Artificial Neural Networks (ANN)

Artificial Neural Network (ANN)

• ANN is a computer system inspired by the human brain
• Just like the brain uses neurons to process information, an ANN uses artificial neurons to recognize patterns and make decisions

Input Layer

• This layer receives raw data such as images, numbers, or text
• Each neuron represents one feature of the data

Hidden Layers

• These layers perform the real learning by identifying patterns
• Each neuron is connected to other neurons and applies mathematical operations to the data
• The more hidden layers present, the deeper the network or deep learning

Output Layer

• The output layer produces the final decision or prediction
• For classifying emails, it gives spam (1) or not spam (0), and for predicting house prices, it provides a number like $350,000

Tabular - MLP (Multi-Layer Perceptron)

• MLP is a type of Artificial Neural Network (ANN) used for tabular data
• An MLP learns patterns in data to make predictions
• It has three main parts: an input layer, hidden layers, and an output layer

Recurrent Neural Network (RNN)

• RNN is a type of Artificial Neural Network (ANN) designed for sequential data, where past information matters for future predictions
• It uniquely has memory, remembering past inputs and uses them to influence current predictions
• A normal neural network would look at each subtitle separately
• An RNN remembers the past subtitles and understands the context

Long Short-Term Memory (LSTM)

• LSTM (Long Short-Term Memory) is a special type of Recurrent Neural Network (RNN) that remembers information for long periods and overcomes the problem of forgetting past data
• LSTM introduces three gates that help manage memory.

Forget Gate

• Decides what to forget from past memory
• When analyzing a book, it may forget old chapters that are no longer relevant

Input Gate

• Decides what new information to store in memory.
• When reading a sentence, it stores new important words

Output Gate

• Decides what information to send as the final output.
• When generating a sentence, it picks the next word based on stored knowledge

Transformer Models

• A Transformer model has two main parts, an encoder which reads and understands the input and a Decoder that Generates the output based on what the encoder learned
• The real magic of Transformers comes from a technique called Self-Attention
• Instead of reading words one-by-one like LSTMs, Transformers look at all words at once and figure out which words matter most in a sentence. Semantic Similarity
• RNN might struggle to know what "it" refers to, but the Transformer learns that "it" refers to "the mat", being much smarter at understanding context

Encoder

• It reads and understands the input

BERT

• BERT (Bidirectional Encoder Representations from Transformers) is a powerful transformer-based model developed by Google in 2018
• Revolutionized Natural Language Processing (NLP) with word-based embeddings, so one vector for each word
• It can understand the meaning of words in a sentence by looking at the context from both directions—left-to-right and right-to-left

Bidirectional Context

• BERT doesn't just read text left-to-right like older models
• It looks at both the words before and after a given word to understand the full context of a sentence
• Example: In the sentence "The bank was full of fish," BERT understands "bank" refers to the side of a river, not a financial institution

Pre-Training

• BERT is pre-trained on a massive amount of text (like Wikipedia), learning useful patterns in language

Masked Language Model (MLM)

• BERT randomly hides words in a sentence and tries to predict them
• Example: The quick brown [MASK] jumps over the lazy dog," predicting that [MASK] is "fox"

Next Sentence Prediction (NSP)

• BERT is also trained to predict whether one sentence follows another, helping it understand relationships between sentences

SBERT sentence-BERT

• SBERT is an improved version of BERT designed specifically for sentence embeddings, transforming sentences into numerical representations that capture their meaning
• Unlike BERT, which is slow for comparing sentences, SBERT is optimized for tasks like similarity detection, clustering, and retrieval
• Employs BERT as a base model to read and process text

Application of Siamese or Triplet Network Structure

• Allows for efficient comparison of sentence meanings

Output Sentence Embeddings

• Converts sentences into fixed-size numerical vectors

Semantic Similarity Computation

• Enables quick computation of cosine similarity between sentences, useful for search engines and chatbots
• If you want to make a search engine for a company, use SBERT and if you have a classification problem, then BERT is a better choice than SBERT

Decoder (GPTs)

• Decoder generates the output based on what the encoder learned

Chaining (Langchain)

• LangChain facilitates the building of applications using language models (LMs) like OpenAI's GPT or LLMs by integrating them into workflows
• LangChain helps build applications that interact with and process natural language data efficiently

