Extra Credit Response

Clare started by introducing the concept of neural networks which is a machine learning technique to classify data. For illustration, she talked about a neural networks she made during a math competition to predict genres of music on Spotify based on different sonic variable inputs like danceability, energy, and valence. For neural networks training, model fitting is the process of applying weights and updating them on different features in order to increase the predicative power of model. This process may take from a few minutes to several days. Like all the model training processes, trained test split method is also used for the validation purpose. Then she talked about image classification process which is different from for the case of classifying music genres. A convolutional networks is needed for image classification as it incorporates different filters to identify different features of the input image. Transfer learning allows people to take the weights that have already been trained and apply those to their own model. Clare also mentioned one of her CNNs to predict road quality based on satellite imagery. Three different transfer learning models were used for her model including VGG 16, Inception V3, ResNet 50. Then she explained the application of freezing, the process of keeping parameters from the transfer learning model and apply to her own model. The idea of stack generalization is introduced by an example of identifying the subject on an image by different people. Since different algorithms extract different features, they may also produce different classification results. Therefore, the overall prediction power can be improved by aggregating predictions from different models into a new model, which is the process of stack generalization.

I think Clare gave a really informative talk and I am able to get an overview of the machine learning technique, convolutional neural networks. She used a lot of examples to help audience get a better understanding of the terms she discussed. For example, the audience interaction activity to identify the subject of an image successfully helped me to get the idea of stack generalization. Although it is quite hard to understand her explanation of the code, generally she did a really good job of introducing neural networks for audience with limited machine learning experiences like me.