networks (RNNS), RNN with memory units - LSTM, Encoders (auto encoders), Decoders, Recursive Neural Networks, Generative Models, GANs (Generative Adversarial Networks), Relational Networks, Adversarial networks (Kausic - CNNs, Chieh-Yang - Word2Vec, Rudra - GAN, Xin - LSTM) Modules and Applications: Vision - CNN for recognition, GAN (and its extensions such as InfoGANs) Language - RNN, LSTM Decision Making - Software and libraries: Tensorflow, Theano, Torch, Keras (a Python library), Caffe, ONNX (an ecosystem for interchangeable AI frameworks) Tuning Parameters: network structure (how many layers), filter size of a CNN, learning rate (adaptive learning rate algorithms), how many epochs (Somak, Rudra, Mo) Reinforcement Learning using NN (Rudra - Recap of Reinforcement Learning Workshop) Recap of EMNLP. Students talking about their own research. Training Methods: SGD, … (ChengXi @UMD - remote presentation) Important Developments: AlphaGo, … Challenges and Questions: How to do incorporate background knowledge? How to make inference based on deep learning explainable? Relational Learning? Planning