Conference Presentation, Keynote, Lecture
Explaining AI
- Human-level performance in perception, computer vision, speech, and natural language tasks is anticipated to be gradually reached.
- Microsoft's Xiaoice social chatbot, currently serving 120 million monthly active users, is projected to expand conversation lengths beyond the typical few turns of existing digital assistants.
- AI technologies are expected to drive innovation across diverse verticals, with AI systems predicted to generate 90% of quarterly earning report summaries for Chinese companies.
- AI adoption is foreseen in systems built by external entities for specific use cases, including the Apple car project and credit decisions by Goldman Sachs.
- Bias within training data, rather than developer intent, is identified as a primary source of AI buyers in machine learning systems.
- Continued reliance on unaddressed biased data is expected to perpetuate issues in AI outcomes.
- Development efforts will focus on creating machine learning models that simultaneously achieve high accuracy and explainability.
- Significant business opportunities are anticipated associated with the development and deployment of AI.
- Concerns exist regarding the future emergence of AI decision-making processes that remain incomprehensible to humans.
- Acceptance of unexplainable or unintelligible AI decisions is explicitly rejected.
- A growing necessity is expected to open the "black box" of complex AI models to clarify the rationale behind specific decisions.
- Two divergent approaches to achieving explainable AI are noted: starting with simple models versus starting with highly accurate models and explaining them subsequently.
- Strategic focus is placed on using very accurate, model-agnostic models and applying local approaches to generate explanations.
- Developers face a profound social responsibility to ensure AI transparency, characterized as a defining challenge for the first generation living with AI.
- The prevailing view is that AI should augment human capabilities rather than replace them.
- Unexplained AI decisions in sensitive domains, including political advertising, medical diagnosis, and military operations, pose significant risks of fatal outcomes.