04 11 19

Page 1

Family name …………………………………………………. First Name ……………………………… ID……………………………………… 1 Scuola di Ingegneria Industriale e dell'Informazione ‐ Politecnico di Milano NeuroEngineering (I) – Prof. Pietro Cerveri Mid‐term exam (type A) November 4th 2019 Multiple choice questions Question 1 (score: 3pts) Given the perceptron learning rule for the training of a two‐class neural classifier, we can assert that: 1. no weight initialization is required 2. only on‐line update can be applied 3. the convergence is assured as the number of examples in the training set is large 4. the scalar product of weights with all the patterns into the first class should be have the same sign 5. the training will succeed whether the output of the activation function in all the neurons is far from saturation Question 2 (score: 3pts) In the training of a fully‐connected ANN, with two hidden layers, using backpropagation and on‐line update, the computation of weight (2,3) into the second hidden layer depends directly on the: 1. output signal of the 2 nd neuron in 1st hidden layer 2. error of all the neurons in the output layer 3. output signal of the 3rd neuron in 2nd hidden layer 4. output signal of the 3rd neuron in 1st hidden layer 5. all the weights of the neurons in the 1st hidden layer Question 3 (score: 3pts) Let consider one ANN for processing a 64 64 gray‐scale image. The first convolutional layer implements 1 0 d i f f e r e n t f e a t u r e m a p s a l l w i t h a f i l t e r s i z e o f 33. Assuming no zero‐padding and a stride of 1 in both directions, we can assert that: 1. the size of the image in output from one feature map is 6464 2. the number of free parameters of each feature map depends on the size of the input image 3. the amount of zero‐padding does not affect the size of the output…