TensorFlow practical Google Depth Learning Framework (i) _ depth learning

Chapter One introduction to Deep learning

1. Artificial sometimes not very good to extract the characteristics of the entity, then there is an automatic way. Yes, one of the key problems in the deep learning solution is to automatically combine simple features into more complex features and use these combination features to solve problems.

2. Depth learning is a branch of machine learning that, in addition to its ability to learn the links between features and tasks, can automatically extract more complex features from simple features.

3. The depth of learning is mainly concerned with how to build an intelligent computer system to solve the problems encountered in artificial intelligence. Computational neuroscience focuses on how to build more accurate models to simulate the work of the human brain.

4. In 1958, Professor Frankrosenblatt presented the Perceptron model (Perceptron), the first model to learn feature weights based on sample data. However, the perceptron can only solve the linear and sub problems, and cannot solve the differences or problems. (Perceptrons:an introductionto Computational Geometry)

5. Distributed knowledge Representation: The core idea is that knowledge and concepts in the real world should be expressed through the presentation of multiple neurons, and that every neuron in the model should be involved in the expression of multiple concepts

6. DataSet 1.    Image DataSet Imagenet 2.    Face recognition DataSet LFW 3.    Handwriting Digital recognition DataSet Mnist 4.    SVHV Data Set (Stanford University) 5.    Timit Data set 6.    Corpus: Wordnet,conceptnet,framenet (describing the relationship between words in natural language) 7.    GloVe (Stanford open source word vector) 8. Sentimenttreebank data Set (Stanford)

7. Depth Learning Algorithm alexnet

8. Image Classification Competition ILSVRC

9. The word vector describes each word as a vector with a relatively low relative dimension, and the corresponding word vector should be closer in space for similar words. So the similarity of the word meaning can be described by the distance in the space. In the word vector, the concept of gender has been implicitly expressed.

Alphago consists mainly of three parts: Monte Carlo Search (Montecarlo tree search,mcts), Valuation Network (valuenetwork), moves network (Policy Network). Monte Carlo search realizes the search for different drop points, it will be more intelligent to find the best drop point according to the valuation network and moves network to judge the drop situation.