AI Practitioner flashcards

AI: machines acting smart; ML: systems learning from data

labeled data, predicting outcomes, like grading papers

unlabeled data, finding patterns, like sorting toys

agent learns by trial and error, rewards and penalties

categorizing data, yes/no, spam/not spam

predicting continuous values, house prices, test scores

grouping similar data, customer segments

layers of nodes, inspired by brain, deep learning

neural networks with many layers, complex patterns

data used to teach the model, like practice problems

tuning model during training, like a quiz during class

evaluating final model performance, like a final exam

creating new features, improving model accuracy

model memorizes training data, performs poorly on new data

model too simple, can't capture data patterns

model reflects biases in training data, unfair outcomes

ensuring equitable outcomes, reducing bias

understanding how AI makes decisions, transparency

creates new content, text, images, code

NLP, computers understanding human language

CV, computers 'seeing' and interpreting images

model makes predictions on new data

set of rules for problem-solving or calculation

settings for the learning process, not learned from data

measures model error, guides learning

optimizing algorithm, finding minimum of loss function

cleaning and transforming data before model training

stores data as embeddings, efficient similarity search

numerical representation of text or image, meaning captured

neural network design, attention mechanism, GenAI backbone

focuses on important parts of input, improves context

LLM, trained on vast text data, generates human-like text

creates new content: text, images, code, music

large pre-trained model, adaptable to many tasks

crafting inputs to get desired GenAI outputs

model performs task without specific examples

model performs task with a few examples

adapting a pre-trained model for a specific task/dataset

GenAI generates plausible but false information

piece of word or character, LLM's basic unit

RLHF, human preferences guide model training

common for sequence-to-sequence tasks, translation

generates output based on previous tokens, LLMs

GPT, decoder-only, predicts next token

RAG, combines GenAI with external data retrieval

initial broad training on large dataset

model learns from examples directly in the prompt

reflects biases in training data, unfair outputs

linking multiple prompts for complex tasks

data generated artificially, not from real-world events

controls randomness/creativity of GenAI output

controls diversity by choosing from top probable tokens

malicious input to override GenAI instructions

GAN, generator vs

processes and generates multiple data types: text, image

number of values in model, indicates complexity/size

XAI, understanding why AI makes certain decisions

filtering harmful or inappropriate GenAI outputs

ensuring GenAI outputs match human values/intent

stores data as embeddings, efficient similarity search

numerical representation of text or image, meaning captured

neural network design, attention mechanism, GenAI backbone

focuses on important parts of input, improves context

LLM, trained on vast text data, generates human-like text

creates new content: text, images, code, music

large pre-trained model, adaptable to many tasks

crafting inputs to get desired GenAI outputs

model performs task without specific examples

model performs task with a few examples

adapting a pre-trained model for a specific task/dataset

GenAI generates plausible but false information

piece of word or character, LLM's basic unit

RLHF, human preferences guide model training

common for sequence-to-sequence tasks, translation

generates output based on previous tokens, LLMs

GPT, decoder-only, predicts next token

RAG, combines GenAI with external data retrieval

initial broad training on large dataset

model learns from examples directly in the prompt

reflects biases in training data, unfair outputs

linking multiple prompts for complex tasks

data generated artificially, not from real-world events

controls randomness/creativity of GenAI output

controls diversity by choosing from top probable tokens

malicious input to override GenAI instructions

GAN, generator vs

processes and generates multiple data types: text, image

number of values in model, indicates complexity/size

XAI, understanding why AI makes certain decisions

filtering harmful or inappropriate GenAI outputs

ensuring GenAI outputs match human values/intent

search by meaning, not just keywords, uses embeddings

condensing long text into shorter, key points

GenAI creating programming code based on prompts

GenAI creating images from text descriptions

AI for conversational interfaces, customer service

GenAI assisting with stories, poems, scripts

AI treats all groups equitably, avoids bias

who is responsible for AI system actions

understandable AI decisions, clear explanations

protecting personal data used by AI

AI systems are protected from attacks, misuse

AI performs reliably, handles unexpected inputs

rules and processes for AI development, deployment

steps to reduce unfairness in AI models

safeguarding user information, anonymization

making AI decisions interpretable to humans

people monitor, intervene in AI operations

inputs designed to trick AI, cause errors

AI performance degrades over time, data changes

building ethical considerations into AI from start

rules for using AI to prevent harm, promote good

AI systems meet legal and regulatory standards

identifying potential harms, likelihood from AI

ensuring AI systems operate without causing harm

AI goals match human values, societal benefit

AI treats all groups equitably, avoids bias

who is responsible for AI system actions

understandable AI decisions, clear explanations

protecting personal data used by AI

AI systems are protected from attacks, misuse

AI performs reliably, handles unexpected inputs

rules and processes for AI development, deployment

steps to reduce unfairness in AI models

safeguarding user information, anonymization

making AI decisions interpretable to humans

people monitor, intervene in AI operations

inputs designed to trick AI, cause errors

AI performance degrades over time, data changes

building ethical considerations into AI from start

rules for using AI to prevent harm, promote good

AI systems meet legal and regulatory standards

identifying potential harms, likelihood from AI

ensuring AI systems operate without causing harm

AI goals match human values, societal benefit