GAMLE SPM

Question 1

Artificial neural networks are?

Answer 1

Parallel distributed computing systems

Question 2

In the delta rule, the weight changes are given by?

Answer 2

By the difference between target and output, multiplied by the input

Question 3

In simulated annealing

Answer 3

At high temperatures the network randomly explores the state space

Question 4

Difference between batch and online (or mini batch learning?

Answer 4

Gradient descent against the global error function in batch learning (computed first across all patterns and then weights are changed.

Stochastic gradient descent against the partial error function (i.e. weights are changed according to the gradient computed for a single example (or for a small number of examples in the minibatch)

Question 5

What does AUC mean, range in between and what does 1 in AUC mean?

Answer 5

Area under the curve. Ranges between 0 and 1. Represents how good the classifier is, where AUC means the predictions are 100 pct. correct (no false positives, 100 pct. true positives).

Question 6

What are variational autoencoders?

Answer 6

Variational Autoencoders (VAEs) are a type of generative model in the field of machine learning and artificial intelligence. They belong to the broader family of autoencoders but have distinctive features, particularly in generating new data samples.

Question 7

what is associative learning (hopfield networks)?

Answer 7

Associative learning, particularly in the context of Hopfield Networks, refers to a type of artificial neural network that is capable of learning and recalling associations or patterns. Hopfield Networks are recurrent neural networks with the ability to store and retrieve specific patterns or memories.

Question 8

What is a discount factor / gamma?

Answer 8

The discount factor, often denoted by the symbol “γ” (gamma), is a parameter used in reinforcement learning algorithms. It determines the importance of future rewards in the decision-making process of an intelligent agent. The discount factor is a value between 0 and 1.

When an agent makes decisions in a sequential environment, it often receives rewards at each time step. The discount factor allows the agent to weigh immediate rewards against future rewards. A discount factor of 1 means that the agent considers future rewards with equal importance to immediate rewards, while a discount factor less than 1 gives less weight to future rewards.

Question 9

How can a gamma / discount factor of 1 be useful?

Answer 9

(Reinforcement learning)

In practical terms, a discount factor less than 1 encourages the agent to prioritize short-term rewards, making its planning horizon finite. This is often useful in scenarios where the long-term consequences of actions are uncertain or less relevant.
Handling Infinite Horizons:

When dealing with infinite horizons, a discount factor less than 1 ensures that the sum of future rewards converges to a finite value. This is particularly important in mathematical formulations of reinforcement learning algorithms.

Question 10

Question 1: What does it mean when a problem is not linearly separable, and which models are used to address this issue?

a) The problem is easy to solve with linear models.
b) The problem is complex and requires non-linear models.
c) The problem has no solution; no models can address it.
d) The problem is linear, and any model can be applied.

Answer 10

Answer 1: The correct answer is b) The problem is complex and requires non-linear models.

Question 11

Question 2: What is the purpose of the Hebb rule?

a) To suppress neural activation
b) To reinforce neurons with correlated connections
c) To randomize weight changes
d) To activate hidden neurons randomly

Answer 11

Answer 2: The correct answer is b) To reinforce neurons with correlated connections.

Question 12

Question 3: In the delta rule, what does the weight change depend on?

a) The learning rate
b) The input
c) The difference between target and output multiplied by bias
d) The activation function

Answer 12

Answer 3: The correct answer is c) The difference between target and output multiplied by bias.

Question 13

Question 4: In contrastive divergence, what method is effective for generalization?

a) Backpropagation
b) Pruning
c) Simulated annealing
d) Reinforcement learning

Answer 13

Answer 4: The correct answer is b) Pruning.

Question 14

How do Boltzmann Machines utilize hidden units to learn higher-order correlations in the data?

Answer 14

Boltzmann Machines use hidden units to learn higher-order correlations in the data by employing stochastic activations, Gibbs sampling, and simulated annealing during the iterative update process.

Question 15

What is the positive phase in the training of Restricted Boltzmann Machines using contrastive divergence?

Answer 15

The positive phase involves presenting the pattern to the network, clamping it to the visible neurons, and computing the correlations between all visible and hidden neurons.

Question 16

Explain the challenges associated with training Recurrent Neural Networks (RNNs).

Answer 16

Training RNNs is challenging due to the issue of vanishing or exploding gradients over many time steps, making it difficult to effectively backpropagate and update weights.

Question 17

How does backpropagation work, and what is its role in training multilayer neural networks?

Answer 17

Backpropagation is a method to train multilayer neural networks by computing gradients and backpropagating them through all layers, adjusting weights to minimize error. It is crucial for finding optimal weights and minimizing the error in the network.

