AI, ML, Deep Learning

May 23, 2023

Under Construction, Watch Your Step

AI Overview

Big surprise! MLOps is helped tremendously by knowing a fair bit about AI.

NLP
CNN vs ANN?
CV
Transfer Learning

AI

AI is

ML A subset of AI dedicated to taking data from the past and training algorithms to create models that can perform highly complex tasks without being explicitly programmed.
- Supervised (labeled): learn from past example to make future predictions
  - Reinforcement Learning ??? Each iteration, weights are changed to minimize error
  - Gradient Decent
- Unsupervised: raw data and look for correlations (grouping)
- Deep Learning A subset of ML that uses artificial neural networks to process more complex patterns than traditional ML. Uses ANNs.
  - ANN (Artificial Neural Networks)(aka NN)
    - Multiple hidden layers (Input Layer - Hidden Layers - Output Layer)
    - Can process labeled and unlabeled data.
    - “Semi-Supervised Learning”: small amount of labeled data, large amount of unlabeled.
      - Labeled helps learn basics of task
      - Unlabeled helps the NN generalize to new examples
  - Generative AI
    - Subset of Deep Learning.
    - Uses ANNs so can process labled and unlabeled data.
    - Uses Semi, supervised, unsupervised learning.
    - Typically involves the Transformer architecture. Essentially, it’s a type of AI that can map long-range dependencies and patterns in large training sets, then use what it learns to produce new content, including text, imagery, audio, and synthetic data.
    - Relies on large models, such as large language models (LLMs) that can classify and generate text, answer questions, and summarize documents
  - LLM (Large Language Models)
    - Subset of Deep Learning

ML/Deep Learning Model Types

Model types: Discriminative vs Generative
Discriminative (aka Predictive) used to classify (is this a dog or a cat or something else)
Generative (aka GenAI) used to generate (create a dog based on all the dog’s you were trained on)
- Part of flow is to check with Discriminative model to see if the generated object passes classification check.
- Uses unstructured content to learn patterns in content.
- NOTE: “model” can also be called a “function” with a multidimensional tensor/matrix with adjustable weights/values.
Classical Supervised and Unsupervised Learning
New Gen AI Supervised, Semi and Unsupervised Learning Generate a Foundation Model (aka a statistical model)
Gen Lang models can be asked questions (prompted)
- PaLM ?
- LaMDA ?
- GPT ?

Transformers

Transformers are “the 2018 revolution” that enabled Generative AI via NLP (Natural Language Processing).

Hallucinations are words or phrases that are generated by the model that are often nonsensical or grammatically incorrect.

Causes of Hallucinations:

Model not trained on enough data.
Model trained on noisy or dirty data.
Model is not given enough context
Model is not given enough constraints

Prompt Design: the quality of the input determines the quality of the output.

Model Types??

text-to-text
- NLP lang input to lang output.
- Learn mapping between pairs, e.g. lang translation
  - Generation
  - Classification
  - Summarization
  - Translation
  - (Re)Search
  - Extraction
  - Clustering
  - Content editing/rewriting
text-to-image
text-to-video and text-to-3D
- video gen
- video editing
- Game assets
text-to-task
- Software agents
- Virtual assistants
- Automation (e.g.: navigate a web GUI)
Foundation Model Large Model pre trained on a vast amount of data designed to be adapted or fine-tuned.

Vertex AI Task Specific Foundation Models

Examples:

Sentiment Analysis: use Language > Extraction > Syntax Analysis > Classification > Sentiment Analysis
Occupancy Analytics: use Vision > Classification > Object Detector > Detection > Occupancy Analytics

GenAI Studio

On Google Cloud. See also SageMaker?

Gen AI App Builder

No code. Visual WYSIWIG.
Create your own
- digital asst.
- knowledge base

PaLM API and MakerSuite

Simplifies Gen Dev Cycle

Intro to LLMs

Video by Google, May,2023
AI > ML > Deep Learning > LLM
Generative AI
LLMs: large, gen purpose lang models that are pre-trained and then fine-tuned
- Large
  - Petabyte scale dataset
  - Large num parameters (called hyperparameters in ML)
- General Purpose
  - commonality of human lang
  - resource restriction (not many orgs can train an LLM)
- Pre-trained and fine-tuned
Benefits of LLMs
- Single LLM can be used for different tasks
- Fine-tune requires minimal field data (aka domain training data)
  - Decent performance with “few-shot” (little data) or “zero-shot” (never been trained but works anyway) scenarios
Performance is continuosly growing with more data and parameters.

