Conference Presentation, Keynote, Product Demonstration
Machine Unlearning: Making AI Forget | RAISE Summit 2024 | Paris
- AI is projected to transform lives, industries, and economies positively, though inherent downsides and risks are expected to become increasingly noticeable as regulatory strictness intensifies.
- Companies face significant risks of high fines for non-compliance, underscored by Stanford research indicating that 83% of foundational models violate the new EU AI Act.
- Ongoing legal proceedings involving copyright claims from the New York Times and artists, alongside demonstrated vulnerabilities where models expose raw training data or SSH keys, validate the necessity of permanent guardrails.
- Despite existing guardrails, they are deemed insufficient for preventing unwanted data retention, driving a strategic shift toward "unlearning" processes rather than full retraining.
- The unlearning process is characterized as taking only minutes, costing less than 5% of retraining expenses, and offering a reduced carbon footprint while remaining scalable for removing data from individual users within millions.
- As regulations tighten, organizations anticipate managing daily, numerous requests for user data deletion, with early pilot design partners already engaged and invitations extended for further collaboration.