Neural Networks Learning
- The aim of this project was to utilise Neural Networks in order to create a model that predicts handwritten digits from images.
The dimensions of the images used in this project were 40x40 pixels and images contain digits from 0 to 9. A small sample of this dataset is illustrated below.
Figure 1: First 100 digits of dataset
A typical structure of Neural Networks is shown in Figure 2, where information of each images is propagated through units and layers.
Figure 2: Structure of Neural Networks
One way to understand how this algorithm “learns” is to visualise information from one unit that belongs to a hidden layer. More specifically, Figure 3 illustrates information from 25 different units in the network.
Figure 3: Hidden layer
The training accuracy of the trained model can be calculated by testing some images with handwritten digits.
Figure 4: Training accuracy.
A more detailed description of this project’s implentation in Matlab can be seen in this github repository: Link to Github repository