BMSO 758Q

Deep learning techniques have been widely used and successful in many domains in academia and industry. For example, an AI trained to classify images of skin lesions as benign lesions or malignant skin cancers achieves the accuracy of board-certified dermatologists. AI is also being applied to create autonomous vehicles (e.g., drones and self-driving cars) to make roads safer. We have also heard many other successful stories by AI in other areas, such as HR recruiting (intelligent resume filtering and ranking), marketing (automated campaign generation and better advertising strategies), and e-commerce (personalized recommender system to increase customer engagement). Another example is that almost every product they have developed within Google involves AI/deep learning. I am sure you have realized and experienced that. Thus, we hope learners can leverage such cutting-edge technologies to tackle their tasks. These advanced AI-based techniques are built upon fundamental neural networks. Different architectures of neural networks have pros and cons and are designed for specific applications.
In this course, we first cover the generic framework of deep learning. To understand every unit involved in this framework, we introduce our first network architecture - a fully connected feedforward neural network, particularly the concept of neurons, activation function, and the structure of neural networks. Next, we move to how to train neural networks. This is not a trivial task because it is a model optimization process involving several vital components. For example, the mechanism used to update parameters (we will learn a very important technique used in almost every deep neural network – gradient descent), which loss function is used to adjust parameters, best combination of hyperparameters (called hyperparameter tuning), and possible overfitting, all need to be taken into consideration.