Inspiration
Machine learning is very much an esoteric field of Computer Science that seems almost unapproachable to learn. However, we were bold enough to tackle the challenge of learning fundamental machine learning concepts. Joseph used his first datathon as an opportunity to apply his strong mathematical skills, while Nicky and Akil strived to expand their abilities in building machine learning models.
What it does
Team 7's Puzzle Solver uses machine learning to correctly guess the order of 2x2 jigsaw puzzles. In fact, it's able to do this across all 24 distinct permutations of a single puzzle. The model that makes this possible is called a Convolutional Neural Network (CNN), which is specialized for image classification and visual pattern recognition.
How we built it
We built our model using a wide variety of tools including numpy, pandas, matplotlib, TensorFlow, and Keras. In order to make it a CNN, we applied concepts of convolution filtering, max pooling, and flattening to the traditional neural network model in order to detect, enhance, and extract the most important features of an image in order to classify it.
We trained our own models on over 50,000 images, with ~2000 samples for each permutation. Utilizing an 80-10-10 split for training, testing, and validation, we were able to achieve an accuracy of ~95% after just a few hours.
Challenges we ran into
One of the biggest challenges we ran into was optimizing the amount of time our model spent on each image, which was ~2 seconds. With our dataset of over 50,000 images, this proved to be incredibly slow and inefficient. However, we solved this by adapting our traditional neural network to a CNN.
Accomplishments that we're proud of
Our entire team was able to learn about neural networks from the ground up, from the concepts of perceptrons, activation and loss functions, and optimizers, to convolutional methods, such as filtering and max pooling. The fact that we were able to accomplish all of this in just 24 hours has made this an invaluable experience for us and are excited to dive deeper into the field of machine learning with future projects and competitions.
Log in or sign up for Devpost to join the conversation.