Published on

AI Foundations

Authors
  • avatar
    Name
    Jegadeesh Vikramanthampi
    Twitter

Notes from the Oracle Certified 2024 AI Foundations Associate certification course.

AI vs ML vs DL

  • Artificial Intelligence (AI) ➠ machines imitating human intelligence.

  • Machine Learning (ML)

    • subset of AI that focuses on creating computer systems that can learn and improve from experience.
    • ML algorithms learn from past data and predict outcomes on new data and or identify trends on past data.
    • uses labeled data for training ML models.
    • use cases include suggesting products while online shopping, movie suggestions, spam mail warning, self-driving cars etc.

  • Deep Learning (DL) ➠ algorithms that learn from complex data using neural networks and predict outcomes and or generate new data.

Types of Machine Learning

  • Supervised Learning

    • classify data and or make predictions.

    • learns relationships between input and output from labeled data.

    • use cases include disease detection, weather forecasting, house price prediction, stock price prediction, spam detection, sentiment analysis, credit scoring etc.

    • Variations

      • Linear Regression
        • used when the output label is continuous - e.g., house price predictor.
        • fits a straight line to the data.
        • Loss ➠ indicates how far the predicted value is from the actual value.
      • Classification
        • supervised ML technique used to categorize or assign data points into predefined classes based on their features or attributes. Uses labeled dataset.
        • used when the output label is categorical - e.g., spam detector.
          • Binary ➠ if prediction value is either true of false.
          • Multi-class ➠ for example, in sentiment analysis, prediction outputs are either positive, negative or neutral.
          • Logistic Regression
            • predicts true or false instead of something continuous such as house prices.
            • unlike linear regression, fits an S-shaped curve called sigmoid function to the data.

  • Unsupervised Learning

    • type of ML that doesn't use any labeled outputs.
    • understands relationships within datasets and extracts trends from data. Can also learn patterns in the data and group similar data items.
    • use cases include detecting fradulent transactions, customer segmentation, outlier detection, targeted marketing campaigns etc.
    • Clustering
      • grouping of similar data items.
      • data items are more similar within a cluster than the items outside of it.
      • data items that are not similar within a cluster are termed as outliers.
      • use cases include market segmentation, outlier analysis, recommendation systems etc.
      • Similarity ➠ how close two data points are to each other - usually a value between 0 and 1.
      • Clustering Workflow
        • prepare the data ➠ remove missing values, normalize data, perform feature scaling etc.
        • create similarity metrics ➠ choose similarity metrics based on nature of data and clustering algorithm used - frequently used are euclidean, hamm, manhattan, cosine etc.
        • run clustering algorithm ➠ use the chosen similarity metrics to cluster the data - common algorithms used are partition, hierarchical, density and distribution based.
        • interpret results and adjust clustering ➠ check quality of clustering output by verifying against expectations iteratively.

  • Reinforcement Learning (RL)

    • type of ML that enables an agent to learn from its interactions with the environment while receiveing feedback in the form of rewards or penalties without any labeled dataset.
    • makes decisions or choices and solves tasks by trial and error.
    • use cases include industrial automation, autonomous vehicles, smart devices, gaming and entertainment, healthcare etc.
    • RL Terminologies
      • Agent ➠ interacts with environment, takes actions and learns from feedback.
      • Environment ➠ external system with which the agent interacts.
      • State ➠ representation of the current situation of the environment at a particular time.
      • Action ➠ possible moves or decisions the agent can take in each state.
      • Policy ➠ strategy the agent uses to decide which action to take in a given state.
    • RL Training Loop
      • goal is to find a policy that will yield a lot of rewards for the agent.
      • optimal policy is learned through training by using algorithms like Q learning or Deep Q learning

Deep Learning (DL)

  • subset of ML that focuses on training Artificial Neural Networks (ANNs) with multiple layers.

  • automatically learns and extracts intricate representations and or features from raw and complex data.

