Neural Network to Clustering
The following code leverages a Neural Network (Inception V3) to classify images using vectors, then runs these image vectors through a Kmeans clustering.
The goal of this exercise is to see how good the NN + K-Means duo is in batching and identifying pictures.
- Image Dataset: cifar10, minist; we could import any other dataset if needed
- Neural Network: InceptionV3. This is trained using the imagenet training set
- Clustering Algorithm: K-Means. Using batches of 200
Results:
- Cifar10: the results for cifar10 was mediocre. We can see that the algorithm did a good job batching trucks together, specifically if the color red was incorporated. Overall the success of this batch was ~80%, meaning 80% of the pictures were accurately batched.
- Minist: despite this dataset subjectively being easier to recognize, the algorithm had trouble batching accordingly. It was 100% accurate when batching the number 1, but often had difficulties with other batches.
Suggested improvements: We lost massive picture quality when resizing our image datasets to be compatible with our Neural Network (150x150 requirement). If we could improve this process I suspect our accuracy would increase.