AI and Machine Learning

Three types of clustering methods

-Hierarchical clustering -Probabilistic clustering -K-means clustering

Define algorithm, parameter, model and epoch

1. Algorithm: Repeatable process used to train a model from a given set of training data 2. Parameter: Internal values inside machine learning that the model derives based on training data 3. Model = algorithm + parameters 4. Epoch: 1 pass through the training data

Name five types of ANN

1. Feed-forward network: --Oldest type. Any unidirectional ANN from input to output 2. Multilayer Perceptron (MLP) --Vanilla neural network 3. Convolutional Neural Network --Example of a deep network --Used for unstructured data, particularly images 4. Recurrent Neural Network 5. Generative Adversarial Network --A pair of deep neural networks. Discriminator network and Generator network. --The generator makes a fake instance, feeds it to the discriminator network that tries to determine if it's real or fake

What is the definition of transparency in the context of AI?

1. For AI developers: Reasons for model's performance are known and understood. 2. For end-users: Sufficient information is published such that the model's performance can be audited

Define hyperparameter, loss function, stochastic gradient descent

1. Hyperparamater: Manually set parameter prior to running algorithm 2. Loss function: Measure of the deviation from the correct output 3. Stochastic gradient descent: Method to reduce loss function

Name three types of machine learning validation methods

1. K-fold cross validation: Method of choice for smaller data. Like a K cup! 2. Leave one out cross-validation: An extreme type of k-fold where n total is equal to the number of instances. 3. Bootstrapping

Name three types of machine learning regression methods

1. Linear regression 2. Multiple linear regression: More than one independent variable 3. Polynomial regression: When independent variables are exponentiated. Best for non-linear data.

Name two types of Reinforcement Learning

1. Markov Decision Process: Use with a known model where states, actions, probabilities of actions generating states, rewards and penalties are evident. 2. Monte Carlo Method: Use when one or more of the elements are unknown. Run through process many times, average outcomes from all previous experiences in a given state.

What is meant by model stability and brittle models?

1. Model stability: The ability of a model to produce similar output over a range of similar inputs, including inputs not previously seen but which are not substantially different from prior inputs. A stable model is a robust model. 2. Unstable (brittle) models: Models that do not produce consistent output when given substantially similar inputs.

Define narrow, general and strong AI

1. Narrow: The machine can perform a single specific task better than a human. 2. General: The machine can perform any intellectual task with the same accuracy as a human. 3. Strong: The machine outperforms humans in many tasks

Define nodes, connections, back-propagation and layers

1. Nodes: Akin to neurons 2. Connections: Akin to synapses 3. Back-propagation: Process where ANN learns whether it made a mistake or not based on output. 4. Layers: Nodes in each layer usually have same activation function. --Input layer: # nodes= # features selected in data --Output layer: # output categories of data --Hidden layer(s) --->Shallow networks usually have 1 --->Deep networks have > 3

What sets should data be separated into during development?

1. Training: Bulk of data 2. Validation: Not always used 3. Testing

What do validation and testing mean in machine learning?

1. Validation: running tests on preliminary (non-final) model. 2. Testing: Final evaluation of a final model where no further changes are expected

What is principal component analysis?

A dimensionality reduction method

What is an artificial neural network?

A machine learning model whose goal is to solve problems the way that a human brain would.

What is a confusion matrix?

A tabulation of the predicted and actual value counts for reach possible class

What is autonomous intelligence

AI is making the decisions (no human in the loop). The term "human-in-the-loop" refers to a human being in the decision-making loop of artificial intelligence tools; that is, AI is not completely autonomous because there is a human in the loop who is monitoring and/or approving or modifying AI decisions before they move forward.

What is augmented intelligence

AI is used to augment and/or assist human in work

What does the level of bias and variance mean for accuracy and fit for a model?

Bias: Accuracy Variance: Precision - High bias: Model is innacurate (underfit). Model is not complex enough. Low bias: Model is accurate (overfit) High variance: Model is imprecise (overfit) Low variance Model is precise (underfit) - Underfit: High bias, low/high variance Overfit: Low bias, high variance

What is a type of sequential ensemble method?

