22 01 20

Page 1

Family name …………………………………………………. First Name ……………………………… ID……………………………………… 1 Scuola di Ingegneria Industriale e dell'Informazione ‐ Politecnico di Milano NeuroEngineering (I) – Prof. Pietro Cerveri I Session January 22nd 2020 Multiple choice questions Question 1 (score: 2pts) Let us consider a perceptron featuring binary activation with n inputs. Such a network can be used as a: 1. data compression model 2. decision support system 3. function fitting 4. data up‐sampling system 5. mechanism to store n‐dimensional patterns SOLUTION The output is binary so that the net can be used to classify th e input into two classes. It is natural therefore to use the net as a decision support system. Function fitting requires linear activation in the output. Storage requires feedback in the net Question 2 (score: 2pts) In the model of the perceptron the correspondence with the neural activity (spiking) of a biological neuron is mapped as: 1. type of activation function 2. frequency of the output signal 3. value of the action potential 4. amplitude of the output signal 5. modulation of the neural threshold SOLUTION The biological encoding of spiking frequency is simplified into the neural model of the perceptron by means of a modulation of the amplitude of the output Question 3 (score: 3pts) You consider a multi‐layer full‐connected FFNN that should undergo training by means of error backpropagation in the delta rule paradigm. We can assert that: 1. over‐fitting may be reduced by increasing the number of neurons in the hidden layers 2. the update of one weight of a hidden neuron is proportional to the value of the activation function derivative of all neurons in the next layer 3. increasing the number of network layers improves the quality of the convergence 4. step activation can be…