Example: PaLM (Pathways Language Model)

April 2022, Google released it.
540 Billion parameters
Dense decoder only model
Leverage the new Pathway system (distributed training on multiple TPU (Tensor Processing Units) V4 pods)

LLM Transformers

Encoding Component > Decoding Component
NN (Neural Networks) circa 2012.
Generative: user can train
- Ex:
  - LaMDA (Lang Model for Dialogue Apps)
  - PaLM
  - GPT

LLM Development vs. Traditional Devlopment

QA (Question Answering)

Traditional QA

Subfield of NLP that deals with answering questions posed in nat lang
Retrieves answers from a given text
- Depending on model, answer can be directly extracted from text or generated from scratch
Requires Domain Knowlege to train

VS Generative QA

Generates based on context
Leverages Text Gen models
No need for domain knowledge

Prompts Design

Prompt Design:

Process of creating prompts that elicit the desired response from a language model.
Instructions and context passed to lang model for specific task.

Prompt Engineering:

Practice of developing and optimizing prompts to efficiently use lang models, for a wide variety.

3 Types of LLM

First 2 are easily confused but very diff.

Generic (or Raw): predict next word (aka token)
- Token is a part of a word, the atomic unit that LLMs work in.
Instruction Tuned: predict a response to the instruction given in the input
Dialog Tuned: have a dialog by predicting the next response.
- subset of Instruction Tuned

Chain of Thought Reasoning

Models are better at getting the right answer when they first output text that explains the reason for the answer.

PETM (Parameter-Efficient Tuning Methods)

Fine-tuning an LLM on your own custom data without duplicating the model. The base model itself is not altered. Instead, a small number of add-on layers are tuned, which can be sswapped in and out at inference time.

Prompt Tuning: easiest PETM

VertexAI

Google Clouds Fully Manage ML platform. VS SageMaker, KubeFlow, Azure ML

Generative AI on Google Cloud

AI Roles

Data Scientist
Data Engineer
DataOps
ML Engineer
MLOps
Prompt Designer
Prompt Engineer

?

AGI
Training
- Gradient Descent
- Sigmoid Functions
- Attention

Types of Models

Open Source

See Hugging Face

BLOOM by BigScience
LLaMA by Meta AI
Flan-T5 by Google
GPT-J by Eleuther AI

Proprietary

OpenAI
co:here
AI21 Labs
Antrhopic

To File

Google Imagen
Dall-E 2

Misc

Libs and Langs

Libraries

Pytorch
TensorFlow
LLM
- LangChain
  - Getting Started with LangChain: A Beginner’s Guide to Building LLM-Powered Applications

LLMs

ChatGPT (OpenAI)
- Plugins 2023/05/17
Google Bard
MS Bing AI

Langs

Python
R Lang

Autonomous Agents

GPT/LLM backed

AutoGPT
- GTP-4 undelying
BabyAGI
AgentGPT
Interactive Simulacra (Stanford)
- NPCs controlled by GPT

AutoML Frameworks

Semantic Kernel

External Tools Access

Thanks to A comprehensive and hands-on guide to autonomous agents with GPT

Autonomous agents can now try to achieve a long-term goal by thinking through the sub-tasks, planning which actions to take, executing the actions with the help of external tools, and reflecting on the results.

Toolformer
JARVIS (HuggingGPT)
VisualChatGPT (TaskMatrix)
ReAct (Reasoning-Acting)
Reflexion

Vector DBs

Used by above

Pinecone
Weaviate
Milvus
Faiss
Chroma

Websites

OpenAI Hugging Face

ML Ops

Cloud Offerings

Certs

Google Generative AI Training

Articles

Building Your Own DevSecOps Knowledge Base with OpenAI, LangChain, and LlamaIndex