LLM - Model

LLM – Large Language Model #

Large Language Models (LLMs) are advanced AI systems designed to process, understand, and generate human-like text.

They learn language by analysing massive amounts of text data, discovering patterns in:

• grammar

• meaning

• context

• relationships between words and sentences

• Built on Deep Learning

• Implemented using Neural Networks

• Based on Transformers

• Often combined with tools like:

  • Retrieval (RAG)
  • Agents
  • External APIs
  • Memory systems

What makes an LLM special? #

• Built using deep neural networks
• Trained on very large datasets (books, articles, code, web text)
• Can perform many tasks without task-specific training
• General-purpose language understanding, not single-task models

Foundation: Transformer Architecture #

LLMs are based on the Transformer Architecture, which allows models to understand context and long-range dependencies in text.

AI Agents

December 15, 2025

AI Agents #

Also referred to as Agentic AI.

AI agents are intelligent systems that can plan, make decisions, and take actions to achieve goals with minimal human intervention.

• A common use case is task automation

• for example booking travel based on a user’s request.

• AI agents typically build on Generative AI and use Large Language Models (LLMs) as the reasoning core.

• Agents often interact with tools (APIs, databases, calendars) to complete multi-step workflows.

Retrieval-Augmented Generation (RAG)

Retrieval-Augmented Generation (RAG) #

Retrieval-Augmented Generation (RAG) is a system design pattern that improves an LLM’s answers by:

1. Retrieving relevant information from an external knowledge source, and then
2. Augmenting the LLM prompt with that retrieved context before generating the final response.

RAG helps an LLM look things up first, then answer using evidence.

Why RAG is Useful #

RAG is commonly used when:

• Your knowledge is in private documents (PDFs, policies, internal wiki)
• You need up-to-date information (things not in the model’s training data)
• You want fewer hallucinations by grounding answers in retrieved sources
• You want traceability (show “where the answer came from”)

RAG does not change the model weights.
It changes what the model sees at inference time by adding retrieved context.

Deep Feedforward Neural Networks (DFNN) for Classification

February 26, 2026

Deep Feedforward Neural Networks (DFNN) or Multi Layer Perceptrons (MLP) for Classification #

A Deep Feedforward Neural Network (DFNN), also called a Multi-Layer Perceptron (MLP), is a neural network with one or more hidden layers where information flows forward only (no recurrence).
For classification, DFNNs learn non-linear decision boundaries by combining hidden layers with non-linear activation functions.

Core idea:

• A single neuron can only learn linear boundaries.
• Adding hidden layers + non-linearity allows DFNNs to solve problems like XOR.

MLP as solution for XOR #

A single perceptron fails on XOR because XOR is not linearly separable.

Communication-Aware Distributed ML

Communication-Aware Distributed ML #

Distributed machine learning is effective only when saved computation exceeds the cost of moving data and synchronising workers. Algorithm design must therefore account for message size, message frequency, barriers, and parameter placement.

Course coverage:

1. Communication overhead, illustrated through distributed k-means
2. Model parallelism when parameters do not fit or compute must be divided
3. Distributed k-nearest neighbours: partitioning, indexing, and approximate search
4. Gradient descent, SGD, and mini-batch optimisation
5. Synchronous versus asynchronous updates

Learning Objectives #

By the end of this page, you should be able to:

Decision Tree

Decision Tree #

A decision tree classifies an example by asking a sequence of questions about its attributes until it reaches a leaf (final decision).

Key takeaway: A decision tree grows by repeatedly splitting the training data into purer subsets using an impurity measure (Entropy / Gini / Classification Error).

• Information Theory
• Entropy Based Decision Tree Construction
• Avoiding Overfitting
• Minimum Description Length
• Handling Continuous valued attributes, missing attributes

Information Theory #

Decision trees need a way to measure: “How mixed are the class labels at a node?”

LLM and Prompt Engineering

LLMs and Prompt Engineering #

A Large Language Model extends neural language modelling through much larger datasets, many more parameters, broad pretraining and adaptation to many downstream tasks. Its central operation remains next-token prediction.

Learning Objectives #

• Explain how neural language modelling develops into an LLM.
• Describe the meaning of large, general-purpose and pretrained.
• Explain how a prompt guides generation.
• Distinguish zero-shot and few-shot prompting.
• Compare prompting with model adaptation.

Big Picture #

flowchart TD
    A["Broad Text Data"] --> B["Large-scale Pretraining"]
    B --> C["General Language Model"]
    C --> D["Prompt or Adaptation"]
    D --> E["Task Output"]

    style A fill:#E1F5FE
    style B fill:#C8E6C9
    style C fill:#FFF9C4
    style D fill:#EDE7F6
    style E fill:#E1F5FE

1. From Neural Language Models to LLMs ☆ #

A neural language model learns a conditional probability for the next token:

Local Search and Optimisation

Local Search and Optimisation #

Local search improves one or a few complete candidate solutions by moving through their neighbourhoods. It is especially useful when the final configuration matters but the path used to reach it does not.

– Local Search Algorithms & Optimisation Problems

• Hill Climbing Search
• Local Beam Search
• Online Search Agents

Learning Objectives #

• distinguish path optimisation from state optimisation
• explain states, neighbours and objective functions
• apply hill climbing and its variants
• recognise local maxima, plateaus and ridges
• explain random-restart and stochastic hill climbing
• describe local and stochastic beam search and trace one beam update

Big Picture #

flowchart TD
    S["Current state"] --> N["Generate neighbours"]
    N --> E["Evaluate values"]
    E --> B["Choose candidate"]
    B --> S
    B --> G["Stop condition"]

    style S fill:#E1F5FE
    style N fill:#FFF9C4
    style E fill:#EDE7F6
    style B fill:#C8E6C9
    style G fill:#C8E6C9

1. Path Optimisation vs State Optimisation ☆ #

In classical route search, the path is the solution. The agent must remember how it travelled from the start to the goal.

Monte Carlo Methods

Monte Carlo Methods #

Monte Carlo (MC) methods learn value functions and policies by averaging returns observed in complete episodes. Unlike Dynamic Programming, they do not require transition probabilities or a complete model of the environment.

1. The Main Idea ☆ #

Suppose the agent repeatedly reaches a state \( s \) . Each time, it records the return received from that point until the episode ends. The average of these returns estimates the state’s value.

Prediction & Forecasting

Prediction & Forecasting #

Prediction and forecasting use statistical models to estimate unknown or future values.

In this module, the focus is on correlation, regression, and time series forecasting.

Key takeaway:
Prediction estimates a value using a model.

Forecasting is prediction where the order of time matters.

• Correlation
• Regression
• Time series analysis
• Components of time series data
• Moving average and weighted moving average
• AR model
• ARMA model
• ARIMA model
• SARIMA and SARIMAX
• VAR and VARMAX
• Simple exponential smoothing

Prediction vs Forecasting ☆ #

Overall Workflow #

flowchart LR
    A[Data] --> B[Explore Pattern]
    B --> C[Choose Model]
    C --> D[Train or Fit]
    D --> E[Validate]
    E --> F[Predict or Forecast]
    F --> G[Interpret Error]

    style A fill:#E1F5FE
    style B fill:#C8E6C9
    style C fill:#FFF9C4
    style D fill:#EDE7F6
    style E fill:#C8E6C9
    style F fill:#E1F5FE
    style G fill:#FFF9C4

Correlation ☆ #

Correlation measures the direction and strength of linear relationship between two variables.