Unit 2: AI Foundations

Artificial Intelligence = ability of machines to mimic the capabilities of

humans.

Human intelligence:

- Learn new skills

- Abstract thinking
- Non verbal cues
- Handle complex situations
- Plan short/long term
- Creativity

Artificial General Intelligence (AGI) – repliate ANY of those capabilities.

AI – using AGI for specific problems

AI enhances SPEED and EFFECTIVENESS of human tasks.

Uses of AI:

- Automation and Decision Making

- Creative Support

AI Domains:

- Language, Vision, Speech, Product Recommendations, Anomaly

Detection, Learn by Reward, Forecasting, Generating Content

Language Tasks:

- Tokenization – converting words to numbers

- Padding – reducing varying sentence length
- Embedding – Dot and Cosine similarity classification to find patterns
- NLP – natural language processing.
- Recurrent Neural Networks – Process data SEQUENTIAL, hidden state
- Long Short-Term Memory – SEQUENTIAL, Uses gates
- Transformers – PARALLEL, uses SELF ATTENTION to understand context

Speech Tasks:

- Sample Rate – 44.1 kHz, rate of sampling audio

- Bit depth - # bits in each of the samples
- Goal: Find correlations of multiple samples.
 - Previous 3 + Variational Autoencoders, Waveform models, Siamese
Networks

Vision Tasks:

- Convolutional Neural Network – detect patterns in images, learning

hierarchical representations of visual features
- YOLO – pocess and detect objects within the image
- Generative Adversarial Network (GAN) – Generate REAL looking
images.

Artificial Intelligence:

- Machines imitate human intelligence

- Machine Learning:
o Algorithms that learn from past data and predict outcomes or
identify trends.
o Deep Learning:
 Learn from complex data using neural networks to predict
outcomes or generate NEW data.

Supervised Machine Learning:

- Learns from labeled data

- Extract rules.

Unsupervised Machine Learning:

- Extracting trends from unlabeled data

- Clustering, dimensionality reduction, etc.

Reinforcement Learning:

- Agent learns to perform actions in an environment.

- Reward or punishment
- Solve by trial/error

Deep Learning:

- Training neural networks with multiple layers to learn features and

rules.
- Can automatically learn by themselves.
- Example of Supervized ML algorithm.
Function Approximation – Technique of estimating an underlying mystery
function using historical observations.

Unit 3: Machine Learning Foundations

Machine Learning – SUBSET of AI that learns and improves from experience.

- Analyze, visualize, and make predictions from data

- Input Features  Output Labels

Supervised ML – Classify data or make predictions

Unsupervised ML – No labels, understand RELATIONSHIPS

Reinforcement ML – make decisions or choices

Continuous Output – Regression. Categorical Output – Classification.

Classification: Binary – one or the other. Otherwise multiclass.

Logistic Regression – Predicts something as true or false. (think logistic

function)

- Y axis of logistic regression is probability of true. (0-1)

Independent vs Dependent Freatures Use scatter plot and linear

regression line

Y-intercept of line can also be called “bias”.

Loss = error. The goal is to minimize the Squared Error.

Anaconda – use for data science, supports python and R and Jupyter
Notebooks

Unsupervised ML Use cases: Market Segmentation, Outlier Analysis

Similarity – how close 2 data points are to each other (value 0-1)

Unsupervised Workflow –

- Prepare (normalize + scaling) , Create

similarity metrics , Run clustering
algorithm , interpret and adjust clustering.
Reinforcement Learning Examples: Autonomous vehicles, Smart Devices,
Industrial Automation, Gaming/Entertainment

Unit 4: Deep Learning Foundations

DL = training Artificial Neural Networks (ANNs) with multiple layers.

- Learn and exract intricate representations from data.

- EXTRACT FEATURES from raw and complex data. (no specify features)
- Internal representation of data built using extracted features
- Parallel processing of data
- Better scalability and performance.

Applications of DL:

- Image classification
- NLP
- Language generation, summary, generative AI, etc.

Select the right DL Algorithm:

- Images/Videos  Convolutional Neural Network (CNN)

- Sequential, Time Series, Natural Language  Transformers, Long-Short-
Term Memory (LSTM), or Recurrent Neural
Network (RNN)
- Images, Text, Audio Generation  Transformer,
Diffusion Models, GANs

ANNs are inspired by the human brain. Use neurons

( single computation unit (input  output) )

Input, Hidden, and Output layers.

Weights – determine the strength of connection between neurons

Activation Function – Works on the weighted sum of inputs to a neuron

and produces an output.

Bias – Additional input to neurons for flexibility.

Backpropagation Algorithm – Training ANNs

- Guess and Compare

- Measure the Error
- Adjust the Guess
 - Update the weights (then repeat)

Sequence Models – input data are sequences. Goal is to find patterns and
make predicitons.

