Abstract
Decentralized strategies are of interest for learning from large-scale data over networks. This paper studies learning over a network of geographically distributed nodes/agents subject to quantization. Each node possesses a private local cost function, collectively contributing to a global cost function, which the proposed methodology aims to minimize. In contrast to many existing literature, the information exchange among nodes is quantized. We adopt a first-order computationally-efficient distributed optimization algorithm (with no extra inner consensus loop) that leverages node-level gradient correction based on local data and network-level gradient aggregation only over nearby nodes. This method only requires balanced networks with no need for stochastic weight design. It can handle log-scale quantized data exchange over possibly time-varying and switching network setups. We analyze convergence over both structured networks (for example, training over data-centers) and ad-hoc multi-agent networks (for example, training over dynamic robotic networks). Through analysis and experimental validation, we show that (i) structured networks generally result in a smaller optimality gap, and (ii) logarithmic quantization leads to smaller optimality gap compared to uniform quantization.
Introduction
In decentralized machine learning, data is often spread across geographically distant nodes, such as edge devices or distributed data centers. To train a global model, these nodes must exchange information—typically gradients or model weights. However, real-world communication channels are constrained by limited bandwidth, necessitating quantization to reduce the number of bits used for data representation.
Existing distributed algorithms face two primary hurdles. First, quantization error in standard uniform quantization (rounding to fixed intervals) often leads to a significant "optimality gap." As the algorithm approaches the minimum and values become smaller, uniform quantization loses the precision required to continue improving. Second, many methods demand network rigidity, requiring "weight-stochastic" matrices where incoming and outgoing link weights must sum to one—a condition difficult to maintain in dynamic networks with frequent link failures.
Proposed Methodology
This work introduces a first-order distributed optimization method characterized by two key innovations: log-scale quantization and weight-balanced design.
Log-Scale Quantization
Unlike uniform quantization, log-scale quantization assigns more bits to represent smaller values and fewer bits to larger ones. Mathematically, this acts as a "sector-bound" nonlinearity. This approach ensures high precision near the optimum where gradients are small, which is critical for achieving convergence to a true minimum rather than stalling due to rounding errors.
Weight-Balanced (WB) Design and Gradient Tracking
The algorithm relaxes network requirements by only requiring the network to be "weight-balanced" (the sum of weights on incoming links equals the sum on outgoing links). This makes the system significantly more robust to time-varying topologies. To ensure global convergence despite local, quantized data, the method employs a gradient tracking mechanism. Each node maintains an auxiliary variable to track the average of the global gradients, guiding the local updates toward the global minimum.
Experimental Validation
The researchers validated the proposed framework using two distinct setups:
- Synthetic Benchmark: A 16-agent network minimizing a non-convex cost function with trigonometric terms, testing the ability to handle complex local functions that aggregate into a strongly convex global function.
- MNIST Classification: A standard image classification task using 12,000 images distributed across 16 nodes, employing logistic regression with a convex regularizer.
Key Results
Experimental results demonstrate that log-scale quantization consistently achieves a significantly lower optimality gap compared to uniform quantization across various quantization levels ($\rho$). While uniform quantization causes the optimality gap to plateau relatively early, log-scale quantization allows the algorithm to maintain linear convergence for much longer, reaching significantly higher precision.
Furthermore, the study highlights that structured networks (e.g., exponential graphs) yield faster convergence and smaller optimality gaps than ad-hoc networks (e.g., random geometric graphs). The method also proved robust to "switching" networks, where the topology changes every few iterations.
Conclusion
By prioritizing precision where it matters most—near the optimum—log-scale quantization offers a superior alternative to uniform methods in bandwidth-constrained distributed learning. The proposed weight-balanced first-order method matches the performance of state-of-the-art non-quantized algorithms while significantly reducing communication overhead. Future work will explore "union-connectivity" assumptions and the application of this framework to distributed reinforcement learning and optimal control.
Summary updated