Intro to Machine Learning Concepts

• Text concerns processing data for machine learning (ML)

ML Concepts

• Machine learning uses mathematical models to make predictions from data at a scale impossible for humans.
• Traditional analytics systematically analyzes large datasets (big data) to identify patterns and trends for actionable insights.
• Traditional analytics uses programming logic to answer questions from data.
• Traditional analytics is suitable for structured data with a limited number of variables.
• Machine learning uses data examples to learn and answer questions.
• Machine learning is useful for unstructured data with complex variables.
• Historical data trains a model, where standardized methods define model rules.
• An algorithm combines a trained dataset to infer outcomes.

Types of ML Models

• Supervised models receive inputs and related outputs for training.
• Supervised models identify fraudulent transactions using past examples.
• Unsupervised models find patterns in training data without assistance.
• Unsupervised models group users with similar viewing patterns to make recommendations.
• Reinforcement models learn from their environment and take actions to maximize rewards.
• Reinforcement models are used to develop self-driving cars.

AI Subcategories

• Artificial intelligence (AI) has various subcategories with increasing complexity like Machine learning, deep learning and Generative AI
• Compute power required increases with the complexity
• In ML, the label or target is what can be predicted
• Features are attributes that can be used to predict the target.
• 1970: Expert systems emerge and research continues to improve decisions using computers. 2000: Data-driven neural networks produce usable solutions for production applications.
• 2010: GPUs accelerate ML model training.
• 2020: the innovation of transformer architecture.

ML Lifecycle

• The ML lifecycle begins by defining a business problem and framing it as an ML problem.
• The lifecycle involves collecting data and preparing it for use in an ML model.
• Model development uses the prepared data
• Developed model is deployed to production.
• The final phase involves monitoring the model in production.
• ML processing includes data processing steps such as ingestion, storage, processing, analysis, and visualization through predictions
• Iterative phases includes identifying business goal, framing problem, processing data, developing, deploying, and monitoring the model

Process Data Phase

• The process data phase involves data collection and preparation.
• Pre-process includes collecting, pre-processing, and engineering features.

Roles in ML Lifecycle

• Data scientists define the business goal and create the ML model.
• Domain experts and business analysts provide insights and knowledge.
• Data engineers collect the data.
• AI/ML architects oversee the entire process.
• ML engineers handle deployment and MLOps

Framing the ML problem to meet the business goal

• Identify what pain the business problem is causing and why it should be resolved by measuring success.
• An example of the statement is that AnyCompany (car insurance ) has seen increase in global fraud costing millions
• Stating the business goal as to improving the process to identify fraud for new claims by the end of calender year.
• Key steps to the ML problem include framing, determining feature and label
• Quantify metrics for data sourcing, evaluate appropriate method and start with simple model
• Determining if ML is the best approch requires reviewing use of the case, relevant training data and cost.
• It is ideal to State the business goal in the form of the problem statement and Involve data scientist for guidance.
• The Business problem drives whether ML is needed or if the simpler solutions can meet the need.

Collecting Data

• Protect data veracity, collect data to test and train model and apply known targets
• Verifying integrity of data sources, protecting ingestion pipeline, validating datasets and auditing the ingestion process for veracity of data.
• 70-80% of data is required for training model, 20-30% for validating
• The data engineer and scientist insure the correct collection of data to support training
• ETL ingestion process involves extracting , transforming, and loadin data
• Labelling process requires adding labels to training data using tool such as SageMaker Ground Truth
• Types of labelling includes, tabular data, natural language processing, images, audio etc

Preprocessing data

• It is ideally the first step in processing data for ML
• involves iterative preparing and modelling and exporatory data analysis
• it transforms the data and handles features and characteristics of it. this impacts development of algorithms.
• Strategies includes: Cleaning, partitioning , scalling , augmenting, balance and formatting data
• Data formatting and conversion example entails converting text to a nuemric field.

More Notes on Preprocessing

• In exploratory data analysis, you find patterns, guide algorithm selection, and inform feature engineering.
• The Data scientist explores and visualies the data to get a feel for the data and the scientist is incharge for data preprocessing to bring data the correct state for training
• Examples include partitioning, balancing, and data formatting.

Feature Engineering

• Feature Engineering is about improving the usefulness of the data. It converts existing features with algorithms
• This approach gives more understanding to the algorithm and improves speed and access.
• Feature creation and transformation generates new features with column sets
• Feature extraction and selection reduce dimensionality by removing measureable value of output
• Improvement existing features , creation, transofmration and extraction reduces dimensionality.

Developing Model

• Consists of building model: training, tining and evaluate
• Process of training and tuning are iterative and continue until the accuracy rate lines upto the business goal and dataset.
• Untagged Test data is to Validate the result from the train data sets.
• Training a model with reasounrce intensive and could require the use of big data frameworks of HPC systems(High performance computing)

Types of accuaracy Metrics

• Positive means :That The target is a dog
• • True postive : The model correctly predicted that the target was present.
• • True negative: The model correctly predicted that the target was not present.
• False postive : The model predicted that the target was present, but it wasn't. and Faise negative: The model predicted that the target was not present, but it was.
• Accuaracy: Measures the total number of correct predictions out of total samples, precision measures the # of true positives and Recall measures the true positive that were compared in all Positives.

Deploying a model and infrastructure for it

• infrastructure is scale to meet production traffic.
• After that the infrastructure is deployed, code with the model is ready.
• the process takes inferenes on data by using production data to provide feedback , this includes , ingestion of new data and individual samples(Input)
• the model provides an output for each training and tuning as well.
• the production model is typically different then the model from training
• to use MLOPs approach with storeaFeatures,data,code,moduel data use following approach -Collect data/ Extract-loading processing,pre process data, engineer features, deploy model and than monitor model.

AWS (Amazon Web Services) ML Infrastructure

• AWS ML infrastructure layers consist of compute, networking and storage
• It also involves framweorks and workflow services
• Service for compute handle power and speed to train and predict in realtime use and tools to simplify. the tools are EMR,DLAMI and Amazon Sage Maker
• AWS includes Sagemaker which provides integrated workbench. Studio , Data wrangler, notebook and processing

Note AWS - compute

• It provides instance that have High Performance for ML, designed deployment , attached GPU which is low cost to Sagemaker instances
• provides the networking interfaces to speed compute for tasks training and running data

AWS- Storage

• Storage service object can be a Amazon S3 , that has thousands transaction
• Amazon EBS- digit millsecond latency
• Amazon S3-Object achieve thousands transactions
• Amazon workfow Amazon es2 to Docker container images and amazon Es3 for managing container apps

AWS and Generative AI

Key Generative AI concepts

• Foundational model- A large pre-trained model with capabilities
• Large language model (LLM): A model that is trained to use using naturally using text
• prompts : Instructions giving give to FM and LLM
• Prompts and engineering : To to develop prompts that produces optimal results
• AWS services to help with foundation such as
• Amazon SageMaker jumpstart, Bedrock,Amazon Q Developer
• AMAZo. is a powered coding system. to help with build up faster , scane security, extrend and maintain .

AI/ML services on AWS:

• AWS provides many AL/MI such as recognizing and process data
• Amazon Rekognition
• detects Text, ,faces,logo,Object and images and video data -Amazon transate
• extract text , print , handwriting, layout, date and data from document.
• Includes insights and insights and connection for nature processing. -It uses natural language to translate large volume of text for content with Amazon transalate