- NLP, Speech Recognition, Music Generation, Gesture Recognition, Time

Series Analysis

Recurrent Neural Network (RNN) – Handle Sequential Data

- Allow info to persist using a feedback loop

- Maintains a hidden state or memory
- Capture dependencies
- Types of RNN Architecture
o 1 to 1 – Standard non-sequential data like FNN
o 1 to many – music generation or sequence generation
o Many to 1 – sentiment analysis
o Many to Many – machine translation / named entity recognition

LSTM – uses specialized memory cell and gating mechanics to capture long
term dependencies of data

- Selectively remembers/forgets information over time.

Steps of LSTM: input processing, Previous memory, Gating Mechanism (input

gate  forget gate  output gate), Update memory (cell state), Output
Generation (for current time step)

Feed Forward Neural Networks (FNN)

- Also called Multi Layer Perceptron (MLP) (the simplest form)

Convolutional Neural Network (CNN) – learn patterns from image or video

RNN – Handle sequential data and use feedback loop

Autoencoders – Unsupervised models for feature extraction, dimensionality

reduction, employed in data compression, anomaly identification.

LSTM – (type of RNN) Find long term dependencies of data

GAN – Generates realistic synthetic data

Transformers – used in NLP

CNN – Used for grid like data.

 - Input layers – accept 3D images with height, width, and depth.
- Feature Extraction Layers – repeating pattern of convolution layer,
ReLu activation function, and pooling layer
- Classification Layer (output)

Convolution Layer – applies convolutions to images using small filters

(kernels)

Activation Layer – learn complex and NON-linear relationships

Pooling Layer – Reduce computational complexity and dimensions of the

feature maps

Limitation of CNN:

- Computationally expensive
- Overfits with limited training data
- Hard to interpret (black box)
- Sensitive to input variations

Applications of CNN – image classification, object detection, image

segmentation, face recognition, medical imaging, autonomous vehicles
(understand what they see), remote sensing.

Unit 5: Generative AI and LLM Foundations

Generative AI: Creates NEW content, part of deep learning

- Learns the underlying patterns to create NEW data matching these

patterns.
- Does not require labeled data in pre-train stage
- Text-Based Gen-AI vs Multimodal (images, audio, video, text) Gen-AI.

Language Model (LM) – probabilistic model of text.

- Uses probabities to decide what the next word is.

- Large in LLM just means # of parameters. (A LOT)
- EOS = end of sentence/sequece.

LLM can answer questions, write stuff, translate stuff

- Based on DL architecture (Transformer)

- Enhanced contextual understanding. NLP
 - Trained on vast language data to recognize patterns
- 100s of millions to billions of parameters.

Parameters – adjustable weights in the neural network. (too many

parameters = overfit)

Model Size – memory to store the parameters.

RNNs maintain a hidden state to allow persistance of information.

RNNs have feedback loops and capture dependencies.

Vanishing Gradient – long range dependencies are harder to capture.

Transformers – look at ALL words in the sentence and understand how all
words relate to each other.

Attention Mechanism – adds context to the text.

- Helps the transformer capture long range dependencies.

Encoder processes input and makes vectors using attention mechanism.

Decoder generates output.

Tokens – part of word, a word, or punctuation.

- # tokens = complexity

Embeddings – NUMERICAL representation of a piece of text converted to

number sequences.

- Piece of text can be from part of a word to a lot of text.

Vector Database – used to do similarity searches. Helps LLMs provide

informed answers.

Retrieval-Augmented Generation (RAG) – the architecture of LLM and vector

database.

Decoder = Models take a sequence and output next word.

- Decoder generates token and sends it back to itself until the whole
output is generated.

Encoder-Decoder Architecture: Encoder + decoder (from above)

Prompt – input or initial text provided to model

Prompt Engineering – refining a prompt to get a particular style of

response

Completion LLMs follow the dataset, which may not always be what the user
wants.

- Instruction tuning is a CRITICAL STEP in LLM alignment.

- Reinforcement Learning from Human Feedback (RLHF) = used to fine
tune LLMs to follow human instructions.

In-context Learning – conditioning an LLM with instructions or demos of the

ideal task.

k-shot prompting – provide k examples of the intended task in the prompt.

- 0-shot prompting = no examples.

Chain-of-Thought Prompting – using reasoning steps and calculation logic

before the final answer.

Hallucination – model generated text that is made up.

- RAG claims to reduce hallucination but there is no known methodology

to reliably reduce hallucination.

Customize LLMs with your data: Prompt Engineering  RAG  Fine Tuning

RAG – language model queries enterprise knowledge bases (DBs, wikis,

vector DBs, etc)

- RAG doesn’t require fine tuning

Fine-Tuning – take pretrained foundational model and provide add’l training

using CUSTOM DATA.
Inference – model receives NEW TEXT as input and generates output based
onw hat it learns from pretraining and fine tuning.

Fine Tuning:

- Optimize on a domain specifc dataset.

