Feed Forward Neural Networks with Pen and Paper
Year: 2024
Type: article
Abstract: The course provides an overview of Feed Forward Neural Networks and their role as universal approximators in the machine learning world. The course starts with an overview of the basic architecture that underpins FFNN, which comprises the main building blocks: input layer, hidden layers, and output layer. This course emphasizes how the data flows forward through the network during forward propagation and also explains how the network learns by backpropagation—the weight adjustment process to minimize the errors. The Universal Approximation Theorem is emphasized here, showing that FFNNs are powerful tools in solving a lot of complex problems—that they can approximate any continuous function provided they have the required number of neurons and suitable forms of activation functions. Practical applications were done in detail: classification, regression, and function approximation tasks of FFNN. The course also covers how common challenges such as overfitting, vanishing/exploding gradients, and key concepts of regularization and optimization of neural networks are handled when training neural networks. It also provides background on a range of activation functions, the role of network depth, and state-of-the-art training methods. By the end of the course, students will know the basics of FFNNs and how they have been used in image recognition, financial forecasting, and function approximation, among other areas, so that they can then apply neural networks to almost any machine learning problem at hand.
Author: Gafsi, Iheb
Institution: National Institute of Applied Science and Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W6893218735
FWCI:
Cites:
Cited by:
Related to: