Flow: Input → Rules/Program → Output
Limitation: This cannot categorize, recognize, predicts.
--------------------------------------------------------------------------------
2. Statistics & Data Mining:
Data mining is the process of finding patterns, correlations, and useful insights using statistics, machine learning, and database systems
Instead of writing rules, people started analyzing historical data.
Still mostly human-driven.
--------------------------------------------------------------------------------
3. DS: Data Science
ML(Machine Learning)- is branch of AI that works with statistical part of AI. It helps computer to solve problem.
DL(Deep Learning)- this is enhancement of ML where computer learn and make decision.
AI(Artificial Intelligence)- make computer intelligent to perform function normally done by human. AI includes Computer Vision, Language Processing, Creative and summarization.
--------------------------------------------------------------------------------
4. Machine Learning (ML), Data is provided so that machine can learn pattern
Major Shift
Instead of Data + Rules → Answer
New approach was Data + Answers → Model (The model learns the rules itself.)
Different algorithm learns different characteristics automatically.
Common ML Algorithms
Linear Regression
Logistic Regression
Decision Trees
Random Forest
SVM
K-Means Clustering
# ML life cycle
Collect Data
↓
Training
↓
Testing/Validation
↓
Deploy Model
↓
Inference (Prediction)
↓
Retraining (if needed)
How Machine Learns?
Supervised Learning
system is trained using labeled datasets.
E.g. AI predict network security threat by training datasets labeled as “safe” and “malicious”
Unsupervised Learning:
don’t use labeled data.
Useful for anomaly detection.
It enables the system to spot unusual traffic patterns that deviate from what’s considered normal,
potentially flagging a security breach without prior knowledge of what to look for.
Reinforcement Learning:
Dynamic learning process where the AI system learns form consequences of its actions,
trial-and-error learning method. In the context of network operation, system can refine network performance by adjusting configurations and polices in response to real-time feedback,
to achieve optimal performance.
--------------------------------------------------------------------------------
5. Neural Networks
Researchers tried to mimic the human brain.
A neuron receives inputs and produces output.
Input Layer
↓
Hidden Layer
↓
Output Layer
Neural networks can learn more complex patterns than traditional ML.
Way machine learns pattern for given data. neural network mimic our brain (neurons - decision maker)
Training:
millions of example --- neural network --- correct answer
--------------------------------------------------------------------------------------------
6. Deep Learning,
Deep Learning is the technique of using large neural networks with many hidden layers to solve complex problems. Here neural networks became very large with many layers:
Neural Network + Many Layers = Deep Learning
Benefit:
Image recognition
Speech recognition
Language understanding
Examples:
Google Photos
Siri
Face Recognition
Input
↓
Hidden Layer
↓
Hidden Layer
↓
Hidden Layer
↓
Hidden Layer
↓
Output
Deep Learning types:
CNN (Convolutional Neural Network) -->Specialized for images
(Face recognition, Medical imaging, CCTV analytics)
RNN (Recurrent Neural Network) Designed for sequential data.
Examples: Speech processing, Time-series forecasting (Mostly replaced by Transformers.
GAN (Generative Adversarial Network) Two neural networks compete:
Generator vs Discriminator, best for Image creation, Synthetic data generation
Transformer : The architecture behind modern LLMs.
Learn by rewards and penalties.(Examples: Robotics, Self-driving cars, Game-playing AI)
----------------------------------------------------------------------------------
7. AI - Mimic human intelligence
7.1 AI Models (Based on What the Model Does)
Predictive Models -->Predict future outcomes. Examples: Device failure prediction, Capacity forecasting, Security threat prediction
Predictive AI = Predicts outcomes based on historical data.
Generative Models --> Generate / Create new content. (Generated text, image, code, audio)
Examples: ChatGPT, Microsoft Copilot, DALL-E, Stable Diffusion
GPT = Generative Pre-trained Transformer.
Classification Models --> Assign items to predefined categories.
(Examples: Spam vs Non-Spam, Malware vs Benign, High Risk vs Low Risk)
Regression Models --> Predict numeric values.
(Examples: Forecast bandwidth usage, Predict sales, Estimate network latency)
Recommendation Models -->Recommend items likely to interest users.
(Examples: Netflix, YouTube, Amazon)
Reinforcement Learning Models -->
(Best action selected after learning) Learn by rewards and penalties.
(Examples: Robotics, Self-driving cars, Game-playing AI)
7.2 AI MODEL Based on Technology
Machine Learning : Traditional ML
Deep Learning
Neural Networks -->Prediction, Recognition, Generation
CNN (Convolutional Neural Network) -->Specialized for images
(Face recognition, Medical imaging, CCTV analytics)
RNN RNN (Recurrent Neural Network) Designed for sequential data.
Examples: Speech processing, Time-series forecasting (Mostly replaced by Transformers.
GAN (Generative Adversarial Network) Two neural networks compete:
Generator vs Discriminator, best for Image creation, Synthetic data generation
Transformer The architecture behind modern LLMs.
Learn by rewards and penalties.(Examples: Robotics, Self-driving cars, Game-playing AI)
GAN
Generative AI Model Types
A GAN consists of two neural networks competing against each other:
Generator → Creates fake content
Discriminator → Tries to detect whether content is real or fake
Diffusion Models
Multimodal Models --> Process multiple data types.
Microsoft Copilot, GPT multimodal, Gemini
The competition helps the generator improve over time.
Base Model --> General-purpose model trained on large datasets. GPT, Llama, Gemini
Fine-Tuned Model --> Specialized for a particular domain. trained on Cisco, Arista, and Aruba documentation.
RAG-Enabled Model Combines: LLM + Company Knowledge Base Used by enterprise chatbots.
AI Agent LLM + Tools + Actions
Agentic AI System LLM + Planning + Reasoning + Tools + Autonomy
----------------------------------------------------------------------------------
8 Generative AI
Traditional ML predicts.
Generative AI creates. Write an email, Create an image, Generate code
Examples:
ChatGPT, Copilot, Gemini, Claude
Think of it as: Predict → Generate → Act ( Autonomously )
----------------------------------------------------------------------------------
AI Agent = Act, Uses "AI" + "tools" to perform tasks.
----------------------------------------------------------------------------------
Agentic AI = (Group of AI Agents working together) Think + Plan + Act ,Purpose: Plans, reasons, and executes multi-step goals with minimal human guidance.
Enterprise enhancements: RAG + Fine-Tuning
Challenges: Hallucination
--------------------------------------------------------------------------------
--------------------------------------------------------------------------------
AI Terms:
--------------------------------------------------------------------------------
LLM: Neural network trained on a massive amount of text data is called LLM (large language model)
E.g chatgpt, cloud or gemini
They are good at predicting what next is ?
computer start to understand language context reasoning and code.
it think like human
--------------------------------------------------------------------------------
Token
LLM don't read words they read token
They are chunk of text
Every model has limit on how many token it can process once. This is called "context windows"
Context window is like short term memory
--------------------------------------------------------------------------------
Embedding
AI convert text into number (called as vector ) this is called vector Embedding
A mathematical representation of data (usually text) converted into vectors so AI can understand similarity and meaning.
In vector data base the word with similar meaning get similar number or vector representation.
Example: Router, Switch, Firewall, These networking terms will have embeddings close to each other compared to words like "Banana."
--------------------------------------------------------------------------------
Vector Database:
A database designed to store and search embeddings efficiently.
A vector database indexes and stores vectro embeddings for fast retrieval and similarity search
unstructured data <model> vector embedding (storing data)
use case: sematic serach: search based on the meaning or context
vector database example: Pinecone, Weaviate, ChromaDB, Azure AI Search, Redis, qdrant, mivus, vespa
--------------------------------------------------------------------------------
Prompt
The instruction or question provided to an AI model.
it is input give to ai model to do something.