Question 18

Describe the structure and purpose of Convolutional Neural Networks (CNNs).

Answer 18

CNNs are designed with convolutional and pooling layers to process data in the form of multiple arrays. Convolutional layers detect local conjunctions of features, while pooling layers merge semantically similar features. Stages of convolutional and pooling layers are stacked, followed by more convolutional and fully connected layers.

Question 19

Explain the storage and retrieval process in Hopfield Networks.

Answer 19

Hopfield Networks are fully recurrent neural networks used to store and retrieve data patterns. The storage is performed by gradually changing connection weights using a Hebbian-like learning rule. Retrieval involves iteratively updating the state of neurons until a stable state is reached, assigning higher probability to observed configurations during training.

Question 20

What is the purpose of Simulated Annealing, and how does it work?

Answer 20

Simulated Annealing is a method used to minimize the energy of a system by heating it up and gradually cooling down. It is employed to find the best energy minima and is particularly useful in optimization problems.

Question 21

Describe the concept of Competitive Learning and how it is implemented.

Answer 21

Competitive Learning involves projecting input patterns to a pool of neurons with lateral inhibitory connections. The neuron with the highest activation gradually dominates others through competitive dynamics. The implementation follows a winner-takes-all approach, where the neuron with the highest activation becomes the winner.

Question 22

What is the significance of Self-Organizing Maps (SOMs) in competitive layer structures?

Answer 22

In competitive layer structures, Self-Organizing Maps impose a topological structure where each neuron forms coalitions with its neighbors. This allows for a more accurate mapping of the input space into a lower-dimensional manifold, enabling neurons to compete with distant neurons but cooperate with close neighbors to represent input patterns effectively.

Question 23

How does the learning process of Self-Organizing Maps (SOMs) work?

Answer 23

In the learning process of SOMs, the input vector is compared to the weight vector of each hidden neuron. The neuron with the closest weights is declared the winner. The weights of the winner neuron and its neighbors are adapted, with nearby neurons receiving similar updates. This process helps respond to nearby input patterns.

Question 24

What is the objective of Principal Component Analysis (PCA) in the context of neural networks?

Answer 24

PCA, or Principal Component Analysis, is a statistical technique that aims to find directions of maximum variability in a given dataset. In the context of neural networks, PCA helps discover a set of linearly uncorrelated variables that explain as much variance as possible, achieved through a rotation of the data to maximize variance in new axes.

Question 25

How does Reinforcement Learning differ from other types of machine learning?

Answer 25

Reinforcement Learning is a type of machine learning that relies on rewarding desired behaviors and punishing undesired ones. Unlike other types, a reinforcement learning agent perceives and interprets its environment, takes actions, and learns through trial and error.

Question 26

Explain the dynamics of Boltzmann Machines during the learning process.

Answer 26

In Boltzmann Machines, the learning process involves starting with an input pattern clamped on visible units. Hidden units have random initial activation, and their activation values are iteratively updated until the network reaches equilibrium (Gibbs sampling). Simulated annealing can be employed to enhance convergence towards good local minima, preventing spurious attractors.

Question 27

How does Contrastive Divergence contribute to training Restricted Boltzmann Machines (RBMs)?

Answer 27

Contrastive Divergence is a method used to train RBMs, employing maximum likelihood learning. It minimizes the discrepancy between the empirical (training set) and model distribution. It involves a positive phase where patterns are presented to the network, and a negative phase where new visible activations are generated based on hidden neuron activations from the positive phase.

Question 28

What challenges are associated with training Recurrent Neural Networks (RNNs)?

Answer 28

Training RNNs is challenging because backpropagated gradients either grow or shrink at each time step, leading to issues of vanishing or exploding gradients over many time steps. This makes it difficult to effectively train RNNs.

Question 29

How does Competitive Learning work in neural networks?

Answer 29

In Competitive Learning, input patterns are projected to a pool of neurons with lateral inhibitory connections and excitatory self-connections. Neurons engage in competitive dynamics where the neuron with the highest activation gradually suppresses others, leading to a stable state where only one neuron is active, known as “winner takes all.” This process facilitates the network’s ability to learn and adapt based on competition.

Question 30

What is the purpose of Simulated Annealing in the context of neural networks?

Answer 30

Simulated Annealing is a method used to find the best energy minima in a system. In neural networks, it involves gradually heating up and cooling down a system, aiding in minimizing energy and optimizing network performance.

Question 31

What is the goal of Competitive Learning, and how does it achieve a stable state?

Answer 31

The goal of Competitive Learning is for neurons to compete for activation. In the winner-takes-all scenario, the neuron with the highest activation gradually suppresses others. This competitive process leads to a stable state where only one neuron remains active, facilitating effective learning.