AI Exam 0 - General

AI Practitioner Exam Prep - General

Abbreviations

APIs = Application Programming Interfaces, BI = Business Intelligence

dbs = databases, docs = documents, ex = example, GPU = Graphics Processing Unit, 

IDE = Integrated Development Environment, ISV = Independent Software Vendor

NIST = National Institute of Standards and Technology

OWASP = Open Web Application Security Project

PII = Personally Identifiable Info, SME = Subject matter expert.

Terms

Classification  = data divided into categories. 

Decision trees = CASE statement algorithm. Good for transparency.

Hallucination = incorrect output but presented as factual.

Latency = response time (after sent request).  Real-time applications require low latency.

Modality = the form in which something exists. In AI, this is the data type (ex: image).

Nondeterminism = can repeat with the same inputs, yet have different outputs.

Bias/Variance Terms

Bias = error or gap in data that skews results. types: selection/sampling bias, interaction/participant bias, reporting/measurement bias, confirmation/automation bias, and historical/societal bias.

Confirmation/automation bias = blindly trusting outputs that reinforce existing stereotypes

Data Augmentation = artificially expands data by creating modified versions of existing data.
Data Balancing = solves imbalanced data. See Augmentation.

Historical/societal bias = data reflects past prejudices or outdated values

Interaction/participant bias = users provide dishonest or prejudiced responses

Overfitting = model is too large from memorizing the data (capturing noise and specific patterns that do not recur) and fails to generalize to unseen examples. bias=low and variance=high. fix by showing more training data cases.

Reporting/measurement bias = one group or outcome is over-reported

Selection/sampling bias = data doesn't represent the whole population

Underfitting = too simple and fails to generalize to unseen examples. bias=high and variance=low.

Variance in math = spread of data points.

Variance in ML = sensitivity to noise or overfitting.

Responsible AI Terms

Explainability = explains WHY - AFTER we get the results. Black box. Good for debugging. 

Fairness - ensures equitable outcomes across all demographic groups.

Governance - ensures MAX societal BENEFIT with MIN RISK.

Interpretability = explains HOW makes decisions, BEFORE it does anything. White box. Good with accountability and builds trust. Ex: Partial Dependence Plots found in Sagemaker Clarify. Same as transparency. 

Robustness = operates well despite uncertainty. Same as veracity. 

Transparency - = Same as Interpretability.

Veracity = Same as robustness.