Question 1

During backpropagation training, the purpose of the delta rule is to make weight adjustments so as to&#10;A) minimize the number of times the training data must pass through the network.&#10;B) minimize the number of times the test data must pass through the network.&#10;C) minimize the sum of absolute differences between computed and actual outputs.&#10;D) minimize the sum of squared error differences between computed and actual output.

Accepted Answer

The purpose of the delta rule in backpropagation training is to adjust the weights in order to minimize the sum of squared error differences between the computed and actual outputs. This is because the sum of squared errors is a common and effective measure of the difference between the expected output and the actual output of a neural network. By minimizing this metric, the network becomes more accurate in making predictions and classifying inputs.

Question 2

Genetic learning can be used to train a feed-forward network. This is accomplished by having each population element represent one possible&#10;A) network configuration of nodes and links.&#10;B) set of training data to be fed through the network.&#10;C) set of network output values.&#10;D) set of network connection weights.

Accepted Answer

Genetic learning is used to train a feed-forward network by having each population element represent one possible set of network connection weights. The genetic algorithm then evolves these weights over generations to improve the network's performance on a given task.

Question 3

With a Kohonen network, the output layer node that wins an input instance is rewarded by having&#10;A) a higher probability of winning the next training instance to be presented.&#10;B) its connect weights modified to more closely match those of the input instance.&#10;C) its connection weights modified to more closey match those of its neighbors.&#10;D) neighoring connection weights modified to become less similar to its own connection weights.

Accepted Answer

In a Kohonen network, the output layer node that wins an input instance has its connection weights modified to more closely match those of the input instance. This is called weight adaptation or weight adjustment. Essentially, the winning node "learns" to recognize input instances that are similar to the one it just won. This helps the network group similar inputs together in the same region of the output layer.

Question 4

A feed-forward neural network is said to be fully connected when&#10;A) all nodes are connected to each other.&#10;B) all nodes at the same layer are connected to each other.&#10;C) all nodes at one layer are connected to the nodes in the next higher layer.&#10;D) all hidden layer nodes are connected to all output layer nodes.

Accepted Answer

In a fully connected feed-forward neural network, all nodes in one layer are connected to the nodes in the next higher layer. This means that each node in one layer is connected to every node in the next layer, which is why it is also called a dense network or fully connected layer. Options A and B, where all nodes are connected to each other, or all nodes in the same layer are connected to each other, respectively, are not true as it would result in very large networks with a large number of connections that might not be necessary. Option D is not true as it suggests that only the nodes in the hidden layer are connected to the output layer, whereas in a fully connected network, all nodes in one layer are connected to every node in the next layer.

Question 5

Which one of the following is not a major strength of the neural network approach?&#10;A) Neural networks work well with datasets containing noisy data.&#10;B) Neural networks can be used for both supervised learning and unsupervised clustering.&#10;C) Neural network learning algorithms are guaranteed to converge to an optimal solution.&#10;D) Neural networks can be used for applications that require a time element to be included in the data.

Accepted Answer

Although neural network learning algorithms generally converge to a good solution, they are not guaranteed to converge to an optimal solution. Local minima, where the algorithm gets stuck, can be a problem. Therefore, it is important to initialize the network and tune the parameters carefully to improve the chances of finding a good solution.

Question 6

This neural network explanation technique is used to determine the relative importance of individual input attributes.&#10;A) sensitivity analysis&#10;B) average member technique&#10;C) mean squared error analysis&#10;D) absolute average technique

Accepted Answer

Sensitivity analysis is used to determine the relative importance of individual input attributes by assessing how much the output changes with small perturbations in the input values. This technique allows for the identification of the input attributes that have the greatest impact on the overall prediction of the model. Average member technique, mean squared error analysis, and absolute average technique are not specifically used for this purpose.

Question 7

A two-layered neural network used for unsupervised clustering.&#10;A) backpropagation network&#10;B) Kohonen network&#10;C) perceptron network&#10;D) aggomerative network

Accepted Answer

Kohonen network, also known as self-organizing map (SOM), is a type of neural network that is widely used for unsupervised clustering. It is a two-layered network consisting of input layer and output layer. The output layer is a grid of nodes with each node representing a cluster. During training, the weights of the input layer are adjusted to minimize the distance between the input data and the closest node in the output layer. The nodes in the output layer are then grouped into clusters based on their similar weights. Kohonen network is known for its ability to preserve the topological structure of the input data, making it a popular choice for data visualization and exploratory data analysis.

Question 8

Neural network training is accomplished by repeatedly passing the training data through the network while&#10;A) individual network weights are modified.&#10;B) training instance attribute values are modified.&#10;C) the ordering of the training instances is modified.&#10;D) individual network nodes have the coefficients on their corresponding functional parameters modified.

Accepted Answer

Neural network training involves modifying the weights of the network based on the training data. This is typically achieved through an optimization process such as gradient descent, which adjusts the weights based on the error between the network's output and the desired output for each training instance. Option B and C do not accurately describe the training process, and D is not a commonly used term in neural network training.

Question 9

The values input into a feed-forward neural network&#10;A) may be categorical or numeric.&#10;B) must be either all categorical or all numeric but not both.&#10;C) must be numeric.&#10;D) must be categorical.

Accepted Answer

The answer of The values input into a feed-forward neural...

Question 10

Epochs represent the total number of&#10;A) input layer nodes.&#10;B) passes of the training data through the network.&#10;C) network nodes.&#10;D) passes of the test data through the network.

Accepted Answer

Epochs represent the total number of passes of the training data through the network. Therefore, the best choice is B. The other options are not related to epochs.

Deck 8: Neural Networks