Learning notes TF053: Recurrent Neural Network, TensorFlow Model Zoo, reinforcement learning, deep forest, deep learning art, tf053tensorflow

Learning notes TF053: Recurrent Neural Network, TensorFlow Model Zoo, reinforcement learning, deep forest, deep learning art, tf053tensorflow

Recurrent Neural Networks. Bytes.

Natural language processing (NLP) applies the network model. Unlike feed-forward neural network (FNN), cyclic networks introduce qualitative loops, and the signal transmission does not disappear and continues to survive. The traditional neural network layer is fully connected, and neurons in the layer are not connected. The input of the hidden layer of the recurrent neural network includes the upper layer output and the previous hidden layer output. Expand in chronological order. The next step is affected by the processing in this step. Backpropagation (BP) algorithm for network training error sharing of parameter weights. Reverse propagation relies on the current layer and previous networks, and uses the backpropagation through time, BPTT algorithm. Load Network Time series signals by layer, feed-forward static neural network to dynamic network.
Supervised Sequence Labelling with Recurrent Neural Networks http://www.cs.toronto.edu /~ Graves/preprinthistory.

The development of recurrent neural networks.

VanillaRNN
-> Enhanced the hidden layer function
-> Simple RNN
-> GRU
-> LSTM
-> CW-RNN
-> Bidirectional deepening Network
-> Bidirectional RNN
-> Keep Bidrectional RNN
-> Combination of the two: DBLSTM
Recurrent Neural Networks, Part 1-Introduction to RNNs http://www.wildml.com/2015/09/recurrent-neural-networks-tutorial-part-1-introduction-to-rnns.

Enhance the hidden layer function.

Simple RNN (Simple RNN, SRNN ). Layer-3 networks: the hidden layer (context layer) adds context units. The context unit node is connected to the hidden layer node and the weight is fixed. Assume that the current t-moment is predicted in three steps P (wm ). The word Wm-1 maps to the word vector, INPUT (t ). Connects to the hidden layer CONTEXT (t-1) of the last training, and the sigmoid activation function generates the current t-moment CONTEXT (t ). Softmax function prediction P (wm ).
LSTM. Generally, the RNN gradient disappears, and the derivative chain law leads to concatenation and the gradient index level disappears. The cell structure is introduced, and RNN improves Long-Short-Term Memory (LSTM) of the model ). The Block has one cell, and the status parameter records the status. Three gates, input gates, and output Gates process input and output parameters, and forget gates set selective forgetting weights.
GRU (Gated Recurrent Unit Recurrent Neural Network ). Different distance words on the hidden layer have different effects on the state of the current hidden layer. The distance weighting of the influence of each previous state on the current hidden layer state is smaller. If an error occurs, only the corresponding word weight is updated. Two doors: reset the door r (new input and pre-memory are combined) and update the door z (leaving the pre-memory ). Reset Door 1, update door 0, and get normal RNN. Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, Yoshua Bengio paper Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling https://arxiv.org/abs/1412.3555.
CW-RNN (Clockwork RNN ). Clock frequency Drives RNN. Hidden Layer groups. Different Hidden Layer groups work at different clock frequencies to address long-time dependencies. Each group processes the input according to the specified clock frequency. Clock time discretization. Different Hidden Layer groups work at different time points. All hidden layers do not work at each step at the same time, accelerating network training. The speed of the large group of neurons in the clock cycle is slow, and the speed of the group is fast. The cycle is large, and the cycle is small. Group hidden layer neurons and record g. Each group has the same number of neurons and records k. Each group is allocated with a clock cycle of tiε {T1, T2,..., Tg }. All neurons in the group are fully connected. Tj> Ti, group j connects to group I cyclically. T1 <T2 <· <Tg. The connection direction is from right to left, and the speed group is slow. Jan Koutnik, Klaus Greff, Faustino Gomez, Jurgen mongodhuber paper "A Clockwork RNN" https://arxiv.org/pdf/1402.3511.pdf

Bidirectional deepening network.

Bidirectional RNN (Bidirectional RNN ). The output is related to the front and back sequence. Original bidirectional RNN, where the two RNN are superimposed up and down. The output is determined by the state of the two RNN hidden layers. Miske Schuster and Kuldip K. Paliwal papers Bidirectional Recurrent Neural Networks. Bidirectional LSTM and bidirectional GRU.
Deep Bidirectional RNN ). The hidden layer overlays multiple layers, and each step inputs a multi-layer network, providing stronger expressive learning capability and requiring more training data. Https://www.cs.toronto.edu of Hybrid Speech Recognition With Deep Bidirectional LSTM by Alex Graves, Navdeep Jaitly and Abdel-rahman Mohamed /~ Graves/asru_2013.pdf.

Training and Learning algorithms: BPTT (Back Propagation Through Time), RTRL (Real-time Recurrent Learning), and EKF (Extended Kalman Filter ).

TensorFlow Model Zoo.

TensorFlow model https://github.com/tensorflow/models. Many image and speech processing models. The checkpoint file can be used as a pre-training model. For example, Inception V1,inception_v1_2016_08_28.tar.gz. Cafe Model Zoo has many trained models for pre-training models to reduce the training time and number of iterations. It can be converted to the TensorFlow model https://github.com/ethereon/caffe-tensorflow.

Reinforcement learning ). AlphaGo policy network ). Reinforcement Learning is between supervised learning and unsupervised learning. There are only few tags (rewards) and there is a delay. Model learning environment behavior. Games, playing games, and games have multiple steps to make continuous decisions. Q-learning, Sarsa, Policy Gradient, Actor Critic. Including algorithm update and decision-making. Deep Q Network (DQN ).

Deep forest. Zhou Zhihua's paper Deep Forest: Towards an Alternative to Deep Neural Networks https://arxiv.org/abs/1702.08835, multi-level cascading forest (multi-grained cascade Forest, gcForest ). Only a small amount of data can be trained. The super parameter is smaller than the deep neural network, and the super parameter has high robustness and is easy to train.

Deep Learning.

In the field of painting, the Neural network Algorithm of Artistic Style (A Neural Algorithm of Artistic Style), Leon A. Gatys, Alexander S. Ecker, Matthias Bethge papers https://arxiv.org/#/1508.06576v2.20. Painting Style migration. Separate the image style and content, and combine different image styles and content to generate styles of content images. Meitu xiuxiu, magic man camera, face cute applications. Https://github.com/anishathalye/neural-style.
Music field. A large amount of MIDI audio melody training data, RNN generates the melody. Https://github.com/tensorflow/magenta.
Deep Learning can create small artistic samples as the seeds of inspiration.

References:
Analysis and Practice of TensorFlow Technology

Welcome to the Shanghai Machine Learning job opportunity, my qingxingfengzi