Chains

• Chains comprise sequences of operations (LLM calls, API requests, and data processing) to perform tasks step by step

Prompt Engineering

• Optimizes prompts for better responses, dynamically modifying them based on input

External Integrations

• Connects LLMs to APIs, databases, and web scraping for real-time data access

Memory and Context

• Maintains conversation history for more coherent and context-aware interactions

Agents

• Dynamic workflows adapt decisions based on context, allowing flexible task execution

Agentic Systems

• Agentic systems are Al-powered systems or frameworks designed to create and manage autonomous agents
• These intelligent software entities can perform tasks, make decisions, and interact without human intervention, working independently towards predefined goals by using Al models like LLMs, decision-making models, or reinforcement learning

Flowise

• Flowise is a platform designed to help create and automate workflows using Al
• Emphasizes making the process of creating intelligent workflows easier by integrating with various AI models seamlessly

Visual Workflow Creation

• Users design workflows by dragging and dropping Al models, tools, and services onto a canvas

Integration

• Flowise integrates Al models, APIs, and external services to perform tasks like data processing, automation, or decision-making

Automation

• It automates responding to customer queries, processing data, or making decisions

Real-Time Execution

• Flowise executes workflows in real-time, pulling data from external sources, running Al models, and triggering actions

CrewAI

• CrewAl is a platform to build and manage AI-driven teams, enabling seamless integration of multiple Al models to autonomously handle various tasks in an orchestrated way to complete complex tasks

AI Agents

• Users create and configure Al agents, which can be individual models or systems, to handle specific tasks

Task Assignment

• CrewAl allows assigning specific tasks to different agents, who then work collaboratively to complete more complex workflows
• CrewAl autonomously manages the workflow to complete tasks, make decisions, and trigger actions

PydanticAI

• PydanticAl combines Pydantic with Al models to improve the interaction between Al models and structured data.

Data Validation

• PydanticAl uses Pydantic's data validation features to reducing errors and improving model performance

AI Integration

• Enables Al models to process validated data, streamlining the management and integration of models within structured data workflows

Model Output Handling

• Validates the output from Al models before processing, ensuring consistency and correctness

Smolagents

• Smolagents is a framework designed for creating small, lightweight Al agents that perform specific tasks autonomously

Small AI Agents

• Smolagents are small, focused Al models or agents that handle particular tasks like processing text, making decisions, or interacting with external systems

Task Assignment

• Each agent is assigned a specific task, enabling them to work collaboratively on complex workflows

Orchestration

• Allows Smolagents to work together in a larger system for efficient and distributed problem-solving without large models

Image - CNN (Convolutional Neural Networks)

• CNNs are deep learning models primarily used for analyzing and processing visual data
• They adaptively learn spatial hierarchies of features from input images, effective for image classification and object detection

Convolutional Layers

• The convolutional layer is where small filters slide over the input image to detect shapes
• These patterns learn to detect specific features in the image

Pooling Layers

• The data is often passed through a pooling layer, which reduces the image's spatial dimensions
• Helps in reducing the computational load

Fully Connected Layers

• The output is flattened and passed through fully connected layers, like traditional neural network layers
• These layers make the final decision

Activation Functions

• CNNs use activation functions like ReLU to introduce non-linearity for complex patterns

Graph - GNN (Graph Neural Networks)

• A GNN is a neural network specifically designed to work with graph-structured data composed of nodes connected by edges

Input Data

• The input to a GNN is a graph, where each node has features and edges represent relationships

Message Passing

• Employs a message-passing process where nodes exchange information with their neighbors

Iteration

• This message-passing process is repeated for several layers to capture information from further neighbors

Node/Graph Representation

• Multiple iterations allow each node to have an updated representation that integrates its features
• This representation is used for node classification, link prediction, or graph classification

Neuron Structure in ANN

• Input layer receives raw data like images, numbers, or text
• Hidden layers perform real learning by identifying patterns, applying mathematical operations
• Output layer produces final decision/prediction like spam (1) or not spam (0), or house price like $350,000

Purpose of Activation Function

• An activation function introduces non-linearity, enabling the model to learn complex patterns, and without it a neural network would behave like a simple linear regression model

