From the course: Deep Learning: Getting Started

Reusing existing network architectures

From the course: Deep Learning: Getting Started

Reusing existing network architectures

“

- [Instructor] Having discussed the concepts of neural networks and how to train them from scratch, let's now discuss some practical aspects of building neural network models. How do we build neural networks for a use case? An interesting fact about neural networks is that most neural network implementations are not designed and built from scratch. Designing a neural network with the right number of layers and nodes is a tedious, iterative, and time consuming process. Fortunately, the neural network community is very active in sharing their work with the rest of the world. They shared their knowledge and experiments for the rest of the community to build upon. To begin with several papers are published on the architectures for neural networks that have been successfully implemented and proven. You can start off your neural network by implementing a related architecture and then fine tune it for your use case. Implementation code for these neural networks is also available in open source repositories. This can be leveraged to create a neural network for your use case. Finally, open source models are available for popular implementations. The models include all the trained parameters and hyper-parameters packaged in standardized formats that are supported by popular deep learning networks. Let's quickly list some of the popular neural network architectures. LeNet5 is one of the earliest open source models that was built for document and handwriting recognition. AlexNet is a conventional neural network or CNN that is used for image recognition. ResNet is also a conventional neural network that overcomes some limitations of traditional CNNs. VGG is another class of popular CNN architecture. LSTM is a recurrent neural network that is used for predicting sequences of text. Transformers is the latest architecture that is revolutionizing generative AI. I strongly recommend researching and reading up on more of these architectures. There are also courses available for conventional neural networks, recurrent neural networks, and transformers.

Contents