Machine Learning Flashcards
Define Machine Learning.
A computer is said to learn from experience E with respect to a class of tasks T and performance measure P if its performance at tasks in T, as measured by P, improves with experience E.
What’s the difference between top down and bottom up?
Bottom up: give the system lots of data to discover for itself
Top down: model all different functions and wire these together, and write manual instructions for solving a task
Is machine learning top down or bottom up?
Bottom - up
What’s the process of using machine learning to reach AI?
- analyse the task manually
- select appropriate model/s
- Training and testing
- Use the same model to solve unseen tasks
What are the 3 pillars of machine learning?
1.Models and algorithms
2.Computation (powerful and cheaper)
3. Massive data warehouse
Explain the model and algorithm pillar of machine learning.
Doesn’t require knowledge engineers to find all functions and ‘wire’ them in
Explain the data pillar of machine learning.
We have massive amounts of data and the more we can feed into machine learning the more successful it will be.
What is data mining?
Exploration and analysis of large quantities of data to discover valid, novel and useful patterns that are understandable
What is a big difference between data mining and machine learning?
Data mining gives understandable patterns so it can be explained why a conclusion has been come to as opposed to through machine learning
What are the two types of learning in machine learning?
Supervised and Unsupervised Learning
Top-down and bottom-up are also known as?
Deduction (top-down)
Induction (bottom-up)
Describe supervised learning.
Learns from labelled data which consists of an input and a desired output
Describe Unsupervised learning.
Learns from unlabelled data to describe a hidden pattern/structure
e.g. to reach a model or regression or grouping
Name three types of supervised learning.
Linear regression, decision tree, ANN
What is reinforcement learning? What is the goal?
Learning over time based on trial and error / awards from actions. Goal is to learn a process to maximise reward
What are some examples of reinforcement learning?
Deep neural networks, Q-learning, deep Q networks
What are the two main differences between supervised and unsupervised learning?
- supervised has a relationship between Y and f(x), unsupervised usually doesn’t.
- supervised uses a labelled data set, unsupervised uses an unlabelled set
In data cleaning: what is the meaning of imputation?
Replacing missing values with realistic estimated values
What is clustering?
The process of grouping unlabelled data based on a similarity or distance metric
What’s the point of a test data set?
To check how well a trained data set performs
What is regression training?
Finding a function based on given data that minimises its mean squared error to ‘fit’ the samples
What is overfitting?
Fitting a model tightly to a small data set which increases the risk of fitting to an anomaly or noise data
What are 3 pros of linear regressions models?
- short training time
- easy to interpret
- easy to implement
What are two cons of linear regression models?
- sensitive to noise and outliers
- cannot handle complicated relationships
What is classification learning?
Learning to predict which set an instance belongs to based on pre-labelled instances
In the decision tree model, what is an internal node?
A decision point where a test is made on an attribute of the data and a decision is made
In the decision tree model, what is a branch?
The course of a decision or action
In the decision tree model, what is a leaf node?
A predicted class label / output reached through a path of decision points (internal nodes) and branches
What are 4 pros of decision tree?
- Reasonable training time
- Can handle large number of features
- Easy to implement
- Easy to interpret
What are 4 cons of decision tree?
- Only simple decision boundaries
- Problems with lots of missing data
- Cannot handle complicated relationships
- Over- complex tree: potential for overfitting
When we only have one dataset, why do we need to split it?
To have data for training and data for testing
What stage happens before the training set is fed into the machine learning method?
Data pre-processing
What is a confusion matrix?
It represents how much data is predicted to each outcome and how many of those are predicted correctly and incorrectly
What is the concept of perceptron in Neural networks?
There is a set of weighted inputs that if they surpass a certain amount, it causes a set of outputs.
What is the differing meaning of positive and negative weights for neural network inputs?
- a positive weight encourages the neuron to fire as a result of that input
- a negative weight discourages the neuron from firing as a result of that input
What are the 4 logic functions?
- Or
- And Not
What does the logical OR function do in ANNs?
one or other of the inputs needs to be one
What is the logical function ‘And Not’ in ANN?
A AND NOT B
means A must have a value of 1 and B a value of 0
