1. Research background and rationale
1958, Rosenblatt proposed Perceptron model (ANN)
In 1986, Hinton proposed a deep neural network with multiple hidden layers (MNN)
In the 2006, Hinton Advanced Confidence Network (DBN), which became the main frame of deep learning.
Then, the efficiency of this algorithm is validated by Bengio Experiment 2.3 classes of depth learning models
2.1 Generating Deep model
DBN as the representative of the detailed introduction. The DBN model is a kind of deep mixed network, which is stacked in series with RBN as basic unit. The training of DBN is to fine-tune the traditional learning algorithm by first-layer training RBN.
2.2 Discriminant Deep Model
The identification of deep network includes deep stack network, convolution neural network and so on. Introduced by CNN as a representative.
In the 1962, Hubel studied the cat's vision theory, and proposed the concept of the receptive field, which was the first successful CNN model in 1984, Fukushima based on the perception of a neural perceptron.
2.3 Hybrid Deep model
3. The new development of the study of deep learning
4. Practical application of deep learning and challenges to face
Applications: speech recognition, image recognition, NLP (Natural language Processing)
Challenges: Theory, modeling, engineering challenges