Gen AI

Large Language Models

A class of Foundation Models, focused on language based tasks like summarization, text generation,

Generative AI's ability to respond to human conversations allows the creation of...

AI-powered chatbots, voice bots, and virtual assistance

How will Generative AI be used to impact healthcare industry

Accelerate drug research, design synthetic gene sequences, and synthetic patient/healthcare data

How will Generative AI be used to impact financial services

Allows for reduced cost by improving customer services, speed up loan approvals, detect fraud, and give personalized financial advice

How does RAG work?

An informational retrieval component is introduced that takes the user input to pull info from a new data source

Best practices in Generative AI

Begin with optimizing internal applications, enhance transparency, implement security, and test extensively

Transformer-based models

Build upon encoder/decoders, and add more layers to the encoder to improve performance on text-based tasks

How do generative AI models work?

Calculate the probability of known and unknown factors occurring together, learning the distribution of the data and the relationships between the data points

Why are Large Language Models important?

Can perform different tasks like answering questions, summarizing documents, translate languages, and complete sentences

What are examples of things that Generative AI can create

Conversations, stories, images, videos, music, etc.

Applications of large language models

Copywriting, Knowledge based answering, text clarification, code generation, and text generation

Benefits of RAG

Cost-effective implementation, Current information, enhanced user trust, and more developer control

Steps of RAG

Create external data outside original training set, retrieve relevant information using queries, add relevant data to user input in context, and update external data

Diffusion Models

Create new data by making random changes to initial data set, changes often called noise. The changes are controlled and subtle

How did large language models initially work?

Each word was placed in a numerical table specifically for that word

What two neural networks do Large Language Models utilize

Encoder and Decoder

How will Generative AI be used to Media and entertainment

Enhance art with AI generated, personalize content/ads, and games can become more personal

What can Generative AI do?

Explore and analyze complex data, discover new trends/patterns, and summarize content

Fine-tuning

Extension of few-shot learning, provide data relevant specifically to the application

What can Generative AI do to optimize business practices

Extract/summarize data, evaluate/optimize scenarios, and generate synthetic data

What improves as GAN continues

Generator continues to make more realistic fake data, while discriminator becomes better at telling the difference between real and fake data

How will Generative AI be used to impact automotive manufacturing

Help optimize design of mechanical parts, respond quickly to customer questions, and create synthetic data to test applications

Why is RAG important?

LLM's are unpredictable, and its use of static training data makes a cutoff on the information it has

Zero-shot learning

LLMs can respond to a range of requests without explicit training, accuracy varies

Variational Autoencoders

Learn a representation of data called latent space, and use two neural networks called encoders and decoders.

Foundation Models

ML models trained on a broad spectrum of generalized/unlabeled data

Latent space

Mathematical representation of its data, with the unique code representing each of the attributes of the data

How do large language models work now?

Multi-dimensional vectors, where words of similar meanings are in similar vector spaces

Retrieval-Augmented Generation

Optimizing output of large language model, so it references a knowledge base outside of training data. Extends LLM to to specific domains or an organizations knowledge base without retraining

Encoder

Part of a Variational Autoencoder that maps input data to a mean/variance, then generates a random sample from a Gaussian distribution

Dangers of knowledge cutoff

Presenting false information when it does not have the answer, presenting out of date responses, or inaccurate information from terminology confusion

Few-shot learning

Providing a few relevant training examples, accuracy improves significantly

What mechanism do transformer-based models use

Self-attention mechanism, which weighs the importance of an input sequence when processing each element

What do transformers perform?

Self-learning, which allows it to understand basic grammar, languages, and knowledge

Difference between RAG and semantic search

Semantic search is understanding the meaning of queries, while RAG implements the data to enhance generated knowledge

Decoder

Takes the random sample and reconstructs it back to data that represents the original input

Generative AI

Type of AI that can create new content and ideas

How does generative AI work?

Use machine learning models to train on large amounts of data

What can Foundation Models do?

Use patterns and relationships to predict the next item of a sequence

How are large language models trained

Using a large amount of data, where the model changes its parameter values until it correctly predicts the next token from a previous set of tokens (trial and error)

Generative Adversarial Networks

Utilizes two neural networks, one called the generator that creates fake data by adding noise, and one called the discriminator that tries to detect what is real and what is fake data

What was a limitation for the original llm

You could not represent relationships between words with similar meaning