Zoo Crew

Poster submission
writeup-p5
writeup-p2
writeup-p4
writeup-p3
writeup-p1
writeup-p7
writeup-p6

IMPORTANT - Please see the try it out section for a link to the google doc paper as it was put out of order in the media section. Thank you!

Previous Checkin:

Introduction: The paper we are implementing studies the performance of different classifiers on the CIFAR-10 dataset and builds an ensemble of classifiers to reach a better performance. The paper shows that, for CIFAR-10, K-Nearest Neighbors (KNN) and Convolutional Neural Network (CNN), on some classes, are mutually exclusive, thus yielding higher accuracy when combined. The paper reduces KNN overfitting using Principal Component Analysis (PCA), and ensembles it with a CNN to increase its accuracy. Their approach improves the best CNN model from 93.33% to 94.03%. Here is a link to the paper. We chose to implement this problem because our group enjoyed the work done with the CIFAR-10 dataset in HW2 and hoped to implement a more sophisticated version of it. We came across this paper and found it to be both manageable and a more sophisticated version of what we worked on for HW2. This is a classification problem. We hope to accurately classify the CIFAR-10 images to the highest accuracy possible.

Challenges: Combining the CNNs and the KNN has proven difficult as we have very little experience dealing with KNNs. Beyond this tweaking the parameters to try to reach the CNN values which the paper shows have taken a bit of time.

Insights: The hyperparameters need tweaking to allow for higher CNN accuracy as well as faster runtimes. The performance is about on par with expectations at this point as classification with images of this scale is difficult and we have yet to merge CNNs or implement the KNN.

Plan: We are on track as of right now. Finalizing the KNN is a top priority and its implementation is critical to the success of the model. As of right now, we are not planning on changing any aspects of the project as we are still trying to emulate the paper.