AI Deployment Issues

Week 6

AI - Definition

Artificial intelligence are the technologies that improve our ability to find patterns in data, make predictions, and recommend actions without explicit human instruction.

XAI - Definition

Explainable AI is an emerging field that aims to make the black-box of AI transparent for the client to trust the AI gathered insight.

XAI - Answered questions

XAI answers questions like:

• Why did the AI system make a specific prediction or decision?

• Why didn’t the AI system do something else?

XAI - Solved AI characteristics

1. AI is opaque

2. AI is malleable

3. AI is iffy

4. AI is unproven

XAI - AI is opaque

Issue: AI models rely on sophisticated math and stats that make it hard for users to understand models’ mechanics/outputs.

Solution: Decision tracing

• XAI exposes how the model makes decisions about individual cases or a subset of cases and resolves unfairness, errors, etc.

XAI - AI is malleable (easily influenced)

Issue: AI models learn from the data they are given - good or bad - without pushback or judgment. Therefore, they can produce biased results.

Solution: Bias remediation

• Use high-quality training data to avoid biased results

• Ensure that training data represents reality (Includes both common occurrences and rare events)

XAI - AI is iffy

Issue: AI models produce probabilistic results. They apply patterns and insights they learn from training data to. new instances with some degree of confidence.

Solution: Boundary setting

• Determine how AI model outcomes need to be scoped, limited, or interpreted.

XAI - AI is unproven

Issue: Managers are uncertain about how and if their company can create meaningful returns from AI investments.

Solution: Value formulation

• Articulate how the AI model outcomes influence decisions, processes, and actions (look at value, cost, and risk)

Scalable AI - Definition

Scaling AI means growing the value created by a trained model (scaling up) and its adaptations (scaling out).

Scalable AI - Scaling up

Increase volume of core model use

• Model should be accurate and trusted.

• How has the model been deployed?

  • How many users use it?

  • How often is it used?

  • How often is it retrained?

Scalable AI - Scaling out

Increase number of recontextualized models

• Model should be transferable and inspiring.

• Does the model apply to new contexts?

  • Geographies?

  • Subject areas?

  • Different products?