- Useful for when model doesn’t perform
task well or when teaching new things
- Adapt to specific style, tone, and learn
domain-specific words/phrases.

Fine Tuning Benefits:

- Improve model performance on specific

tasks
- Improve model efficiency

Unit 6: OCI AI Portfolio

Left consumes Right for [SaaS Apps, AI Services, Infrastructure, Data]

OCI console – browser-based interface, acces to notebook and service

features

Rest API – access to service functionality, requires programming

Language SDKs – provides programming language SDKs

CLI – quick access and full functionality, no scripting

Pretrained Models – Language Detection, Sentiment Analysis, Key Phrase

Extraction

Custom Models – Named entity recognition, Text Classification

Speech – convert media files into text (JSON and SRT format)

Digital Assistant – ai driven interfaces that help users achieve tasks with NLP

Services (7x): Generative AI, AI Agent Platform, Digital Assistant,

Language, Speech, Vision, Document Understanding

OCI Data Science:

- Accelerated: automated workflow, open-source libraries, streamlined

approach to building models, Collaborative, Enterprise-Grade
- Build, train, deploy ML models. Use Jupyter
Notebook
- Notebook session – contains jupyter notebook, libraries, etc
- Conda environment – code environment
- Accelerated Data Science (ADS) SDK
- Model Deployments – deploy models as HTTP or API infrastructure
- Jobs – Define and run repeatable ML tasks and workflow.

GPU – Graphics Processing Unit:

- Parallel computing for large datasets. Optimize for DL

Remote Direct Memory Access (RDMA) – data transfer, bypass CPU (low
latency)

Supercluster – many GPUs with RDMA.

OCI uses Non-blocking Network Fabric OCI RDMA Supercluster

Clos Fabric - multistage circuit-switching network

OCI RDMA Supercluster is lossless and low latency. (because networks are
very local)

Control Plane – deploys workloads only as much as needed to customers.

Optimize distribution for efficiency.

Flow Collision – two flows collide on a single link.

AI guiding principles: Legal, ethical, robust (technical and social)

AI must be regulated by policy, national, and international law

Human ethics: respect for human dignity, freedom for individuals, respect for
democracy, justice, and law, and have equality.

AI Ethics: respect human autonomy, prevent harm, fairness, explicability

Responsible AI Requirements:

- Human centric and human oversight

- Technical robust and safety
- Privacy and data protection
- Transparency, diversity, nondiscrimination, fairness +
Accountability

Steps: Set up governance  Develop policies and procedures  Ensure

compliance

AI is only as good as the data it is trained on.

Unit 7: OCI Generative AI Service

Gen-AI Service:

- Fully managed service, provides customizable LLMs (single API)

- Choice of pretrained models
- Flexible fine-tuning
- Dedicated AI clusters
o GPU and RDMA to host resources (separate from other GPUs)

Foundational Models:

- Chat – questions, conversational response

o Instruction-following models
- Embedding – text  vector embeddings
o Semantic Search Multilingual Model

Fine-tuning: optimization on a SMALLER, domain-SPECIFC dataset

T-Few Fine Tuning – fast and efficient customizations

Preamble Override – defines the AI’s self-identity (context)

Temperature = how random it is

Endpoint – used to host and serve fine-tuned models.

Playground – visually explore and test pre-trained and fine-tuned models

(no code).

Oracle Database 23ai

- SQL support for vector generation

- Vector Data Type
- Similarilty search
- Approx. search indices

Used with Gen-AI Pipelines

Distance Functions – used to determine similarity

Vector Search finds top K closest matches to a query item

Organization: In memory or Neighbor Partitions

Target Accuracy – specify how much accuracy needed for result

Similarity Search Over Joins:

- Go through multiple tables

Select AI:

- Use human langauge to query data

- Just ask a question.
- The AI will find the data you need.
- Can view data in different ways
- Can access the AI generated SQL used to retrieve data.

Select AI is: Simple, Futre-enabled, Secure

You can choose which AI model to use. Then specify schemas, tables, or
views for processing.

Unit 8: OCI AI Services

Language:

- Detect language
 - Identify entities (like date, time, currency)
- Identiy sentiment of parts of the text
- Identify key phrases
- Classify general topic from list of 600 categories and subcategories

Speech:

- Treanscription (DL) (w/o data science)

o With time stamps
- Processes data in oject storage
- Multiple languages (Spanish, Portuguese, English)
- Batching support (many files at once)
- Fast processing. (10 hr in <10 min)
- Confidence scores
- Punctuates transcription
- SRT close captions
- Normalization – make transcribed text more readable (one hundred
 100)
- Profanity filtering (hide, mask, tag)

Vision:

- Image Analysis:
o Object Detection (detect objects inside image)
o Image Classification (label scene)
- Document AI:
o Works with document images
o Text recognition
o Document Classification (10 different possible types)
o Language Detection (analyze visual features of text)
o Table Extraction
o Key Value Extraction