Short Talks
summary
Open-source NVFlare is production-ready since 2021 and runs the 20-hospital COVID oxygen predictor, supports on‑prem IP protection, and surfaces in 50–60+ cancer projects; Eli Lilly shares base models with startups via Flare for fine-tuning. Google workflows push computation into TEEs for verifiable privacy or emit synthetic data for re‑use. A modular mixture‑of‑experts branching lets institutions add or drop experts without retraining. Flower shows a 1.3B Collective One trained on ~1M users and a low‑bandwidth optimizer that beats DDP under constrained links.
AI-written summary · about
chapters (49)
- 0:00Federated learning with NVIDIA Flare
- 1:01NVIDIA Flare overview and production use
- 3:15Key Flare features for federated learning
- 4:01Hierarchical federated learning at scale
- 5:08Industry use cases for NVIDIA Flare
- 7:18Eli Lilly’s federated learning case study
- 8:43Cancer AI Alliance medical research
- 9:34Federated AI for Therapeutic Engineering
- 10:18Taiwan clinical trial validation platform
- 11:58SWIFT fraud detection with federated learning
- 13:03Self-driving car training across countries
- 14:35Earth-2 weather modeling data mesh
- 15:21Federated learning challenges and innovation
- 16:08Simplifying federated learning
- 16:51Legal, regulatory and security hurdles
- 18:27Data governance and incentive models
- 21:28Google’s trusted execution environments
- 23:49LLM challenges in federated learning
- 25:02TEE-based learning for foundation models
- 26:32Synthetic vs direct evaluation in federated learning
- 27:43Prompt tuning and personalization in LLMs
- 29:14Google’s verifiably private federated compute
- 31:17Collaborative learning for foundation models
- 33:16Why collaborative learning is essential
- 34:01Fragmented data in LLM training
- 37:15Limitations of traditional federated learning
- 39:12Modular architecture for LLM pre-training
- 40:22Mixture of experts for foundation models
- 43:51Q&A on collaborative learning
- 44:49Privacy and security in federated systems
- 45:11Differential privacy and cryptographic tools
- 46:36Protecting servers from malicious clients
- 47:45Designing primitives for federated frameworks
- 48:30Federated statistics and simpler tasks
- 49:16Learning distributions with Earth Mover Distance
- 51:47Instance optimality in distribution learning
- 54:24Beyond worst-case federated statistics
- 55:03Flower’s transition to federated models
- 56:32Real-world federated learning success
- 58:50Flower’s growth and community
- 1:00:18Federated foundation models with Flower
- 1:01:23FlowerTune LLM Leaderboard
- 1:02:30Pre-training with Flower’s Photon system
- 1:03:22Desynced adaptive optimizers for LLMs
- 1:04:32Flower’s stack and collaborations
- 1:05:20Vana partnership for data DAOs
- 1:06:06Government partnerships for sovereign AI
- 1:06:56Adoption of Flower’s innovations
- 1:07:17Flower AI Summit 2025 closing
related talks
31mFrom Theory to Practice: Advances in Real-World Federated Learning
30mFederated Learning from simulation to production
39mOpenFL: Building Better AI Models with Private Data
39mFlare & Lantern: Accelerators for Spark and Deep Learning
26mOpenFL (Federated Learning) Building better AI models with private data
32mInnovation in the Age of Regulation: Federated Learning with Flower
27mFederated Learning with Azure Machine Learning NVIDIA FLARE and MONAI
32mOpenFL: A Federated Learning Project to Power (and Secure) Your Projects