Oort

This component quantifies the potential contribution of a client's data to improving the global model. It primarily uses training loss as a proxy for data utility. Clients with higher local loss are assumed to possess data that is more informative or challenging for the current model, making their updates more valuable for convergence.

• Core Metric: Local training loss on the client's dataset.
• Dynamic Scoring: Utility scores are updated after each participation round.
• Objective: Prioritizes clients whose data will most effectively reduce the global loss function.

This module monitors and predicts the resource consumption of client participation. It profiles clients based on real-time and historical metrics to avoid stragglers that delay round completion.

• Profiled Metrics: Includes computation time (CPU/GPU capability), communication latency (network bandwidth), and device availability (battery, online status).
• Predictive Modeling: Uses historical data to forecast a client's completion time for a given training task.
• Objective: Selects clients that can complete training within a target wall-clock time, improving overall system throughput.

The core decision engine that selects the participant cohort for each training round. It formulates client selection as an online optimization problem, balancing the scores from the Utility Estimator and Efficiency Profiler.

• Utility-Efficiency Trade-off: Employs a configurable parameter (α) to weight the importance of statistical utility versus system efficiency.
• Online Learning: Adapts selection strategy based on feedback from previous rounds (e.g., actual completion times, global model improvement).
• Output: A ranked list or probabilistic distribution for selecting the optimal set of clients per round.

A persistent, updatable repository that maintains profiles for all clients in the federation. This is the source of truth for the other components.

• Stored Data: Historical loss values, resource profiles (compute speed, bandwidth), participation history, and data distribution signatures (if available).
• Update Mechanism: Profiles are refreshed after each client interaction (e.g., after a client submits an update, its completion time and new loss are recorded).
• Role: Enables informed selection without requiring real-time interrogation of all clients, which is infeasible at scale.

A mechanism to prevent the selection policy from becoming myopic. It ensures the system occasionally selects sub-optimal clients (in terms of current scores) to gather new information and adapt to changing conditions.

• Exploration: Selects clients with uncertain profiles or those not recently participated to update their metadata.
• Exploitation: Primarily selects clients with high utility-efficiency scores.
• Balance: Often implemented using techniques like ε-greedy or upper confidence bound (UCB) algorithms from multi-armed bandit theory.

An optional but common enhancement where clients are grouped into tiers based on their performance profiles (e.g., fast, medium, slow). Oort can incorporate this to enforce fairness and maintain diversity.

• Tier Formation: Clients are clustered by their historical training completion times.
• Stratified Selection: The scheduler selects a minimum number of participants from each tier in every round.
• Benefit: Prevents model bias towards data only from the fastest devices and improves robustness by ensuring updates from a diverse set of system capabilities.

Oort accelerates the training of diagnostic models across hospitals without sharing sensitive patient data. It selects clients (hospitals) based on both the statistical utility of their local loss (ensuring diverse medical cases are represented) and system efficiency (prioritizing institutions with available GPU clusters and stable bandwidth). This reduces the time to converge on a robust, globally accurate model for tasks like detecting pathologies in medical imaging, while respecting strict HIPAA and GDPR compliance through federated learning's inherent privacy.

For next-word prediction models trained on user devices, Oort manages extreme heterogeneity. It profiles devices for resource efficiency—selecting those with sufficient battery, idle state, and strong Wi-Fi—while also ensuring statistical utility by prioritizing users whose typing data shows high loss (indicating the current model performs poorly for their dialect or slang). This balances model personalization with battery preservation, improving user experience without draining device resources.

Oort coordinates model updates from a global fleet of vehicles. It selects vehicles for a training round by optimizing for:

• Statistical Utility: Vehicles in novel geographic or weather conditions (high loss) provide valuable edge-case data.
• System Efficiency: Vehicles parked and charging, with high-bandwidth cellular connections, are prioritized to upload large model updates. This enables the continuous, privacy-preserving improvement of perception and navigation models based on real-world driving data, accelerating the adaptation to new environments.

In factories with thousands of sensors on manufacturing equipment, Oort selects which sensor gateways participate in training a failure prediction model. It uses client profiling to identify gateways with:

• High utility (sensors on machinery showing anomalous telemetry).
• High efficiency (ample compute, connected via wired Ethernet, not during peak production). This ensures the global model quickly learns from emerging failure patterns while avoiding disruption to critical operational networks, minimizing downtime and maintenance costs.

