Skip to content

Latest commit

 

History

History
325 lines (163 loc) · 6.01 KB

cs-229-deep-learning.md

File metadata and controls

325 lines (163 loc) · 6.01 KB

Deep learning translation [webpage]


1. Deep Learning cheatsheet


2. Neural Networks


3. Neural networks are a class of models that are built with layers. Commonly used types of neural networks include convolutional and recurrent neural networks.


4. Architecture ― The vocabulary around neural networks architectures is described in the figure below:


5. [Input layer, hidden layer, output layer]


6. By noting i the ith layer of the network and j the jth hidden unit of the layer, we have:


7. where we note w, b, z the weight, bias and output respectively.


8. Activation function ― Activation functions are used at the end of a hidden unit to introduce non-linear complexities to the model. Here are the most common ones:


9. [Sigmoid, Tanh, ReLU, Leaky ReLU]


10. Cross-entropy loss ― In the context of neural networks, the cross-entropy loss L(z,y) is commonly used and is defined as follows:


11. Learning rate ― The learning rate, often noted α or sometimes η, indicates at which pace the weights get updated. This can be fixed or adaptively changed. The current most popular method is called Adam, which is a method that adapts the learning rate.


12. Backpropagation ― Backpropagation is a method to update the weights in the neural network by taking into account the actual output and the desired output. The derivative with respect to weight w is computed using chain rule and is of the following form:


13. As a result, the weight is updated as follows:


14. Updating weights ― In a neural network, weights are updated as follows:


15. Step 1: Take a batch of training data.


16. Step 2: Perform forward propagation to obtain the corresponding loss.


17. Step 3: Backpropagate the loss to get the gradients.


18. Step 4: Use the gradients to update the weights of the network.


19. Dropout ― Dropout is a technique meant at preventing overfitting the training data by dropping out units in a neural network. In practice, neurons are either dropped with probability p or kept with probability 1−p


20. Convolutional Neural Networks


21. Convolutional layer requirement ― By noting W the input volume size, F the size of the convolutional layer neurons, P the amount of zero padding, then the number of neurons N that fit in a given volume is such that:


22. Batch normalization ― It is a step of hyperparameter γ,β that normalizes the batch {xi}. By noting μB,σ2B the mean and variance of that we want to correct to the batch, it is done as follows:


23. It is usually done after a fully connected/convolutional layer and before a non-linearity layer and aims at allowing higher learning rates and reducing the strong dependence on initialization.


24. Recurrent Neural Networks


25. Types of gates ― Here are the different types of gates that we encounter in a typical recurrent neural network:


26. [Input gate, forget gate, gate, output gate]


27. [Write to cell or not?, Erase a cell or not?, How much to write to cell?, How much to reveal cell?]


28. LSTM ― A long short-term memory (LSTM) network is a type of RNN model that avoids the vanishing gradient problem by adding 'forget' gates.


29. Reinforcement Learning and Control


30. The goal of reinforcement learning is for an agent to learn how to evolve in an environment.


31. Definitions


32. Markov decision processes ― A Markov decision process (MDP) is a 5-tuple (S,A,{Psa},γ,R) where:


33. S is the set of states


34. A is the set of actions


35. {Psa} are the state transition probabilities for s∈S and a∈A


36. γ∈[0,1[ is the discount factor


37. R:S×A⟶R or R:S⟶R is the reward function that the algorithm wants to maximize


38. Policy ― A policy π is a function π:S⟶A that maps states to actions.


39. Remark: we say that we execute a given policy π if given a state s we take the action a=π(s).


40. Value function ― For a given policy π and a given state s, we define the value function Vπ as follows:


41. Bellman equation ― The optimal Bellman equations characterizes the value function Vπ∗ of the optimal policy π∗:


42. Remark: we note that the optimal policy π∗ for a given state s is such that:


43. Value iteration algorithm ― The value iteration algorithm is in two steps:


44. 1) We initialize the value:


45. 2) We iterate the value based on the values before:


46. Maximum likelihood estimate ― The maximum likelihood estimates for the state transition probabilities are as follows:


47. times took action a in state s and got to s′


48. times took action a in state s


49. Q-learning ― Q-learning is a model-free estimation of Q, which is done as follows:


50. View PDF version on GitHub


51. [Neural Networks, Architecture, Activation function, Backpropagation, Dropout]


52. [Convolutional Neural Networks, Convolutional layer, Batch normalization]


53. [Recurrent Neural Networks, Gates, LSTM]


54. [Reinforcement learning, Markov decision processes, Value/policy iteration, Approximate dynamic programming, Policy search]