MLTrain 1
Schedule
- 24 Sep 2016
24 Sep 2016
Registration
Session 1: TensorFlow paper
- Symbolic coding: From imperative to declarative programming
- Tensor as first class citizen
- The computational graph model
- ControlFlow: Mutating variables
- Managing hardware
- Device communication model
- Distributed computing
- Lossy data compression
- Automatic Differentiation
- Synchrony vs Asynchrony
- Debugging
Session 2: Creating and Visualizing a Computational Graph
- Variables vs placeholders
- Operations
- Creating a computational session
- Managing Hardware, cpus and gpus
- Using TensorBoard for Visualization
- Breaking declaratively: Mutating variables with controlFlow.
Break
Session 3: Linear Algebra with TF
- Sparse/Dense Matrix/Vectors operations
- Kronecker Products in TF
- From Matrices to Tensors
- Tensor Tiling: the map operator of TF
- Reductions on Tensors
- Thinking in batch
- Limitations of TF
Lunch
Session 4: Optimization in TF
- Creating a symbolic objective function
- Computing the gradients
- Build your own simple gradient Descent optimizer
- The mini-batch
- Inside the optimizer TF class
- Tweaking predefined optimizer by touching the gradients
- Presentation and Parameter tuning of famous optimizers: AdamOptimizer, RmsPropOptimizer, and AdaGrad
- Build your first linear regression in 3 lines!
Break
Session 5: Coding ML algorithms
- Continue building linear models (logistic)
- Using different objectives: L1, L2, Cross Entropy, Hinge Loss, and Maximum Likelihood
- Adding L1/L2 regularization
- Compare your implementation with the tensorFlow Linear module
- Regularizing with dropout
- Code your first multilayer perceptron (MLP)
- Debugging your model with TensorBoard
Session 6: Wide and deep modeling
- Presentation of the paper Wide and Deep recommender systems
- Build your first wide and deep model
- Beyond recsys, extending to wide and deep classifiers