Final Stuff 2

Question 1

what is LIME?

Answer 1

a technique that aims to create an interpretable model local to a data point

Question 2

why does the vanishing/exploding gradient problem occur?

Answer 2

we are multiplying gradients as we move to earlier layers, multiplying lots of small/large numbers together

Question 3

what is grokking?

Answer 3

when test error falls way after training error has fallen, model suddenly generalizes well.

Question 4

what are support vectors in SVM

Answer 4

the data points that are closest to the hyperplane

Question 5

how to set up confusion matrix

Answer 5

actual on top, predicted on the side
positive positive in the top left

Question 6

what is the loss function for logistic regression?

Answer 6

binary cross entropy, it is strictly convex

Question 7

types of transformers

Answer 7

encoder, decoder, encoder-decoder

Question 8

false positive rate

Answer 8

FP / (FP + TN)
second column

Question 9

2 examples of non parametric models

Answer 9

knn, decision trees
don’t make assumptions about the underlying distribution

Question 10

recall/true positive rate

Answer 10

also called true positive rate
TP / (TP + FN)
out of all the positives, how many was the model able to get correct (recall)
first column

Question 11

false positive

Answer 11

we predict positive, but that is false (actually negative)

Question 12

advantages of MSE and MAE

Answer 12

MSE - differentiable, good for learning
MAE - result is interpretable, simple, less prone to outliers

Question 13

what is calibration?

Answer 13

making sure the output of the model represents how confident it is

Question 14

three svm kernels

Answer 14

linear, polynomial, RBF

Question 15

what are sampling layers in cnns?

Answer 15

pooling layers

Question 16

types of autoencoders

Answer 16

deep (multiple layers), sparse, variational, etc.

Question 17

precision

Answer 17

TP / (TP + FP)
how many predicted positives were truly positive?
across first row

Question 18

accuracy

Answer 18

correct vs total
diagonal vs total

Question 19

what are scaling laws?

Answer 19

take in model size and data and try to predict loss
how much will throwing resources at the model improve it?

Question 20

MSE and MAE formulas

Answer 20

true - predicted, square/take absolute value, sum up, divide by n

Question 21

normalization vs regularization

Answer 21

normalization - making sure weights are same scale
regularization - makes sure model doesn’t overfit

Question 22

false negative

Answer 22

we predict negative, but that is false (actually positive)

Question 23

types of normalization

Answer 23

batch norm, L2 norm

Question 24

what is out of bag evaluation

Answer 24

evaluate tree on the data that wasn’t used for training

Question 25

high level of how to train an LLM

Answer 25

pretraining - predict next token supervised fine tuning - train on prompts and good responses reinforcement learning with human feedback - humans rate responses

Question 26

why use kernels in SVM?

Answer 26

transform the data into a higher dimensional space where the data is now linearly seperable

Question 27

what are proxy models?

Answer 27

models that behave similar to complex models