The qualit of output depend upon quality of input (prompt)
Example: Generate a Cisco IOS configuration for OSPF.
--------------------------------------------------------------------------------
Inference:
An inference is a conclusion or educated guess that you form by combining known facts, clues, and your own background knowledge.
The process of using a trained model to generate predictions or answers.
Example: Asking Copilot a question and receiving a response.
--------------------------------------------------------------------------------
Transformer:
The neural network architecture behind modern LLMs.
Example: GPT, Llama, Claude, and Gemini are Transformer-based models.
--------------------------------------------------------------------------------
Multimodal AI:
AI capable of understanding multiple types of data such as text, images, audio, and video.
Example: Uploading a network diagram and asking Copilot to explain it.
--------------------------------------------------------------------------------
Retrieval Augmented Generation (RAG)
A technique where an LLM first retrieves information from external sources and then generates its answer using that information.
Example: A company chatbot searching internal network documentation before answering a support question.
RAG- Retrieval-Augmented Generation (Providing data)
Augmented means made greater, larger, stronger, or added to in number, amount, or value
LLM know lot - they only know public internet data they don't know our personal data
step 1: retrieve relevant information from your own documents
step 2: inject that info into prompt as additional info or context
Retrieval-Augmented Generation (RAG) is a technique that helps alleviate the issue of hallucination and out-of-date answers in LLMs.
RAG combines the capabilities of a retrieval system and a generative model.
Rag benefit:
reduce hallucinating
keep knowledge up to date
cost effective
maintain data privacy
RAG has two pipeline:
Ingestion pipe line
information/data --- extract data --- split into chunk == embedding vector database (semantic search)
Retrieval pipeline
user - query - embedding api - embedding - semantic search- context - pass it to mode - augmentation - generation (answer)
When given a prompt, the retrieval system first fetches relevant documents or pieces of information from a large database or knowledge source. Then, the generative model uses this retrieved information to generate a response. This process ensures that the responses are grounded in the retrieved documents, which can help improve factual accuracy and reduce the likelihood of hallucination. The key advantage of RAG is that it anchors the model's responses to real-world information contained in the retrieved documents, making the responses more accurate and less prone to fabricating information.
This capability makes RAG systems particularly valuable for navigating complex technical documentation.
Instead of manually sifting through dense material, users can rely on the RAG system to extract relevant details and present them in a digestible format.
This not only speeds up the process of finding information but also ensures that the guidance provided is accurate and directly relevant to the task at hand,
as it's grounded in the actual content of the documentation.
You can leverage your own RAG system with most modern LLMs by simply drag-and-dropping a PDF document into the prompt and directing the LLM to consult the documentation when giving you the response. ChatGPT, Claude, and Gemini all support file uploads.
Here model did not learn your data during learning but learn real time
--------------------------------------------------------------------------------
Fine-Tuning
Additional training of a pre-trained model using specialized data to improve performance in a specific domain.
It actually change model.
take General-purpose llm and train it with personal data.
Example: Fine-tuning an LLM on Cisco TAC cases to make it a network-support assistant.
--------------------------------------------------------------------------------
RAG VS Fine-Tuning:
RAG: handing a consultant a briefing document before a meeting
Fine-Tuning: Hiring someone who spent three year working in your industry.
--------------------------------------------------------------------------------
Hallucination
When an AI confidently generates information that sounds correct but is actually incorrect or fabricated.
Example: An AI claiming Cisco IOS-XE 20.1 exists when no such version exists.
AI makes up stuff that are factually not true.
LLM are good at predicting next Plausible text or token but not predict fact.
(Plausible means seeming reasonable, believable, or likely to be true)
It is not bug.
--------------------------------------------------------------------------
Technology Purpose
Embeddings Convert text into vectors
Vector DB Store embeddings
RAG Retrieve relevant knowledge
MCP Connect AI to tools and systems
Agent Uses LLM + RAG + MCP + reasoning
