A common approach for transfer learning is to freeze the bottom layers and fine$-$tune the top layers, why?

Consider an exhibition where different products are presented with never$-$seen$-$before shapes and designs. You have collected few pictures of these products and are asked to design a classifier to classify each product into a different class. You decide to use a domain adaptation approach with help of hand drawn sketches. Among the different domain adaptation approaches we discussed in the class, which one is the best choice for the considered problem and why?

How can we use hierarchical clustering for network compression? Calculate the compression ratio achieved using this approach?

Explain the role of temperature in softening the softmax scores?

$q,$

Code: CSL$7590$

Deep Learning

Page $2$ of $2$

Consider the autoencoder shown below where the bottleneck layer has higher dimension as compared to the input. Explain the problem associated with this architecture and suggest the solutions.

hat$(x{)}_i=f(W^{**}h+c)$