A shortcut to getting started with AI is to understand and use various network models and create your own applications based on actual scenarios. Neural network models are the common wealth of mankind.
Neural Networks
Neural networks can be divided into three main types: feedforward neural networks, feedback neural networks, and graph neural networks.
Feedforward neural network is a simple neural network, also known as multi-layer perceptron (MLP), in which different neurons belong to different layers and are determined by the input It is composed of layer - hidden layer - output layer. The signal is transmitted in one direction from the input layer to the output layer without feedback in the middle. Its purpose is to fit a certain function, which is represented by a directed acyclic graph. Feedforward neural networks include activation functions (sigmoid function, tanh function, etc.), loss functions (mean deviation loss function, cross-entropy loss function, etc.), optimization algorithms (BP algorithm), etc. Commonly used model structures include: perceptron, BP neural network, fully connected neural network, convolutional neural network, RBF neural network, etc.
The output of a feedback neural network is not only related to the current input and network weights, but also to the previous input to the network. It is a directed cyclic graph or an undirected graph, with strong associative memory capabilities and optimized computing capabilities. Commonly used model structures include: RNN, Hopfield network, restricted Boltzmann machine, LSTM, etc.
A graph is a collection of functions organized according to a graph structure in a topological space for relational reasoning, including social networks, knowledge graphs, molecular graph neural networks, etc. Graph neural networks usually have the following network models: graph convolutional networks, graph autoencoders, graph generation networks, graph spatiotemporal networks, and graph attention networks.
Image Identification
AlexNet、GoogLeNet、VGG、ResNet、ResNeXt、DenseNet和SENet。
Text Categorization
Tf-Idf, Word2Vec and BERT.
Speech Recognition
- ASRT: Chinese speech recognition system based on deep learning, using convolutional neural network (CNN) and connectivity temporal classification (CTC) methods, using a large number of Chinese speech data sets for training, and transcribing sounds into Chinese pinyin
- DeepSpeechRecognition: Use CNN+DNN+CNN
- end2end_chinese_speech_recognition: a small open source model
- Wenet: A relatively mature model for industrial open source implementation
- whisper: openai’s speech model
- pycorrector: speech error correction
- pyttsx3: text to speech
- Ali's Paraformer
references
ALL