COSI-101: Introduction to Artificial Intelligence: Final Exam

What are the three limitations of network embedding?

1. Lack of Parameter Sharing: Each node's embedding is learned without leveraging shared parameters. 2. Inability to Forecast Unseen Data: They are based on static snapshots of the network. 3. Inability to Utilize Node Features: There is no way to leverage node features.

Biconditional Elimination

A SLE which allows a biconditional to be split into two implications. P↔Q is equivalent to (P→Q)∧(Q→P)

Double-negation Elimination

A SLE which allows a double negation to be removed from a proposition. ¬(¬P) is equivalent to P

Contraposition

A SLE which allows an implication to be replaced with its contrapositive. P→Q is equivalent to ¬Q→¬P

Implication Elimination

A SLE which allows an implication to be rewritten as a disjunction. P→Q is equivalent to ¬P∨Q

Distributivity of ˄ over ˅

A SLE which allows one to distribute a conjunction across a disjunction. P∧(Q∨R) is equivalent to (P∧Q)∨(P∧R)

Distributivity of ˅ over ˄

A SLE which allows one to distribute a disjunction across a conjunction. P∨(Q∧R) is equivalent to (P∨Q)∧(P∨R)

De Morgan's Law

A SLE which allows one to distribute negations across conjunctions and disjunctions. ¬(P∧Q) is equivalent to ¬P∨¬Q and ¬(P∨Q) is equivalent to ¬P∧¬Q

Associativity of ˄

A SLE which allows one to group conjunctions in any order. (P∧Q)∧R equivalent to P∧(Q∧R)

Associativity of ˅

A SLE which allows one to group disjunctions in any order. (P∨Q)∨R equivalent to P∨(Q∨R)

Commutativity of ˄

A SLE which allows one to swap the order in conjunctions without changing the outcome. P∧Q equivalent to Q∧PQ∧P

Commutativity of ˅

A SLE which allows one to swap the order in disjunctions without changing the outcome. P∨Q equivalent to Q∨P

Implication (Þ)

A binary operator indicating that if the first proposition (antecedent) is true, then the second proposition (consequent) must also be true. It is false only when the antecedent is true and the consequent is false.

Disjunction (˅)

A binary operator that combines two propositions and is true if at least one of the propositions is true.

Conjunction (˄)

A binary operator that combines two propositions and is true only if both propositions are true.

Bi-conditional (iff, Û)

A binary operator that is true if both propositions have the same truth value, either both true or both false. It represents the phrase "if and only if".

Propositional Logic

A branch of logic that deals with propositions (statements that can be true or false) and their combinations using logical connectives. It allows for the construction of complex sentences and reasoning about their truth values.

Decision Networks

A classification of networks which combine Bayesian networks with additional nodes for actions and utilities. They include chance nodes (random variables), decision nodes (choices of actions), and utility nodes (agent's utility function).

k-means Clustering

A clustering algorithm that partitions data into k distinct clusters based on feature similarity. It assigns data points to the nearest cluster center and adjusts centers based on the mean of the points.

Actuators (A)

A component of the PEAS framework which defines the components through which the agent affects its environment. They are the means by which the agent takes action.

Environment (E)

A component of the PEAS framework which defines the context or surroundings in which the agent operates. It includes everything that the agent interacts with and needs to consider while making decisions.

Performance Measure (P)

A component of the PEAS framework which defines the criteria for success in an agent's environment. It is a metric used to evaluate how well an agent is achieving its goals or objectives.

Sensors (S)

A component of the PEAS framework which defines the way in which an agent is able to perceive its environment. They are the means through which an agent gathers information about its surroundings to make informed decisions.

Search Algorithm

A computational process that takes as search problem as input and returns either a solution or an indication of failure.

Conjunctive Normal Form (CNF)

A concept in propositional logic where every sentence is equivalent to a conjunction of clauses.

Graph (Network)

A data type used to determine complex systems of interacting entities. It is used to model relationships in various domains, such as social networks and biological networks.

Recurrent Neural Network (RNN)

A family of neural networks designed for processing sequential data, capable of handling sequences of variable length. It is suitable for tasks involving time series data, speech recognition, and language modeling.

Machine Learning

A field of AI where algorithms learn from data. Algorithms improve performance at specific tasks by learning through experience.

Four-phase Problem-solving Process (FPP)

A framework used in artificial intelligence to guide an agent in solving problems through systematic steps.

Heuristic Function

A function in search algorithms that estimates the cost or distance from a node to the goal state, guiding the search towards the goal efficiently. It calculates a value (h(n)) indicating how close a state is to the goal, influencing the search algorithm's efficiency.

Bayesian Network

A graphical model representing a set of variables and their conditional dependencies via a directed acyclic graph, with conditional probabilities attached to each node.

Multi-Task Learking

A learning paradigm in machine learning where multiple learning tasks are solved at the same time, using a shared representation. It improves generalization by leveraging domain-specific information contained in the training signals of related tasks.

