Work

100% Self-made AI : image recognition

Year: 2023
Type: dissertation
Abstract: Artificial Intelligence has seen major improvements in recent years. Multimodal large language models such as ChatGPT can answer almost any question, and produce almost any wanted text piece and Image text models such as DALL-E can produce almost any picture with a given prompt. Because the importance and usefulness of AI software seem to only rise. This work aims to understand how AI works and how it is trained, by constructed ex-ample AI. \n \nThe constructed example will be neural network image classification AI. The work will first go through the necessary general theoretical background information, after which the program will be constructed with Python using the Numpy library for calculations and Pyplot for drawing images. The build program achieved 74% accuracy for the data, for which it was trained and 66.6% accuracy for the test data. The main limiting factor of this work was the need to reduce the total image dataset from 25000 images to only 9000 images, and the need to additionally scale images to 50x50 resolution. This was done to save memory and reduce the model training time.
Author: Karppinen, Valtteri
Language: English
Open Access status: green
OpenAlex ID: W6987188195
Cites: 8
Cited by:
Related to: 10