27 Oct Consider a combined Lenet-5 and a single-layer RNN based visual captioning system that is trained to generate the sequence of small characters corresponding to ten digits, 0 to 9.

Consider a combined Lenet-5 and a single-layer RNN based visual captioning system that is trained to generate the sequence of small characters corresponding to ten digits, 0 to 9. For example, an input image of ‘7’ will generate the output character sequences {s, e, v, e, n}. An input image that is not of a digit will generate the output {n, o, n, e}. Assuming one hot representation is used for both input and output, 1. What is the minimum number of input nodes and minimum number of output node required in the RNN’? 2. Assuming linear combinations of the output of last convolution layer (after subsampling and unrclling) is used to initialize the RNN hidden layer, how many trainable parameters will be needed, excluding the CNN convolution parameters? Assume 50 hidden nodes are used in RNN. Show all steps clearly. No attention is used. 3. Over how many time steps, does the loss function need to be evaluated during training?