Introduction to AI Flashcards

A sequence of instructions designed to solve a specific problem, such as predicting user preferences.

A method for analyzing data to make conclusions, often without relying on machine learning.

ML where models learn by trial and error to achieve a goal, guided by feedback. Example: Netflix recommendations improving with user interaction.

ML performance depends on the quality of training data.

Intermediate layers in an ANN that process and transform input data. More layers allow for learning complex features.

The final layer in an ANN that generates the result, such as identifying a digit in an image.

Units in each ANN layer responsible for processing information via weights, biases, and activations.

The value a node holds, representing the strength of its output. Higher values indicate stronger signals.

Mathematical factors used in ANN layers to adjust learning and refine predictions.

The interdisciplinary field involving the design, construction, and operation of robots to perform tasks autonomously or with human-like capabilities.

An AI field teaching machines to interpret and analyze visual information from images and videos.

Specialized neural networks designed for image and video processing, identifying patterns and features in layers.

AI models using two competing networks to create and refine realistic images, videos, or text.

AI systems capable of creating new content, such as text, images, videos, audio, or 3D models.

An AI system trained to predict the next word in a sentence based on context.

Neural networks trained on extensive text datasets to generate human-like text, like ChatGPT.

AI models that generate images or videos by refining random noise into detailed visuals, as used by DALL-E.

Generative AI approaches that combine multiple techniques, such as LLMs and GANs, for enhanced content generation.

A field of computer science focused on enabling computers to understand, interpret, and generate human language.

Early systems used grammar rules to process text but were limited by their manual nature.

Introduced probabilistic methods for understanding language by analyzing patterns and word frequencies.

Numerical representations of words or sentences in high-dimensional space, capturing relationships and meanings.

Deep learning models that identify complex patterns for tasks like translation and text generation.

Predict missing words regardless of their position in a sentence. Example: “Water (blank) at 0 degrees.”

Predict the next word in a sequence based on preceding words. Example: GPT models.

Larger training datasets and more parameters lead to increasingly versatile and intelligent models.

Estimate word probabilities based on preceding n-1 words, lacking context for subtle nuances in language.

Neural networks that retain sequential information, improving context understanding but struggling with long text sequences.

Improved RNNs using gate architectures to retain relevant information over longer sequences.

A breakthrough architecture focusing attention on relevant words or tokens, improving scalability and handling long-range dependencies.

Allows transformers to assign importance scores to words, enabling models to focus on crucial parts of input sequences.

Selecting neural network architecture, depth, and parameters to define the model’s capabilities.

Gathering, cleaning, and structuring data for training while addressing ethical concerns like bias and diversity.

Training models on large datasets to create a raw version capable of general tasks.

Includes supervised finetuning and incorporating human feedback to refine the model's performance.

Evaluating the model’s accuracy, speed, and ethical behavior to ensure readiness for end-user deployment.

Guiding models using instructions and examples without modifying their weights.

Attaching a database to the model for expanded context without altering weights, improving contextual learning.

Training an existing model with additional data to update its weights, making it more specialized or efficient.

Massive AI models trained on diverse data, enabling general-purpose capabilities across various domains.

Foundation models that process multiple data types, such as text, images, and videos, extending beyond traditional language tasks.

When AI generates false or inaccurate responses due to errors in prediction or training data.

When AI produces varying answers to the same question, often due to hardware or hosting differences.

Using prompts like "Only respond if you know the answer" or optimizing hardware allocation for consistent outputs.

As LLMs process most publicly available data, new and unique training content becomes harder to source.

A bridge enabling communication between a client (e.g., app) and a server to request and retrieve data.

Store and organize unstructured data (e.g., text, images) as numerical embeddings in high-dimensional spaces.

Indexes vectors to group similar data, enabling fast retrieval and contextual searches (e.g., YouTube video recommendations).

Vector databases give LLMs memory by storing past interactions, enhancing contextual understanding in chatbots and assistants.

Community-driven, freely accessible AI models that foster collaboration and innovation (e.g., Meta’s Llama).

Proprietary models offering plug-and-play convenience, robust APIs, customer support, and stronger data security.

A platform for sharing ML models, datasets, and applications, often called the GitHub of Machine Learning.

Simplifies access to pre-trained models and facilitates creating ML pipelines.

An open-source framework for integrating multiple AI models and external data sources into apps using modular components.

Using one AI model to evaluate the output of another for tasks like coding or open-ended questions.

A professional aligning AI projects with business strategies to maximize impact and ensure effective implementation.

Builds foundational AI models, requiring advanced technical and research skills.

Integrates foundation models into products, optimizing performance with techniques like fine-tuning and RAG.