  • internal representation of data is built using extracted features and DL algorithms allow parallel processing of data leading to better scalability and performance.

  • Types of DL algorithms/models and their use cases

    DL Algorithm/ModelsUse Cases
    Convolutional Neural Networks (CNN)

    Image classification and segmentation, object detection and facial recognition.

    Detect patterns and features.

    Recurrent Neural Networks (RNN)

    Process sequential, time series and natural language data.

    Has a feedback loop to capture hidden states to manage temporal dependencies.

    Autoencoders

    Unsupervised learning models used for feature extraction and dimensionality reduction.

    Commonly employed in data compression and anomaly detection.

    Transformers

    Process sequential, time series and natural language data.

    Generative tasks like Question Answering (QA), text summarization and text to image generation.

    Long Short-Term Memory (LSTM)

    Specialized RNN variant that can process sequential, time series and natural language data.

    Can handle long-term dependencies in sequential data.

    Diffusion ModelsGenerating images, text to image generation etc.
    Generative Adversarial Networks (GAN)Generating audio, images, text to image generation etc.

  • Artificial Neural Network (ANN)

    • inspired by the human brain and consists of interconnected nodes called neurons.
    • Building Blocks of ANN
      • Layers ➠ input, hidden and output layers receive inputs, transform them and produce outputs.
      • Neurons ➠ computational unit from an input that produces output.
      • Weights ➠ determines the strength of connection between the neurons.
      • Activation Function ➠ works on the weighted sum of inputs to a neuron and produces an output.
      • Bias ➠ additional input to a neuron that allows a certain degree of flexibility.
    • ANNs are trained using backpropagation algorithm.
      • Backpropagation algorithm ➠ adjusting the weights of neurons based on a calculation during training to correct errors in output.

  • Sequence Models

    • input data is in the form of sequences and the goal is to find patterns and make predictions.

    • Use Cases

      • Natural Language Processing (NLP) ➠ sentiment analysis, machine translation, text generation etc.
      • Speech Recognition ➠ convert speech/audio to text.
      • Music Generation ➠ DL sequence models can create new music compositions.
      • Gesture Recognition ➠ sign language recognition.
      • Time Series Analysis ➠ in the field of finance and weather, used to predict future values.

    • Recurrent Neural Network (RNN)

      • type of sequential DL model that handles sequential data.
      • unlike Feed Forward Neural Network (FNN), allows information to persist using a feedback loop.
      • a key feature of RNN is the ability to maintain a hidden state/memory that is updated after each element in the input sequence is processed.
      • hidden state is used as the input for the network for the next time step allowing the model to capture dependencies and patterns across data spent across time.
      • RNNs do not perform well with capturing long term dependencies - due to vanishing gradients problem that can be addressed by using Long Short-Term Memory (LSTM) models.
      • Types of RNN
        • One-to-One ➠ similar to feed forward neural network and not suitable for sequential data.
        • One-to-Many ➠ provides multiple output values for one input value. Use cases includes music or sequence generation.
        • Many-to-One ➠ provides one output value after receiving multiple input values. e.g., sentiment analysis based on reviews.
        • Many-to-Many ➠ provides multiple output values for multiple input values. e.g., machine translation and entity recognition.

    • Long Short-Term Memory (LSTM)

      • works by using a specialized memory cell and gating mechanisms to capture long-term dependencies in sequential data.
      • overcomes the vanishing gradients problem by selectively remembering or forgetting information over time.
      • LSTM Workflow
        • Input Processing ➠ inputs current data point in the sequence.
        • Previous Memory ➠ receives the previous hidden state values that represent what the system had remembered and forgot up to the current point in the sequence.
        • Gating Mechanism ➠ similar to filters that control the flow of information between cells in the system.
          • Input Gate ➠ decides what new information from the current input should be added to the memory cell.
          • Forget Gate ➠ decides what information in the current memory cell should be discarded or forgotten.
          • Output Gate ➠ regulates how much of the current memory cell should be exposed as the output of the current step.
        • Updating Memory ➠ updates the cell state by using information from both input and forget gate.
        • Output Generation ➠ LSTM uses the output to produce the current hidden state which becomes the output of the LSTM for the next time step.