Banks collaborate to train a fraud detection model without exposing transaction details. Oort's selection is crucial for regulatory compliance and model efficacy. It selects bank clients offering high statistical utility (those with recently detected novel fraud patterns) while enforcing system efficiency by preferring data centers with dedicated secure line availability over slower, less secure connections. This joint optimization allows for rapid adaptation to new fraud tactics while maintaining the stringent security and latency requirements of financial networks.

For enterprises training a model across different business units or geographic divisions (silos), Oort acts as an intelligent orchestrator. It selects silos by evaluating:

• Data Utility: Divisions with non-IID data distributions that challenge the current global model.
• Resource Profiles: Silo IT infrastructure capacity and scheduled maintenance windows. This enables efficient knowledge consolidation from disparate parts of the organization—such as training a unified sales forecast model—while respecting internal data governance policies and optimizing for available computational budgets.

In federated learning, statistical utility quantifies the expected contribution of a client's local data to improving the global model's accuracy. It is a core optimization target in client selection.

• Primary Metric: Often measured by the training loss on a client's local dataset. A higher loss indicates the model performs poorly on that client's data, suggesting its inclusion would provide greater learning signal.
• Contrast with System Utility: While system utility focuses on speed and resource efficiency, statistical utility focuses on model quality and convergence rate.
• Oort's Application: Oort uses training loss as its key proxy for statistical utility, actively selecting clients whose data will most reduce the global loss.

System efficiency refers to the computational and temporal costs of executing a federated learning round, focusing on the resources of the participating edge devices.

• Key Factors: Includes client availability, computation speed (CPU/GPU), network bandwidth, battery level, and memory constraints.
• Objective: To minimize round completion time and prevent stragglers (slow devices) from bottlenecking the entire training process.
• Oort's Application: Oort profiles clients for these resource metrics and factors them into its selection to ensure timely round completion without sacrificing model quality.

FedCS is a foundational client selection framework designed to handle resource heterogeneity. It prioritizes system efficiency by selecting clients that can complete training within a deadline.

• Core Mechanism: Clients report their resource profiles (compute, bandwidth). The server selects the maximum number of clients estimated to finish local training and upload updates by a set deadline.
• Primary Focus: Straggler mitigation and reliable round completion. It is less focused on the statistical quality of the selected cohort compared to Oort.
• Contrast with Oort: While FedCS is primarily resource-aware, Oort is a multi-objective framework that jointly optimizes for both resources (like FedCS) and statistical utility.

Power-of-choice is a simple but effective heuristic for client selection that balances randomness with utility-based picking.

• Algorithm: In each round, the server randomly samples a small pool of d clients. From this pool, it selects the k clients with the highest utility (e.g., largest local dataset size or highest loss).
• Advantage: Provides a probabilistic guarantee of better statistical utility than pure random selection, with minimal overhead.
• Relation to Oort: Can be seen as a simplified, single-metric (utility) version of selection. Oort generalizes this by using a multi-armed bandit to learn a more complex, dual-objective utility function over time, rather than a simple per-round ranking.

Client profiling is the systematic collection and maintenance of metadata about federated learning clients. This data is essential for informed selection decisions.

• Profiled Data: Includes hardware specifications (CPU, RAM), network conditions (bandwidth, latency), power status, data statistics (dataset size, distribution), and historical behavior (reliability, dropout rate).
• Role in Selection: Profiling data feeds the utility functions of frameworks like Oort and FedCS. Accurate profiles allow the server to predict training duration and contribution value.
• Oort's Implementation: Oort maintains a lightweight, continuously updated profile for each client, tracking both resource metrics (for system efficiency) and training loss (for statistical utility).

In client selection, a Multi-Armed Bandit (MAB) is an online learning framework used to sequentially choose clients to maximize a reward, balancing exploration and exploitation.

• Arms = Clients: Each client is an "arm" that can be pulled (selected).
• Reward: A composite function reflecting the benefit of selecting that client (e.g., model improvement per unit time).
• Exploration vs. Exploitation: The algorithm must explore new or under-sampled clients to learn their value, while exploiting known high-performing clients to maximize immediate reward.
• Oort's Core Algorithm: Oort formulates client selection as a contextual bandit problem. The "context" is the client's profile (resources, loss). It learns a policy that maps this context to a selection probability, dynamically optimizing for its dual objectives.