Probability

A measurement of the likelihood of a specific outcome or event in effort to quantify uncertainty.

Linear Regression

A method to predict a continuous outcome variable based on one or more predictor variables.

Learning Decision Trees

A model in the form of a tree structure that represents decisions and their possible consequences. It uses information gain and entropy to select the best attributes for splitting the data.

Multi-hop Similarity

A network embedding method that considers multi-hop (k-hop) neighbors of a node for determining similarity.

Adjacency-based Similarity

A network embedding method to determine similarity based on the edge weights between nodes in a network.

Random-walk Optimization

A network embedding method to refine node embeddings in a network.

Random-walk Approach

A network embedding method which involved running short random walks from each node and collecting the multisite of visited nodes. It helps to understand the node similarities and network structure.

Execution

A phase in the FPP in which the agent executes the actions in the found solution sequence, one at a time.

Search

A phase in the FPP in which the agent simulates sequences of actions in its model, searching for a sequence that achieves the goal. This sequence is considered the solution.

Goal Formulation

A phase in the FPP in which the objectives the agent needs to achieve are defined.

Problem Formulation

A phase in the FPP in which the states and actions necessary for the agent to reach its goal are defined.

Bayes Rule

A principle used for revising existing predictions or theories (posterior probabilities) in light of new or additional evidence. P(A∣B)=P(B∣A)P(A)/P(B)

Node

A representation of a state, in the context of a search tree.

Edge

A representation of an action that leads to the next state, in the context of a search tree.

Complex Sentences

A sentence that is built from simpler sentences using logical connectives like negation (¬), conjunction (˄), disjunction (˅), implication (Þ), and bi-conditional (iff, Û).

Conditional Independence

A situation where two variables are independent of each other given a third variables, which means the occurrence of one variable does not affect the probability of the other variable once the third variable is known or given.

Greedy Best-First Search

A specific type of best-first search that prioritizes expanding the node with the lowest heuristic value (h(n)), which appears to be closest to the goal.

Reached

A state of a node indicating that it has been expanded.

Search Tree

A structure that forms various paths from the initial state, with the goal of finding a path that reaches the desired goal state.

Deep Learning

A subset of machine learning where neural networks learn from a large amount of data. It differentiates itself from traditional machine learning by its emphasis on learning data representations and feature engineering.

Logic

A systematic method of reasoning that involves constructing and analyzing arguments based on rules. It is used in AI to formalize reasoning, enabling agents to deduce new information and make decisions based on given facts and rules.

Neighborhood Aggregation

A technique where nodes aggregate information from their neighbors and apply a neural network to the aggregated data. It is used in GNNs to update node embeddings based on neighborhood information.

Simple Reflex Agents

A type of agent that chooses actions solely based on the current percept, disregarding the history of past percepts. These agents react to immediate input without considering the broader context of their actions.

Graph Neural Network (GNN)

A type of neural network that extends traditional deep learning methods to graph data. It can handle more complex graph structures for tasks like node classification and link prediction.

Feed-forward Neural Networks

A type of neural network where connections between nodes do not form cycles. It includes perceptrons and multi-layer perceptrons, used for various machine learning tasks.

Convolutional Neural Networks (CNNs)

A type of neural network which is particularly effective at processing data with a grid-like topology, such as images. It uses convolutional layers to capture spatial feature, making them powerful for tasks like image recognition.

Negation (¬)

A unary operator that inverts the truth value of a proposition. If a statement is true, its negation is false, and vice versa.

Human Agent

An agent which uses organs like eyes and ears as sensors, and hands, legs, and vocal tract as actuators.

Robotic Agent

An agent which utilizes devices like cameras and infrared range finders for sensing, and various motors for actions.

Problem-solving Agent

An agent whose purpose is to find a sequence of actions that leads to a desirable goal state. The computational process it undertakes to achieve this is known as a search.

Bagging (Bootstrap Aggregating)

An ensemble method in machine learning that trains multiple models and aggregates their results. It reduces variance and helps to avoid overfitting, but typically involves the creation of multiple datasets through bootstrap sampling.

Agent

An entity that perceives its environment through sensors and acts upon that environment through actuators.

A* Search

An informed search strategy characterized by the function f(n) = g(n) + h(n), where g(n) is the path cost from the initial state to the node, and h(n) is the estimated cost of the shortest path from the node to a goal state.

Unary Operator

An operator that operates on only one operand or argument. In logic, it modifies the truth value of a single proposition.

Binary Operator

An operator that operates on two operands or arguments. In logic, it is used to combine the truth values of two propositions.

Gradient Descent

An optimization algorithm that iteratively moves towards the minimum of a function. It is used in machine learning to find the best parameters for minimizing a cost function.

Semantics

Defines the meaning or truth of sentences in a given world, interpreting and assigning meaning to each sentence.

