Karo Castro-Wunsch, Ramaneek Gill, Ishtiaque Khaled, James Yuan
Implementing and Evaluating CNNs, attention mechanism enabled LSTMs and GANs for the task of high dimensionality image classification. The classification task is to identify which scans of breasts have cancer. The project is being done as a submission to the Digital Mammography Dream Challange.
Ramona Mirtorabi, Karo Castro-Wunsch
We explored different models for the classification of water pipes in Ontario. In comparing models we are trying to build a model that is more consistent and has a higher performance than the system of expert advice currently in place.
Karo Castro-Wunsch, Andrew Petersen, Alireza Ahadi
Published: Special Interest Group on Computer Science Education (SIGCSE) 2017
This work provides two contributions: first, a partial reproduction of previously published results, but in a different context, and second, an exploration of the efficacy of neural networks in solving this problem. Our findings confirm the importance of two features (the number of steps required to solve a problem and the correctness of key problems), indicate that machine learning techniques are relatively stable across contexts (both across terms in a single course and across courses), and suggest that neural network based approaches are as effective as the best Bayesian and decision tree methods. Furthermore, neural networks can be tuned to be reliably pessimistic, so they may serve a complementary role in solving the problem of identifying students who need assistance.
Karo Castro-Wunsch, Michael Guerzhoy
An investigation into automated song transposition. Segmentation and mixing via STFT based audio analysis and recomposition proved to have mixed results where some songs were correctly segmented whereas others with less regular structures and more prevalent vocals were. Style Transfer to translate a song from one instrument to another was implemented with an seq to seq model. The model was able to reproduce music in short time windows.
K Castro-Wunsch, W Maga, C Anton
Published: Thirtieth AAAI Conference on Artificial Intelligence 2016
Computers and Games 2016
URSCA MacEwan Journal
We investigated Parameterized Poker Squares to approximate an optimal game playing agent. We organized our inquiry along three dimensions: partial hand representation, search algorithms, and partial hand utility learning. For each dimension we implemented and evaluated several designs, among which we selected the best strategies to use for BeeMo, our final product. BeeMo uses a parallel flat Monte-Carlo search. The search is guided by a heuristic based on hand patterns utilities, which are learned through an iterative improvement method involving Monte-Carlo simulations and optimized greedy search.
Vicky Bilbily, Karo Castro-Wunsch
We engaged in retinal jousting with a broad range of flora.