Student name: Patrick Whelan
Project title: The analysis and subsequent building of virtual terrain representations with neural networks
Course: BSc (Hons) Computer Science
The aim of the project is to teach computers (neural networks) how natural terrain images (height-maps) look and then training the computers to generate entirely new terrain images from what they’ve learned. This is a relatively new area of study, and a very exciting topic. The project tested the computer’s generated terrain both algorithmically and by having human participants choose between real and computer-generated terrains and decide which were real. Artificial intelligence has progressed in recent years and reading about these innovations inspired me to finally take the plunge and learn just how these systems work and create my own. The type of AI I focused on has the goal of learning from input data to create entirely new output data. This could mean creating entirely new music, art or literature when given the works of famous composers, artists and writers. It’s always amazes me to see an image generated entirely by a computer.
Many of the challenges faced were due to this area being in its infancy, so access to relevant journals and articles were limited. A lot of experimentation was required before I felt confident in my abilities.
Python was used for creating the AI (which is a Deep Convolutional Generative Adversarial Neural Network [DCGAN]) using the TensorFlow and Keras libraries, the most popular libraries for artificial intelligence work.
Training of the neural network was done on Google Collab’s servers.
Blender 3D was used to create 3D models from the 2D images.
The project was very successful in demonstrating that neural networks produced more realistic terrain images than traditional methods. The AI created was successful, it was given black & white height-map images, it read the land gradient data and successfully generated new height map images that tricked participants tasked with discerning whether the images they were seeing was computer generated or not. The images generated by the AI also performed well when analysed against more traditional methods of generating height-maps (i.e. using noise functions).
The AI could be refined to produce more accurate images and images with a greater resolution.
My time at University
I have studied a broad range of modules, including; computer science core concepts, programming, mathematical models, statistics, databases, web systems and network security. During my time at EHU I had an assigned mentor who was an older student, who was extremely helpful. The lecturers are always happy to explain something again (even if it’s the 100th time!).
Societies have been a large part of my university experience. They helped to kick-start my social life at university and as I became one of the older members it was a great experience to support new members.
Services at Edge Hill have always been great whether that be the academic departments helping me wrap my head around things, the accommodation team helping me sort out problems with my room or friends with mental health issues seeking counselling in the new library.