Transition Model

Describes how each action changes the state.

Spatial Features

Features that represent the spatial arrangement and relationships of elements in data, often in two or three dimensions. They are crucial in tasks like image and video analysis where the position and arrangement of pixels are significant.

Antecedent

In a conditional statement, it is the first proposition or clause. It represents the "if" part of an "if-then" statement.

Consequent

In a conditional statement, it is the second proposition or clause. It represents the "then" part of an "if-then" statement.

Underfitting

Occurs when a model cannot capture the underlying trend of the data, often due to its simplicity or lack of sufficient parameters. It is an indication of a need for increased model complexity or better feature engineering.

Overfitting

Occurs when a model learns the training data too well, including noise and outliers, at the expense of generalizing to new data. Regularization and cross-validation are common strategies to combat overfitting.

Low Capacity

Refers to a model with limited ability to fit complex functions, usually due to fewer parameters or lower complexity. These models might struggle to capture all relevant patterns in the data, leading to underfitting.

High Capacity

Refers to a model's ability to fit a wide variety of functions due to having a large number of parameters or high complexity. While these models can capture complex patterns in the data, they risk overfitting if not properly regularized.

Standard Logical Equivalence (SLE)

Refers to commonly accepted equivalences between logical expressions that hold true in all cases. These equivalences are used as tools to simplify or transform logical expressions, especially in proofs and problem-solving.

Logical Agents in Games

Refers to the application of logic in creating agents that operation within game environments, utilizing logical rules and strategies.

Cycles (in Neural Networks)

Refers to the presence of loops in the network architecture, where outputs of some neurons are fed back as inputs to others. They are commonly seen in recurrent neural networks (RNNs), allowing them to maintain a form of 'memory' of previous inputs in the sequence.

Agent Programs

Software that dictates how an agent behaves. The challenge in AI is to develop agent programs that produce rational behavior efficiently, without relying on large, exhaustive lookup tables.

Rational Agents

Systems that can be considered intelligent due to their ability to make decisions based on their perceptions and goals.

Task Environment

The context in which an agent operates, described using the PEAS (Performance, Environment, Actuators, Sensors) framework. This framework helps in defining the nature of the environment and the agent's objectives within it.

Graph Representation Learning

The goal is to learn task-independent feature representations in graphs. It focuses on learning how to represent graph structures for effective analysis and prediction.

When should the gradient descent method be used?

The gradient descent method should be used for optimization problems where the goal is to minimize a cost function. It's particularly useful in machine learning for finding the optimal parameters of a model, like in linear regression or neural networks.

Total Probability

The principle that the sum of the probabilities of all mutually exclusive and exhaustive outcomes of a random variable equals one.

Conditional Probability

The probability of an event given that another event has occurred. P(X=x∣Y=y)=P(x∧y)/P(y) If X and Y are independent, then P(x∣y)=P(x)

Marginal Probability

The probability of an event irrespective of the outcomes of other variables. P(x)=Σy​P(x∧y)

Joint Probability

The probability of two events occurring together. For independent variables X and Y, P(x∧y)=P(x)P(y)

Exact Inference in Bayesian Network

The process of calculating the exact probability distribution for a set of variables with a Bayesian network. Although theoretically accurate, it's often computationally expensive or impractical for large networks.

Approximate Inference for Bayesian Networks

The process of estimating the probability distribution for a set of variables with a Bayesian network. It is often employed with randomized sampling algorithms, like Monte Carlo algorithms, to provide approximate answers where the accuracy depends on the sample size.

Feature Engineering

The process of selecting, modifying, and creating features that are used to train machine learning models. It involved using domain knowledge to create feature that make machine learning algorithms work better.

Expand

The process of taking a node and generating new nodes, known as child nodes or successor nodes, based on the available actions for that state. Each child node considers the current node as its parent.

Contrapositive

The result of the negation of both the antecedent and consequent and then reversing them.

Syntax

The rules for constructing well-formed sentences in a representation language, specifying which symbols and combinations are allowed.

Frontier

The set of all nodes that have been generated but not yet expanded.

Perceptrons

The simplest type of artificial neuron that computes a weighted sum of its input and applies an activation function. It is used as a basic building block in neural networks, capable of performing simple classifications.

What is the reason behind the name backpropagation?

The term "back propagation" refers to the method of training artificial neural networks through backward propagation of errors. It is named so because the error is calculated at the output and distributed backwards through the network's layers. This process involves adjusting the weights of the network in a way that minimizes the error.

Informed (Heuristic) Search Strategies

These strategies use domain-specific knowledge (heuristics) to more efficiently find solutions compared to uninformed strategies. They involve a heuristic function that provides an estimate to guide the search.

Logical Equivalence

Two propositions or statements are logically equivalent if they always have the same truth value in every possible scenario. They are used to determine if two different expressions essentially express the same thing.