  • Convolutional Neural Network (CNN)

    • type of DL model designed to process grid-like data. e.g., images and videos.
    • CNN Layers
      • Input Layer ➠ accepts images with height, width and depth representing RGB color channels.
      • Feature Extraction Layers
        • automatically learns and extracts relevant patterns and features from input images.
        • Building Blocks
          • Convolutional Layer ➠ applies convolutional operations to the input image using small filters called kernels.
          • Activation Function ➠ allows the network to learn more complex and non-linear relationships in the data.
          • Pooling Layer ➠ helps to reduce the computational complexity and spatial dimensions of the feature maps generated by the convolutional layers.
          • Fully Connected Layer ➠ responsible for making final predications or classifications based on the learned features.
          • Softmax Layer ➠ converts the output of the last fully connected layers to probability scores.
          • Dropout Layer ➠ regularization technique used to prevent overfitting in the network. Similar to that of a quality checker.
      • Classification Layers ➠ fully connected output layer where classification occurs.
    • CNN Limitations
      • requires massive data and computations.
      • subject to overfitting - happens with limited training data.
      • typically considered as black box models thereby difficult to interpret.
      • sensitive to small changes in the input leading to unstable predictions.
    • CNN Applications
      • image classification, object detection (draws bounding boxes around objects in an image)
      • perform pixel-level image segmentation.
      • face recognition tasks.
      • medical image diagnosis.
      • autonomous vehicle development (identify pedestrians, traffic signs and other vehicles)
      • analyze satellite images and remote sensing data (for tasks like land cover classification and environmental monitoring)

OCI AI Portfolio

  • AI Services

    • Digital Assistant
      • platform to create and deploy digital assistants that are AI-driven interfaces to help users accomplish a variety of tasks with natural language conversations.
      • interacts with the user, lists what it can do, routes user's requests to skills and handles disambiguation.
    • Language
      • pretrained models trained on industry data to perform language analysis with no data science expertise required.
      • Pretrained Models ➠ supports language detection, sentiment analysis and key phrase extraction.
      • Custom Models ➠ named entity recognition and text classification.
      • Text Translation ➠ natural machine translation is used to translate text across various languages.
    • Vision
      • provides two main capabilities ➠ image analysis and document AI
        • Image Analysis Variations
          • Object Detection ➠ feature that detects objects inside an image using a bounding box and assign a label to each object with an accuracy percentage.
          • Image Classification ➠ assigns classification labels to the image by identifying the major features in the scene.
      • both pretrained and custom models are available for use. Users can retrain the models with own unique data to fit specific needs.
    • Speech
      • transcribes audio and video files using DL techniques. No data science expertise required.
      • processes data directly in object storage and generates timestamped accurate transcriptions.
      • converts media files to readable texts and text transcriptions stored in JSON and SRT formats.
      • supports english, spanish and portuguese currently and more languages on the way.
      • batching support for multiple files in one call and fast processing (10 hrs of audio in < 10 mins)
      • makes transcribed text more redable using normalization. Provides profanity filtering as well.
    • Document Understanding (DU)
      • user can upload documents to detect and classify text and objects in them.
      • can process individual files or batches of documents.
      • in optical character recognition (OCR), DU can detect and recognize text in a document.
      • in text extraction DU provides word and line-level text and the bounding box coordinates of where the text is found.
      • in key value extraction DU extracts a predefined list of key-value pairs from receipts, invoices, passports and driver IDs.
      • in table extraction DU extracts contents in tabular format, maintaining the row and column relationship of cells.
      • in document classification DU classifies documents into different types.