Steps in Training a Neural Network

• Initialize weights and biases
• Pass input data to compute simple predictions
• Compute loss function to measure error like (MSE) or Cross-entropy classification
• Calculate gradients of the loss with respect to weights using chain rule
• Update weights from the gradient using optimization algorithm like SGD or Adam
• Use either Batch Gradient Descent/Stochastic Gradient Descent through the sample or hybrid

Assess the Model's Performance

• MSE (Mean Squared Error) predicts regression by measuring average of squared differences between predicted and actual values
• Cross-Entropy predicts classification tasks by measuring difference between the true class distribution

Batch Gradient Descent vs. Stochastic Gradient Descent

• Batch Gradient Descent (BGD) updates weights after processing the entire dataset while Stochastic Gradient Descent (SGD) updates weights after each

Optimal Values for Parameters such as Weights and Biases via Stochastic Gradient Descent (SGD) and Adam

• Optimization algorithms minimize the loss function by iteratively updating parameters through backpropagation, and combining momentum and adaptive learning rates

Loss Function

• It quantifies model error that guides weight adjustments during training
• For regression tasks, Mean Squared Error (MSE) is used, and for classification, Cross-Entropy Loss is preferred

Hyperparameters in Neural Networks

• Controls the training process, includes learning rate, batch size, number of layers, number of neurons
• Proper tuning of hyperparameters impacts model performance because poor results can effect the ability to generalize

Parameters in Neural Networks

• Parameters in a neural network are values learned during training, including weights and biases
• These are updated through optimization algorithms to minimize the loss function

Neural Network Format Selection

• For a tabular project use MLPs (Multi-Layer Perceptrons), for image processing use CNNs, for sequential data like time series or text use RNNs and for relationships in graph-based data use GNNs

Setting the Loss Function

• Choose the loss function to align itself with the problem's objective to ensure effective optimization and learning
• MSE is common for regression and Cross-Entropy is used for binary classification

Gradient Descent

• Gradient Descent is an optimization algorithm to minimize a function by adjusting parameters
• Gradients are the derivatives of the loss function, representing slope
• Updating parameters reduces error and finds the optimal value

Mean Squared Error (MSE)

• MSE measures the average squared difference between predicted and actual values in regression tasks
• Sensitively penalizes large errors more heavily due to squaring

Neural Network Backpropagation

• Backpropagation is an algorithm to update weights by computing the gradient of the loss function
• It involves a forward pass and a backward pass, using optimizers like SGD

Forward and Backward Pass

• Forward pass passes inputs through layers and applies biases and functions
• Backward pass calculates gradients ensuring the model learns from its mistakes

###Data Splitting

• It splits data into training and testing sets that ensure the model generalizes well to unseen data to learn patterns
• Without splitting, the model could overfit

Classification

• Binary Classification: Predicts one of two classes and Uses sigmoid
• Multi-Class Classification: Predicts one of multiple classes and Uses softmax
• Multi-Label Classification: Assigns multiple labels and Uses sigmoid

Loss Function Selection

• Regression: Uses Mean Squared Error (MSE)
• Classification:
  • Binary Classification: Uses Cross-Entropy Loss (Log Loss)
  • Multi-Class Classification: Uses Cross-Entropy Loss

Input Layer Activation Function in MLP

• A linear activation function

Activation Funtion in CNN

• CNNs use a linear activation (no transformation)
• ReLU is often applied in hidden layers to introduce non-linearity and prevent vanishing gradients

RNN Activation Function

• RNNs typically use linear in the input layer
• Hidden layers apply tanh or ReLU, where tanh helps with smooth gradient flow and ReLU helps mitigate vanishing gradient problem

Role of Learning Rate

• The learning rate controls step size during weight updates
• High learning rate risks overshooting the optimal solution, while a low learning rate ensures stability
• Adaptive methods like Adam dynamically adjust the learning rate

Neural Network Construction via Pytorch

• Define a class that inherits from nn.Module
• Initialize layers like nn.Linear for fully connected layers, while the forward method defines how data flows
• Use ReLU for non-lineraity
• Trained unsing Adam/ loss function

