From: mk_thisisit
Professor Artur Dubrawski, a professor in Computer Science at Carnegie Mellon University, discusses the emergence and benefits of a new paradigm in artificial intelligence (AI): distributed AI and federated learning [00:00:30], [00:25:17].
Limitations of the Current AI Paradigm
The current paradigm of artificial intelligence is considered “impractical” [00:00:36], particularly for large organizations with extensive and scattered data resources [00:23:30], [00:23:42]. This paradigm typically involves:
- Centralized Data Collection Collecting vast amounts of data from different sources into one location [00:24:20].
- Data Harmonization Reconciling and building a unified dataset for model training [00:24:22], [00:24:25].
- Model Export Exporting the trained model for use across various locations [00:24:31].
This approach is becoming unsustainable due to:
- Data Sharing Reluctance Institutions like hospital systems often resist sharing data with competitors due to legal justifications related to patient rights and safety [00:23:59], [00:24:05], [00:24:08].
- Time and Effort The process of collecting, harmonizing, and training models on large datasets is time-consuming and requires significant effort [00:25:02], [00:25:05].
- Outdated Results By the time models are trained and deployed, the tactical data they learned from may already be outdated, rendering the results ineffective [00:25:08], [00:25:11]. This was observed in the war in Ukraine, where rapid response to changing tactics is crucial [00:24:43], [00:24:48].
The New Paradigm: Distributed AI
To address these limitations, a new paradigm of distributed artificial intelligence is being developed [00:25:17]. This approach leverages federated learning, a concept that has existed for several years [00:25:43], [00:25:45].
Federated Learning Mechanism
Instead of moving large datasets to a central location, federated learning operates by:
- Local Model Training Models are trained directly on the data where it resides [00:25:34].
- Model Sharing Only “small models” or their updates are transmitted between nodes or to a central aggregator, not the raw data itself [00:00:40], [00:25:39], [00:26:23], [00:26:25].
Advantages
This distributed approach offers several benefits:
- Data Privacy and Security By not transmitting large amounts of sensitive data, it addresses concerns about data sharing, especially important in fields like medicine where patient data rights are paramount [00:24:05], [00:26:23].
- Efficiency It avoids the cumbersome process of downloading and harmonizing vast datasets, making deployment faster and more efficient [00:26:23].
- Adaptability Models can learn locally on arbitrarily constructed formats and databases, meaning database compatibility is no longer an obstacle [00:26:38], [00:26:40], [00:26:43].
- Timeliness It allows for faster responses to changing conditions, as demonstrated by the analogy of military tactics [00:24:48].
Current Status and Future Prospects
Professor Dubrawski’s team, along with collaborators and sponsors, has a “Pro Concept” for this new paradigm [00:26:56]. He is convinced that “cool practical solutions” will emerge in the near future, allowing the overcoming of current limitations in artificial intelligence applications [00:27:00], [00:27:06]. The focus is on developing ethical AI models that can save lives, particularly in the medical sector [00:22:59], [00:23:03], [00:23:06]. This includes applications in intensive care, where AI can process increasing amounts of patient physiological data from sensors and extract operationally valuable information for doctors and nurses [00:27:43], [00:27:50], [00:27:54], [00:28:36].