Classification accuracy is the most straightforward measure for the performance of a classifier. Explain the reason why we need so many other performance measures such as sensitivity, specificity, area under the curve, etc. (5%) Provide an example where classification accuracy might be misleading but those other measurements are more appropriate. (5%)
List the limitations of bag-of-words representation for document analysis. (5%)
Describe what is Term Frequency-Inverse Document Frequency (TF-IDF) weighting and what problem it tries to solve? (5%)
Describe the purpose of Principal Component Analysis (PCA) and how it works (5%). Do you think PCA is a dimensionality reduction technique? (5%)
Describe the purpose of t-SNE and how it works (10%)
Use your own language to describe what is deep learning (5%)
For regular feedforward artificial neural network, which includes multiple layers of artificial neurons to map the input feature variables to the output variables, the neurons in adjacent layers are fully connected with each other. The mapping from any specific layer (except the top layer) to one layer above includes two operations: 1) a linear transformation (weighted sum); 2) a nonlinear activation. Please explain why the nonlinear activation is needed (hint: think of what would happen if there’s no such nonlinear activation). (10%)
As introduced in the class, in convolutional neural network (CNN), a specific neuron in each layer (except the top) is connected to a subset of neurons of the layer below. This subset corresponds to a specific spatial local region in the neuron matrix. Explain the benefits of this comparing with full cross-layer connections. (10%)
Suppose we have an image illustrated as follows The size of the image is 7×8, where each cell represents a pixel. The image is simply black and white, where the white pixel is with intensity value 255, and the black pixel is with intensity value
If we use the following convolution filter to convolve the entire image, please write down the resultant image after the convolution. Please explain what are the effects of these convolution filters. (20% = 5%x4) (a) (b) (c) (d)
Both PCA and Autoencoder can be viewed as methods aiming at learning novel representations of the data vectors. Please perform a head-to-head comparison between PCA and Autoencoder and explain the potential pros and cons of them both. (5%)
Autoencoder include two phases: the encoding phase learns a compact representation of the original data vector through a multi-layer fully connected feedforward neural network, the decoding phase maps the compact representation back to the original data space with the transpose (reverse) of the encoding network. Autoencoder is designed for vector based data representations. Now think about images. If we want to design a CNN based autoencoder, which first transforms the input image into a compact vector based representation through several layers of CNN including convolution+activation and pooling layers in the encoding phase and then decode such representation back to the original image space. The model parameters will be optimized through the minimizing reconstruction loss between the original image and decoded image. Please describe how you want to design the decoding phase and the key operations. (10%