Mitigate Overfitting

• Dropout randomly deactivates neurons, and TM models have dropout incorporated
• L1/L2 regularization (Lasso/Ridge) penalizes large weights

Saving Models for Future Use

• Save for future use, save the model's state_dict, the architecture if not easily constructible, and the optimizer state

Preventing Overfitting

• Use a larger dataset, reduce the model complexity, and preform hyperparameter tuning prevent overfitting in Neural Networks

Multilayer Perceptron (MLP)

• Sequential data (e.g., time series, speech) which MLPs cannot maintain

Data Importance

• RNNs are sensitive to length, scale, and noise

RNN Function

• RNN function as a type of neural network designed for sequence data
• Allows networks to capture dependencies in sequences like text, speech, or time-series data

Time Series Analysis

• RNNs are suited for series analysis because they can maintain hidden states that capture temporal patterns over time to predict future values such as forecasting and sensor data analysis

RNN Challenges & Pitfalls

• Can learning long-term dependencies

LSTM Networks

• An LSTM (Long Short-Term Memory) is a form of an ANN designed to overcome the vanishing gradient problem in RNN
• It utilizes specialized gates to control the flow of information, allowing it to remember and forget data, making it more effective

LSTM purposes

• LSTMs are beneficial for tasks such as text generation, because they can model long-range dependencies to improve quality
• Manage information through their gate mechanisms regulating flow, retaining important elements

Gradient Challenges & Vanishing

• LSTMs utilize a cell state to enable grads to flow more easily, preserving information

Attention Mechanisms

• Neural networks allow models to focus on relevant parts of the input sequence when making predictions
• Useful for machine translation and text summarization models because they heavily weigh important words

Transformer Structure

• Based on stacked layers of self-attention and feedforward neural networks
• Encoder processes input sequences while the decoder generates an output

Text Embeddings

• Word embeddings represent individual words using Word2Vec
• Sentence embeddings represent entire sentences and can be generated using models like BERT or SBERT

Pre-Training

• Pre-training involves training a model on a large, generic corpus to learn grammar and word relationships
• Masked Language Modeling (MLM) approach is used in models like BERT

BERT Structure

• Stack of transformer encoder layers
• The input is tokenized text, and passes to generate contextualized word embeddings
• Birectional approach meaning it considers context

Transformer Models

• Create text embeddings by tokenizing the input and passing it through layers of self-attention and feedforward

Word vs. Sentence

• Word Embeddings represent individual words, capturing semantic meaning and relationships with other words in a space.
• Sentence Embeddings represent entire sentences, encapsulating the meaning of the sentence and often used in tasks for sentence classification and simularity.

Attention Mechanism

• How does the attention mechanism work in neural networks?

BERT & SBERT

• BERT output dimensions for each token are 768-dimensional
• For SBERT, it's fixed and an SBERT the output d is a 768-dimensional vector, representing the entire sentence rather than individual tokens.

Adjusting a Language Model

• For specific tasks this enables specialization improving its task-specific performance

Optimizing Large Language Models

• Optimized through techniques like fine-tuning, parameter-efficient

Prompt Engineering

• LLMS in designing effective input prompts enabling desired outputs

FineTuning approach LLMs

• parameter efficient this focuses on modifying only a small subset of the model's parameters

Catastrophic

• Catastrophic forgetting occurs when data previously learned has been overfitted
• It can be mitigated by preserving knowledge

Retrieval-Augmented Generation

• RAG) work in LLMs this combines generative models it retrieves relevant information it handles questions requiring external knowledge

Graphs and Prompts

• Prompt Engineering,helps the prompt's understanding and it reduces ambguity

GNN Information

• Information flows edges nodes updates.

LLM Agent

• LLM are enhanced its able external environments
• Standarised and handle Multi-Step.

Sequential memory

• Is used for sequential the planning.

Complex Questions?

• can anayalse and simulate decision makeing ?

LM Agent

• Benifits include Enhancment and provide a context
• Key components include the language model's
• This gives tasks and provides contexts multiple .

Memory & Task

• Short term the long memory helps that over improves agent task the performance.

Chain and Tree Thought

• Planning for LLMs helps the to aproches and approch and solve their probelms.

Tools

• Access or interact externally the tools that enhance those to query or retrieve the respinses dynamically.