DAT566: Module 7 Classical And Modern AI

What are generative models?

Models which can generate new data instances

Deep generative models use neural networks to do so

What are Large Language Models?

Deep leaning models designed to generate and manipulate human language

LLM’s are trained to generate text by predicting the next token given all previous tokens

A mechanism called attention improves prediction by learning relationships between tokens

What is the main mechanism used by Transformer-based LLMs to generate text?

Autoregressive next-token prediction

How do vector databases answer queries?

By measuring the distance between embeddings of queries and documents in the database

What are commonly used training phases for LLMs?

Pre-training on a large text corpus to learn general language patterns

Task-specific fine-tuning on labeled data

Reinforcement learning with human feedback (RLHF) to refine model behavior

What is Retrieval Augmented Generation (RAG)?

In-context learning is when we include relevant data in the prompt

RAG is how we decide what additional information to put in the prompt

What are the three main components of a transformer?

Encoder: Turns input sequence (i.e. sentences) into embeddings

Decoder: Generates output sequences (i.e translated sentences) from embeddings

Positional encoding: Tells the model about the order of the tokens

What is a Vision Transformer (ViT)?

A breakthrough method of image visualization compared to CNNs

ViT splits the image into patches treating each patch as a token

ViT includes position encoding to indicate where each patch is located

What is the main idea behind diffusion models?

Denoising images by reversing a noise process

What is the main idea of transfer learning?

Using knowledge from one task to improve another

What is the main idea of multi-task learning?

By training a model to solve multiple related tasks at the same time, you can improve its sample efficiency and prediction accuracy compared to training multiple separate models.

What is contrast learning?

A form of supervised learning used to teach a model to distinguish between similar and dissimilar data points

What is representation learning?

Learning meaningful embeddings that capture the underlying structure of the data

Broad concept used to learn useful data representations

What does GOFAI (Good Old-Fashioned AI) primarily rely on?

Symbolic reasoning and rules

What is Neuro-symbolic AI?

A combination of classical and modern AI paradigms

Learn to generate potential solutions then check and accept/reject the proposed solutions

What is the Rosenblatt Perception?

Earliest example of a neural network model

Limited: His algorithm for training perceptions only worked for a single layer of perceptions

Which event marked a major resurgence in AI interest due to deep learning?

AlexNet winning the ImageNet competition in 2012

What does the term “AI winter” primarily refer to?

A period in the history of AI characterized by reduced interest and funding in AI