AIML 131 mid term

Question 1

What is overfitting in the context of supervised ML?

Answer 1

When the model is too complex and learns too much detail.

Question 2

What are error surfaces in neural networks?

Answer 2

Represent the error based on different weight combinations.

Question 3

What is the concept of ‘feature space’ in ML?

Answer 3

The space of all possible combinations of input features.

Question 4

How are modern neural networks organized?

Answer 4

Organized into layers, with input and output layers.

Question 5

What is the purpose of dimensionality reduction models in unsupervised learning?

Answer 5

Simplifying the representation of data.

Question 6

What are some examples of supervised ML models for regression?

Answer 6

Linear regression, polynomial regression

Question 7

What is gradient descent in the context of neural networks?

Answer 7

Optimization method to minimize the error by adjusting weights.

Question 8

How does DBSCAN clustering work?

Answer 8

Identifying dense regions based on radius and point count.

Question 9

How do reinforcement learners improve over time?

Answer 9

By trying outputs at random and receiving feedback to adjust actions.

Question 10

How are input words represented in LLMs?

Answer 10

By creating word embeddings that cluster similar meanings.

Question 11

How does the attention mechanism address the bottleneck problem in LLMs?

Answer 11

It provides direct access to important input words for each output word.

Question 12

What is ‘parameter space’ in ML?

Answer 12

The space of all possible combinations of model parameters.

Question 13

How do transformers improve upon the attention mechanism?

Answer 13

By replacing recurrent networks with self-attention mechanisms.

Question 14

How does reinforcement learning differ from supervised learning?

Answer 14

In reinforcement learning, the algorithm learns behaviours based on feedback.

Question 15

How does supervised learning work?

Answer 15

Machine learns to map inputs to outputs with labeled examples.

Question 16

What role do rewards and punishments play in reinforcement learning?

Answer 16

They guide the learner to improve its actions.

Question 17

How do you evaluate a regression model?

Answer 17

Using error measures like SSE, MSE, MAE, RMSE.

Question 18

What is SSE, MSE, MAE, RMSE?

Answer 18

Sum Squared Error, Mean Squared Error, Mean Absolute Error, root mean squared error

Question 19

What is the bottleneck problem in recurrent networks?

Answer 19

They remember recent words better and struggle with long dependencies.

Question 20

How does k-means clustering work?

Answer 20

Choosing k cluster centroids, assigning items, and recomputing centroids.

Question 21

What are some key concepts related to language models (LLMs)?

Answer 21

Predicting the next word, learning from large corpora, self-supervision.

Question 22

What are clustering models used for in unsupervised learning?

Answer 22

Identifying different types of input.

Question 23

What is principal component analysis (PCA)?

Answer 23

Algorithm to reduce the dimensions of data while preserving variance.

Question 24

How do you evaluate a classifier model?

Answer 24

Using accuracy, error rate, and confusion matrix for binary classifiers.

Question 25

What are recurrent networks and how are they used in LLMs?

Answer 25

Networks that remember earlier words in sequential data.

Question 26

What are transformers in the context of LLMs?

Answer 26

Models that use self-attention mechanisms for encoding and decoding text.

Question 27

What are word embeddings used for in LLMs?

Answer 27

To represent words in an n-dimensional space.

Question 28

How are neural networks structured?

Answer 28

Organized into layers of neuron-like units with connections.

Question 29

How does an autoencoder neural network work for dimensionality reduction?

Answer 29

It learns to map each input back onto itself through a bottleneck layer.

Question 30

What are some examples of supervised ML models for classification?

Answer 30

k-nearest neighbour, decision trees

Question 31

What is underfitting in the context of supervised ML?

Answer 31

When the model is too simple and doesn’t learn real structure.

Question 32

What are word representations in LLMs?

Answer 32

Words represented in a one-hot scheme with probabilities for the next word.

Question 33

How does training a neural network work?

Answer 33

Adjusting weights to improve performance using training data.

Question 34

How do unsupervised ML models differ from supervised ML models?

Answer 34

They do not have labeled target outputs in training data.

Question 35

How do neural networks learn to predict the next word in natural language text?

Answer 35

By training on large corpora, predicting probability distributions.

Question 36

What is reinforcement learning?

Answer 36

Algorithm learns good outputs through trial and error with rewards and punishments.