Extra Credit Opportunity
This talk is about neural network, feature extracting, transfer modelling and stacked generalization. Neural network is a type of machine learning model that can classify data, it has input layer, hidden layer and output layer. Clare first uses an example of Spotify in which different types of music produces different sonic variable input and there are many artists in Spotify, neural network can then be used to predict music genre. Taking a proportion of data for training, the train test split can show how accurate the prediction is. Weights are applied to vectors these weights are eventually updated to make the model better (model-fitting). Clare then shift from general data to image data and introduced the idea of convolutional neural networks and terms like filters which help identify relevant features and assign proper weight. Her lab attempted to use algorithms to predict road quality based on merely satellite images instead of having to actually drive to a remote road. Then she used the cancer cell image net as an example to illustrate how to decide what layer to freeze. If image data are not available in a sufficiently large number, a problem of overfitting would appear and this problem can be solved with layer freezing and only use the weights from the transfer learning model. When analyzing an image, different people are likely to apply different algorithms that leads to different outcomes. Since we won’t use a single model to make any prediction, it is useful to take various features from multiple models. If any one single model is known to outperform others, the more training it, the more accurate the prediction could be, so heavier weight could be place on it. This is how stacked generalization works.
I think Clare’s talk is pretty good although I haven’t fully understood everything she went over yet. She used a lot of data examples to explain each new term and support her views. This enables those who know little about data science and machine learning to follow her pace. It is also very nice of her to interact with audiences when illustrating how computers or algorithms can be better predictive of an image with low resolution than human eyes. However, there are also some flaws. As professor Davis noted, the picture that Clare chose when talking about the quality of roads is more like a blurred picture of a good road instead of a picture of a road with poor quality. This confused me when I was watching the recording. Regarding the transfer modelling part, although Clare said that we should choose data as similar as possible, but I still don’t quite know in what sense would the data be similar. Generally, I really appreciate her for giving such an informative talk, which made me learn more about machine learning.