Paul Borrego

1.2

Course objectives

  • Identify examples of real-world AI applications
  • Know when and when it is not useful
  • Know the capabilites of the AI

AI:
Use cases:

  • Retail: Product review summaries, pricing optimization, virtual try-ons, layout optimization
  • Healthcare: Personalize medicine and scan medical imaging
  • Life Sciences: Scanning protein folds and other complex biological formations en masse for better understanding, drug discovery
  • Financial Services: Fraud detection, debt collection, portfolio management
  • Manufacturing: Product optimization, predictive maintinence, product design
    Examples:
  • Intelligent document processing (IDP):
    • Application that classifies information from unstructured data into summaries
  • Fraud detection:
    • By scanning many transactions it can flag those that are suspicious

ML:
More useful when rules are complex and not very binary (Like determining what is spam in email)

Use cases:

  • Supervised:
    • Classification (identifying new data by analyzing old) and regression (predicting new instances based on existing data)
      • Fraud detection and image classification
      • Weather prediction or any forward predictions
  • Unsupervised
    • Clustering (Finds similar features between points) and dimensionality reduction
      • Clustering includes recommended systems
      • Dimensionality is for compression
  • Reinforcement
    • Useful when you can give correct answer but path is not known

Generative AI:
Automate tedious tasks and can analyze data freeing up time for workers to do more creative work

Use cases:

  • Wider range as this AI is more adaptable
  • Very responsive and can be used in live service
  • Good at simplifying complex tasks

Challanges:

  • Regulatory violations:
    • Can generate outputs that exposes personally identifiable information (PII)
    • To reduce risk, make sure data is anonymized and the agent is only given whats needed
  • Social risk:
    • Might create unwated content
    • Just test
  • Data security:
    • Ensure data that you share with model does not violate privacy laws
  • Hallucinations:
    • These models can't be trusted 100% off the time.
    • Users should be aware that they are using an FM and this can result in halucinations
  • Nondeterminism:
    • These are nondeterministic models and will give different outputs per input
    • You can run the model multiple times for consistency

Business metrics for generative AI:

  • User satisfaction: Gather feedback and see what customers think
  • Average revenue per user (ARPU): Calculated profits per user attributed to AI. Sees if this product is generating more money if it cost
  • Cross-domain performance: Ensures that different uses are all efective. (Ex: montoring if AI is effective on a multidomain e-commerce platform)
  • Conversion Rate: convert clicks or sign ups to purchases
  • Efficiency: How effective the model is getting its answers