Worksheetsiav ML Study Group Pop Quiz 1
Total questions: 15
Worksheet time: 30mins
Put the three items "Artificial Intelligence", "Machine Learning" and "Deep Learning" in order from broadest to most specific.
Machine Learning encompasses AI which encompasses Deep Learning
Deep Learning encompasses Machine Learning which encompasses AI
Machine Learning encompasses Deep Learning which encompasses AI
AI encompasses Machine Learning which encompasses Deep Learning
Please check all the boxes which are types of machine learning.
Robotic
Unsupervised
Programmatic
Supervised
Analytic
Those model factors which are extrinsic to (NOT included in) the data are called:
Classes
Hyperparameters
Externalities
Adjustment factors
The modelling approach which yields continuous, numerical values such as stock prices, box office receipts or (x,y) co-ordinates is known as:
(a)
The modelling approach which yields discrete, qualitative results such as next word in a sentence, positive/negative facial recognition or category for a customer to churn is:
(a)
To make a successful classification model we will need data which has:
Features that can be quantified
Labels that are unknown
Labels that are known
A broad variety of classes
A method to measure similarity
Why do we square each of the error terms when computing the full Mean Squared Error of a model?
Squaring is more accurate than cubing
Subtracting the true observation might give a negative result, so we square them to get a positive term
When we divide by the m number of observations, Euclidean distances increase for dimensional spaces greater than 5
Squaring the error terms individually leads to larger values, which helps improve the model's accuracy.
Identify all components of the Coefficient of Determination (also called R-squared) below:
Sum of Squared Error (SSE)
Variation of Square Error (VSE)
Mean Squared Error (MSE)
Total Sum of Squares (TSS)
Which of the following is NOT an appropriate cross-validation approach?
K-fold Cross Validation
Stratified Cross Validation
Vector Re-alignment Cross Validation
Leave One Out Cross Validation
We can use polynomial features in a regression model:
if the model also has classification properties
to increase robustness
to drive the r-squared value closer to 1
to capture non-linear effects
The higher the degree of a polynomial regression, the more complex the model and:
at lower degrees we see visual signs of bias, the predictions are too rigid to capture the curve patterns of the data
at higher degrees we see visual signs of bias, the predictions are too rigid to capture the curve patterns of the data
at higher degrees we see visual signs of variance, the predictions fluctuate wildly because of model hypersensitivity
at lower degrees we see visual signs of variance, the predictions fluctuate wildly because of model hypersensitivity
The three sources of error for our models are:
entropy
reductive error
bias
variance
irreducible error
We regularize features because
Trick question: We don't regularize features
it penalizes high-valued coefficients, reduces parameters and shrinks the model, reducing complexity.
Irregular features cannot be represented with real numbers
it penalizes low-valued coefficients, expands parameters and grows the model, reducing simplicity
Why does the level of support in Supervised ML models vary between interpretation vs prediction?
Sometimes the dramatically increased complexity required for very accurate predictions makes the model almost impossible to interpret
It does not, and you can get easily-interpreted models which make very accurate predictions given a skilled-enough programmer
This is a natural feature of the Python programming language
Why split data into training, test and validation sets? Why not just train and test and validate on the whole dataset as-is?
We need to split because the model should be tested against new, unseen data, not just the data it trained on.
We need to split because the testing and validations steps are also training steps too, so they improve model accuracy
We need to split because the fork() system call in UNIX is more highly performant the more highly parallelized our code is
We need to split because machine learning models can be very taxing computationally, and it is a safe way to avoid shortage of memory errors
