Intro to Machine Learning Concepts

ML Concepts

• Machine learning uses mathematical models to make predictions from data.
• Machine learning works at a scale that is difficult or impossible for humans.
• ML uses examples from large amounts of data.
• ML is useful where variables are complex or data is unstructured.
• Traditional analytics systematically analyzes large datasets (big data) to identify patterns and trends.
• Traditional analytics answers questions using programming logic.
• Analytics are useful for structured data with a limited number of variables.

Algorithms & Models

• Algorithms train ML models.
• Data, historical in nature, is used to train a model.
• Algorithms are standard methods that define model rules.
• A model is a function, combining a trained dataset and an algorithm to infer outcomes.
• These result in a prediction of what should happen.

Types of ML Models

• There are three general types of ML models including supervised, unsupervised, and reinforcement learning.
• Supervised learning occurs when the model is given inputs and related outputs for training.
• An example of supervised learning is identifying fraudulent transactions using past examples.
• Unsupervised learning occurs when a model finds patterns in the training data without help, also known as labels.
• An example of unsupervised learning is grouping users with similar viewing patterns for recommendations.
• Reinforcement learning occurs when the model learns from its environment and takes actions to maximize rewards.
• An example of reinforcement learning is the task of developing self-driving cars.

Subcategories of AI

• Artificial intelligence is the overarching category.
• An extension of artificial intelligence is machine learning.
• Deep learning is a subcategory of machine learning.
• Generative AI is a subcategory of deep learning.
• Each category can solve more complex problems than the previous one.
• They each require more compute power than the previous one.

The Evolution of ML

• In 1970, expert systems emerged along with academic research to improve decision-making with computers.
• In 1990, cloud computing emerged beginning to reduce costs of compute.
• In 2000, data-driven neural networks produced usable solutions for production applications.
• In 2010, graphic processing units (GPUs) accelerated the speed of training ML models.
• In 2020, the transformer architecture was innovated.

ML Data Concepts

• In Machine Learning, there are some key concepts in play which include:
  • Features include variables such as months as a customer, age, umbrella limit, and vehicle claim.
  • A Target, also known as a Label, is what a model is trying to predict

Key TakeawaysML

• ML models are functions that combine a trained dataset and an algorithm to predict outcomes.
• Three general types of machine learning: supervised, unsupervised, and reinforcement.
• Deep learning is a subcategory of ML using neural networks.
• Generative AI is a subcategory of Deep learning which can generate content and is trained on large amounts of data.
• The label, or target, is what is trying to be predicted in ML, while features are the attributes used to predict the target.

The ML Lifecycle

• ML lifecycle includes data ingestion, storage, processing, the analysis and visualization.
• This results in predictions.
• The ML lifecycle is iterative and has phases that include:
  • Identifying the business goal
  • Framing the ML problem
  • Processing data
  • Developing a model
  • Deploying the model
  • Monitoring the model

Process Data Phase

• The process data phase includes collecting and preparing data.
• When processing data, you also collect data, pre-process data, engineer features, and prepare data.
• Common roles in the ML lifecycle include data scientists, AI/ML architects, domain experts, business analysts, data engineers, ML engineers, and MLOps engineers.

The ML Lifecycle, Continued

• It starts with identifying a problem framing it as an ML problem.
• The next step is to collect data and prepare it for use in an ML model.
• Model development takes the prepared data and hands it off for deployment.
• The final phase is monitoring the model.

Framing the ML Problem

• Working backward starts with asking what the business problem is.
• Then, you should ask what pain the problem is causing.
• Next, you should ask why the problem needs to be resolved.
• You should also ask what will happen if it is not solved.
• Finally, decide how to measure success.
• A real world example occurred at AnyCompany, a car insurance provider.
• They saw an increase in global insurance fraud costing them millions of dollars in financial losses, administrative overhead, and investigation activity.
• The business goal was to improve their process to predict fraud for new claims by the end of the calendar year.

Key Steps in Framing ML

• Determine what will be observed (feature) and what will be predicted (label or target).
• Establish observable, quantifiable metrics.
• Create a strategy for data sourcing.
• Evaluate whether ML is the correct approach.
• Start with a simple model and iterate.

Determining the Best Approach

• Start by reviewing the potential use case.
• If business rules can be hard coded and the number of variables is limited, build a BI or analytics solution.
• If these conditions are not present, consider an ML solution.
• Ensure enough relevant, high-quality training data is available.
• Ensure ML can provide the necessary level of accuracy, latency, and transparency.
• Assess whether the organization can support the cost and resources to build and sustain the ML solution.
• Finally, assess whether the financial cost of implementing the ML solution is a good match to the impact of the problem.

Framing Final Thoughts

• State the business in the form of a problem statement.
• Data scientists work with domain experts to determine the appropriate ML approach.
• Determine if ML is even needed or if a simpler solution can meet the need.

Collecting Data

• Data collection is the first step in processing data.
• Once collected, Data is cleaned, processed, and transformed.
• The data engineer builds a collection pipeline.
• The data engineer extracts data from data sources, loads them into an Amazon S3 data lake, and transforms them.
• From here, the data scientist pre-processes the data and engineers features.
• Key steps for data collection: protect data veracity, collect enough data and apply labels.

