Quiz Statistics

Which of your quiz scores will be dropped?

Which statements did David Hartley assert or imply in Observations on Man? (select all that apply)

Match the models to their definitions. 

1.- The McCulloch and Pitts model

2.- Is one of David Hartley’s Observations 3.- Frank Rosenblatt

4.- Associationism Theory by Aristotle

5.- Alexander Bain

6.- Lawrence Kubie

7.- Hebbian Learning

8.- Marvin Minsky and Seymour Papert

 A.___ known for his Connectionism Theories.

B.___ These are his four laws of association: The law of contiguity, The law of frequency, The law of similarity, The law of contrast.

C.___ Their mechanism known as the multilayer Perceptron could perform all Boolean functions including the XOR.

D.___ Modeled the memory as a circular network.

E.___ A proposed mechanism to update weights between neurons in a neural network.

F.___ made the first algorithmically described neural network with a learning algorithm that was provably convergent in some cases.

G.___ Our brain represents compound or connected ideas by connecting our memories with our current senses.

H.___ A Logical Calculus of the Ideas Immanent in Nervous Activity, a mathematical model of a neuron.

What term does the following sentence describe? "The finding that neurons whose connections form cycles can model memory."

Suppose you have a neural network with only two inputs I1 and I2, and two output neurons O1 and O2. There are no hidden units. All neuron outputs are Boolean. The network has been trained to model the following function:

The vast majority of the data set has I1 = 1, I2 = 0. If we apply Hebbian learning with this trained network on our data set, will the behavior of the network eventually change? If so, what will happen first?

A majority function is a Boolean function of N variables that produces a 1 if at least N/2 of the inputs are 1. Which of the following are true? (select all that apply)

How many neurons will the smallest (in terms of neurons) network that implements the truth table shown by the following Karnaugh map need?

How does the number of weights in an XOR network with 1 hidden layer grow with the number of inputs to the network?

We would like to compose an MLP that can represent a three-dimensional version of the cross-hash function shown below (the illustration is 2-dimension because it is difficult to depict the 3-D version of the map).  The input is now a three-dimensional vector. The output map comprises alternating cuboids of yellow and grey, where yellow represents a region where the function must output 1, and grey cuboids are regions where it must output 0. 

If the network is restricted to only two hidden layers (fixed total depth of 3 including output layer), which of the following is true of the number of neurons required to model the function where the network must output a 1 if the input is in the yellow regions, but 0 if it’s in the grey regions?

Suppose you are using a neural network to recognize MNIST handwritten digits in a 28x28 grid of pixels. The neural network contains neurons that input and output REAL values. Almost any activation function would work, but for now assume that all neurons use the sigmoid activation function.

The network architecture contains two parts, an encoder and a decoder. The output of the encoder is passed as the input of the decoder:

Encoder: 784-dimensional input, 1-dimensional output. Contains as many neurons and hidden layers as needed.

Decoder: 1-dimensional input, 10-dimensional output. Contains as many neurons and hidden layers as needed.

This is a horrible architecture in practice because of the 1-dimensional bottleneck in the middle. In theory though, does there exist a network of this form that correctly classifies all the available MNIST digits?

What term does the following sentence describe? "A modification of Hebbian learning addressing its lack of weight reduction and its instability."

Under which conditions will the perceptron graph below fire? (select all that apply). In the figure the number within the circle is the threshold.  The numbers on the edges are weights.  The perceptron fires if the total incoming signal is at least equal to the threshold.

Is the following statement true or false?

"For any Boolean function on the real plane, which outputs a 1 within a single connected but NON-CONVEX region in the plane (i.e. not a convex polygon), you need at least three layers (including the output layer) to model the function EXACTLY"

According to the university’s academic integrity policies, which of the following practices are NOT allowed in this course? Select all that apply.

We sometimes say that neural networks are connectionist machines as opposed to Von Neumann machines. Which of the following describe why we make this distinction? (select all that apply)

How does the number of weights in an XOR network with 2 hidden layers grow with the number of inputs to the network?

What is the implication of Shannon’s theorem on network size (as a function of input size) for Boolean functions? (select all that apply)

Is the following statement true or false? An activation function must always be a squashing function.

Suppose the data used in a classification task is 10-dimensional and positive in all dimensions. You have two neural networks. The first uses threshold activation functions for hidden layers, and the second uses softplus activation functions for hidden layers. In both networks, there are 2000 neurons in the first hidden layer, 8 neurons in the second hidden layer, and a huge number of neurons for all later layers. It turns out that the first network can never achieve perfect classification. What about the second network?

Which are NOT one of Aristotle's laws of association? (select all that apply)

How does the number of weights in an XOR network with arbitrarily many hidden layers grow with the number of inputs to the network?

Which of the following are impossible in theory? Assume all networks are finite in size, though they can be as large as needed. (select all that apply)

When was the first connectionist network model proposed?

Is the following statement true or false? Hebbian learning allows reduction in weights and learning is bounded.

Under which conditions will the perceptron graph below fire? (select all that apply).

In the figure the number within the circle is the threshold.  The numbers on the edges are weights.  The perceptron fires if the total incoming signal is at least equal to the threshold.

If the first layer of a network operating on real-valued inputs does not have sufficient neurons to capture all the basic features of the function (map), which of the following activation functions gives subsequent layers the best chance to capture the features missed by the first layer?

Under which conditions will the perceptron graph below fire? (select all that apply)  In the figure the number within the circle is the threshold.  The numbers on the edges are weights.  The perceptron fires if the total incoming signal is at least equal to the threshold.

A neural network can compose any function with real-valued inputs (with either Boolean or real-valued outputs) perfectly, given a sufficient number of neurons

Which of the following statements are true about MLPs? (select all that apply)