  • ML Services

    • Core Principles of OCI Data Science

      • Accelerated ➠ allow data scientists to work how they want and provide access to automated workflows, best of open-source libraries and a streamlined approach to building models.
      • Collaborative ➠ enable data science teams to work together with ways to share and reproduce models in a structured and secured way for enterprise-grade results.
      • Enterprise-Grade ➠ provide a fully managed platform built to meet the needs of the modern enterprise.

    • Data Science Features and Terminology

      • Projects ➠ containers that enable data science teams to organize their work.
      • Notebook Sessions ➠ provides a JupyterLab environment with pre-installed open source libraries and ability to add others.
      • Conda Environments ➠ open source environment and package management system created for python programs.
      • Acceleratd Data Science (ADS) SDK ➠ python library to simplify the steps in data science workflow.
      • Models ➠ mathematical representation of your data and business process.
      • Model Catalog ➠ a place to store, track, share and manage models. Centralized and managed repository of model artifacts.
      • Model Deployments ➠ used to deploy models stored in model catalog as HTTP endpoints on managed infrastructure.
      • Jobs ➠ used to define and run a repeatable machine learning task on fully managed infrastructure.

  • AI Infrastructure Overview

    • GPU Architecture
      • allows many processes at the same time with parallel computing.
      • performs many computations very rapidly and crunches through extremely large datasets at tremendous speeds.

  • Responsible AI

    • should follow applicable laws.
    • should be ethical.
    • should be robust both from a technical and social perspective.
    • Cycles and Roles
      • setup governance to be put in place.
      • develop policies and procedures.
      • ensure compliance.
      • typical roles involved are developers, deployers and end users.
    • Challenges in Healthcare
      • ensuring fairness and lack of bias.
      • ensuring transparency and explainability.
      • regular evaluation.

  • OCI Generative AI Service

    • Vector Search Using Oracle Database 23ai
      • simple to combine relational and AI vector search in single query.
      • efficient orchestration of Gen-AI pipelines.
      • SQL support for vector generation.
      • vector data type - dimension format can be either INT8, FLOAT32, FLOAT64
      • simple vector spec - to support newer ML embedding models in the future. Schema can stay same and apps don't need to be rewritten.
      • similarity search using SQL syntax and functions.
        • Vector Distance Function ➠ distance between two vectors is smaller for entities that are more similar.
        • Supported Distance Functions ➠ cosine (default), euclidean, euclidean_squared, hamming, manhattan and dot.
        • Vector Search SQL ➠ used to find the top k closest matches to a given query item.
        • Vector Index Syntax
          • similar to basic index creation. Organization for an index is based on in-memory fitment.
          • Use IN MEMORY NEIGHBOR GRAPH if the index data will fit in-memory. Else use NEIGHBOR PARTITIONS
          • TARGET ACCURACY clause is used to indicate the default accuracy the index should provide for similarity search queries.
      • Approximate Search Indexes ➠ use the new APPROXIMATE keyword in the row limiting fetch clause to perform a similarity search using a vector index.
      • database-native vector embedding generation.

  • Autonomous Database Select AI

    • simplest way to get answers about your business.
    • use your language to query data. No need to understand where and how your data is stored to gain insights.
    • autonomous database manages the entire query process to produce your answer.
    • Select AI translates your language into Oracle SQL Language using Large Language Models (LLMs)
    • developing apps with select AI is simple, future-enabled and secure.
    • Select AI workflow
      • first, use one or more select AI profiles to choose an LLM to generate a DB query from natural language.
      • then specify schemas, tables or views to participate in the processing.
    • SQL query generation process flow
      • Select an AI profile ➠ choose between different LLMs for prompt generation.
      • Produce prompt ➠ LLM generates prompt for the question asked.
      • Generate SQL using LLM ➠ LLM generates the SQL using the model.
      • Return query results ➠ LLM then returns the result set or SQL to tools like SQL developer or an application.

References and Further Reading