Protecting Data veracity

• This involves verifying data sources, protecting the ingestion pipeline, validating datasets and auditing.
• You always want good data.
• You don't want bad data as the foundation of the model.

Collecting data

• It is important to have enough data for training and testing purposes.
• Training data is 70-80% of data and teaches models
• Test data, which is 20-30% serves to validate models
• When put into production data ,predictions are made in your live application.
• This also feeds additional training and tuning.

Key Takeaways

• Collecting data is similar to extract, load and transform (ELT) processing.
• It is important the data engineer and data scientist ensure there is enough of the correct data to support training and testing models.
• At times, labels must be added to training data.

Applying Labels

• Labeling provides context for ML models.
• Humans add label values to training data and ML models then learn from the labeled data to identify patterns.
• Common types of labeling include photos, tabular data, natural language processing, and audio processing.
• Amazon SageMaker Ground Truth allows the offloading of labeling work with an expert workforce.

Pre-processing

• A data scientist prepares this data iteratively.
• Exploratory data analysis is done while extracting data to find any important data.
• Preprocessing data includes cleaning, partitioning, scaling, augmenting, and balancing.

PreProcessing Strategies

• Cleaning the data involves removing outliers and duplicates as well as inaccurate data.
• Partitioning prevents models from overfitting during training and validation.
• Scaling helps keep target values close to normally distributed for ML algorithms to work efficiently.
• Augmenting synthesizes additional data and it helps to stop overfitting.
• Balancing helps mitigate imbalances in feature values.
• Formatting and converting helps ensure all the various inputs are able to work well with the algorithms.

Exploratory Data Analysis

• It allows you to find patterns, select relevant algorithms, and inform feature engineering.
• Often a clustering algorithm is run to identify customer segments that might be relevant to predicting behavior.

Key Takeaways - Preprocessing Data

• Data preprocessing puts data into the correct shape and quality for training.
• Data scientists performs preprocessing using a combination of techniques/expertise.
• Exploring and visualizing data helps data scientists get a feel for the data.
• Examples of preprocessing strategies include partitioning, balancing, and data formatting.

Feature Engineering

• It exists to improve usefulness of data.
• It is performed by refining models and choosing relevant algorithms
• Its process is to convert already existing features into better ones to allow for improved predictive accuracy.
• Improved features allow algorithms to understand datasets better.

Creation and Transformation

• The practice of binnning values is when values that could fall in continous ranges are put in to fixed number values to create categories for the algorithm to analyze

Extraction and Reduction

• This allows removal of features that don't add meaning for predictions to come from.

Key Takeaways - Feature Engineering

• It improves how useful features were previously to improve predictive outcomes.
• Transformation focuses on adding better more useful data in.
• Extraction and selection are the process of reducing dimensionality.

Model Development

• Model building includes training, tuning, and evaluating.
• Training uses the training dataset.
• Tuning involves monitoring error rates, with the goal of reaching the desired accuracy level.
• Evaluation involves validating results with unlabeled test data.
• There are accuracy metrics with 4 values:
  • True positive, correct target.
  • True negative, target was not present.
  • False positive, model predicts target, but it wasn't.
  • False negative, model predicts target isn't present, but it was.

Model Building Key Takeaways

• Model development consists of all aspects of the models themselves.
• Training aims for accuracy and being in line with the overall business goal.
• Unlabeled test data is important to find results as well.
• Building resoucre intensive applications can lead to big data and having to use big big data frameworks.

Model Deployment

• Model deployment is composed of code in an operational environment.
• Production models should handle data, and individual samples should be able to be put as inputs.
• The models generate outputs for each of the inputs.
• MLOps is an approach taken during the lifecycle to ensure models are used and taken care of.

Key Takeaways - Model Deployment

• Training infra is very diffrennt from training infra, and automation allows the models to work very well.
• Automating lifecycle is very helpful in creating models.
• MLOps takes the streamlined data and uses resources that it deems best.

ML on AWS

• AWS infrastructure has compute, networks,storage, framework layers, and work services.
• Services are designed to work together to train models, networks allow data and models to translate.
• AWS' workflow services will simplify building the model and managing them.
• EC2, GPU and network systems ensure models work to their full capacity.

Sagemaker

• Studio simplifies visual interfaces for all ml models.
• Processing simplifies experience.
• Canvas is no-code and provides business analysts a tool to provide great AI with their own data.

Sagemaker and Model building

• Allows models from S3 buckets to be brought to sagemaker.
• Enbales built-in algorithms.
• Enbales pretrained models to have extra assistance.

Sagemaker - Key Takeaways

• Fully automated service.
• Models are built using studio or notebooks.
• Low code solutions provide the power to create great models from small business analysts.

Generative AI

• Are deep learning models.
• Allow for models to do better and better.
• Prompt is instruction given to model that will in turn create perfect results.

Generative AI - Concepts

• Bedrock is an AI service from amazon focused on Generative models with APIs.
• Jumpstart is for more beginner level users who are getting jumpstarted into the concept.
• Q Dev is for security devs who work with code and want to detect security flaws with the AI.

Q- Dev

• Allows users to scan code, has a powerful AI for building solutions, and is secure by design.
• You can code by chatting with your integrated development environment, you can get instant guidance from questions, test your code with suggestions, and make your models more concise.