Boosting. Think of two nitro boosted cars driving in squence, one behind the other

Name five types of machine learning classification methods

Classification is a supervised machine learning method where the model tries to predict the correct label of a given input data. In classification, the model is fully trained using the training data, and then it is evaluated on test data before being used to perform prediction on new unseen data.1. Logistic regression: Sepsis prediction models 2. Naive Bayes classifier 3. Support vector machine: See picture 4. K-nearest neighbor: No model (NOT same as K-means clustering). 5. Decision trees: Most common

What is the most common type of machine learning classification method?

Decision trees

What is XAI?

Explainability of how the AI tool works.

What are ensemble methods

Groups of more than one model used together to improve performance. Used on output of high variance (imprecise) and low bias (low inaccuracy)

What is Automation bias?

Human assumes computer is right even when it doesn't make sense

What is Polanyi's Paradox?

Human decision-making is beyond explicit understanding or description. IE; our gut feeling as to why we make a decision

Name the types of AI learning

I. Supervised learning: Goal is to train model to generate known answers, patterns or relationships --Fully supervised: All data labeled to same extent --Semi-supervised: Some data labeled, others not --Weakly supervised: Small amount of data have detailed labels, others don't II. Unsupervised: Data which not been classified or labeled. Goal is for model to discover new patterns or relationships III. Reinforcement: Model is trained on how to reach a complex goal. Ex; IBM Watson on Jeopardy IV. Transfer learning: Bulk of data are from a different domain than the domain intended for the model's use. IE; use domestic animals to train for wild animals. Why? Data is expensive.

Define Instance, Label and Feature

Instance: Row Label: Cell Feature: Column

What is a brittle model?

It does not produce consistent output when given similar inputs. Example of changing very small parts of a picture and changing AI to recognize guacamole vs cat.

What is the goal of AI Reinforcement Learning?

Learn policy (value function) through trial-and-error optimizing long-term reward. Uses unlabeled data like unsupervised learning but uses outcomes to affect model training like supervised learning.

Tell me more about decision trees

Most common supervised machine learning classification method. Model follows branches based on whether or not a threshold is acheived. -Advantages: Each node is easily explainable as if-then cutoffs. -Disadvantages: Susceptible to overfitting (high variance) because small variations in data can cause branches to be created that are not useful. Can be mitigated by preventing nodes to be formed unless statistically significant (ie; pruning the tree)

What is a weak learner?

Performance in prediction is just better than random chance

What is a type of a parallel ensemble method?

Random Forest. Commonly used for 30-day hospital readmission algorithms

What is the difference between machine learning regression and classification?

Regression helps predict a continuous quantity and classification predicts discrete class labels

Expert Systems

Rule-based systems that encode human knowledge in the form of if/then rules. These rules, relationships and ontologies are explicitly coded or programmed into a knowledge base. Machine learning, on the other hand, not based on human knowledge-base or specified rules.

What is the WHO requirement for AI transparency?

Sufficient information regarding use, data used must be published before the design or deployment of an AI technology.

What is irreducible error?

This is due to noise in the data that can be reduced by cleaning up the data.

What are dimensionality reduction methods?

Typically used in pre-processing stage to prepare data for other models. Ranks the importance of dimensions or features to reduce multicolinearity. Principle component analysis is a technique.

Describe under and overfitting

Underfit: Model does not accurately predict output for the data fed to it. Overfit: Model fits training data but not new data. Most common problem. Maybe because too many features per instance (should be 1:10). Maybe training data was too complex.

What is one of the disadvantages of an artificial neural network?

Unravelling the pathways after training is completed can be difficult or impossible. The black box problem

What is a convolution neural net work particularly good at?

Unstructured data, in particular images. Example of a deep network

What are clustering methods?

Unsupervised machine learning algorithms. -Exclusive (hard) clustering - an instance can only belong to one cluster. -Fuzzy (soft) clustering - an instance can have more than 1 cluster assignment

Other names for "Feature" in machine learning

Vector, attribute, dimension

Has the FDA approved an AI device?

Yes, in 2018.