Interpretable Data Fusion for Distributed Learning: A Representative Approach via Gradient Matching

Mengchen Fan, Baocheng Geng, Keren Li, Xueqian Wang, Pramod K. Varshney

In the world of distributed machine learning, Federated Learning (FL) is the reigning champion of privacy. By sharing model weight updates instead of raw data, it allows multiple parties to train a global model without ever exposing sensitive information. However, FL has a major "black box" problem: a vector of weight updates is gibberish to a human operator. If a client’s data is biased, corrupted, or malicious, there is no intuitive way to see it.

A new paper, "Interpretable Data Fusion for Distributed Learning: A Representative Approach via Gradient Matching," proposes a shift in how we think about distributed updates. Instead of sending abstract math, what if clients sent a single, synthesized "representative" data point that summarizes their entire local dataset?

TL;DR

The authors introduce a distributed learning framework that replaces black-box weight updates with human-interpretable "representative" data points. By using gradient matching to synthesize these points, the method achieves accuracy and convergence rates that meet or exceed standard Federated Learning (FedAvg) while allowing humans to visually inspect what the model is learning.

What Problem Does This Paper Address?

Traditional distributed learning focuses almost entirely on accuracy and privacy, often neglecting human-machine collaboration. In critical environments—like healthcare or defense—exclusive reliance on automated systems is risky.

The core challenge is that current distributed methods (like FedAvg) are not interpretable. When a client sends a model update, a human supervisor cannot look at those numbers and understand the underlying data distribution. This makes it nearly impossible to detect outliers or provide human oversight during the training process. The authors aim to bridge this gap by creating a "digestible" format for data fusion that maintains privacy but remains comprehensible to people.

Methods

The researchers move away from sharing weights and instead focus on Representative Data Points.

  1. Gradient Matching: The goal is to create a virtual data point $(x^r, y^r)$ that, when fed into the model, produces a gradient nearly identical to the average gradient of the entire raw local batch.
  2. Optimization with Perturbation: To keep the representative recognizable, they start with the mean of the local data and add a small, constrained perturbation ($\delta$). They solve an optimization problem to find the $\delta$ that minimizes the "representative loss"—the difference between the synthetic point's gradient and the true batch gradient.
  3. Residual Error Correction: Recognizing that one point can't perfectly mimic a whole batch, they introduce a residual term ($\tau$). This term tracks the gradient discrepancy from previous rounds and injects it into the current optimization, ensuring the model doesn't drift away from the "ground truth" of the raw data.
  4. Distributed Fusion: In the distributed setting, clients compute these representatives locally and send them (along with the batch size) to a central server, which then updates the global model.

Datasets

The approach was evaluated on two distinct types of data:

  • Image Segmentation Dataset (UCI): A compact multivariate dataset with numeric attributes.
  • MNIST: The classic benchmark of 60,000 handwritten digits.
  • Non-IID Setup: To simulate real-world challenges, the authors used a "non-Independent and Identically Distributed" strategy, where each client only held images of a single label (e.g., Client A only has '0's, Client B only has '1's).

Key Insights and Ideas

  • The "Aha" Moment (Visual Interpretability): The most striking finding is that the synthesized representatives actually look like the data they represent. An MNIST representative for the digit '0' looks like a slightly noisy '0'. This allows a human to "see" the data fusion process in real-time.
  • Regularization via Synthesis: Interestingly, the process of condensing data into a representative acts as a form of regularization. This leads to much smoother convergence curves compared to FedAvg, which often suffers from high volatility due to simple gradient averaging.
  • Complexity vs. Fidelity: As models get more complex (moving from a simple MLP to a CNN), the representative requires more perturbations to match the gradients. This creates a trade-off: more complex models are more accurate but produce "noisier" representatives that are slightly harder for humans to interpret.

Key Results

  • Outperforming FedAvg: On the MNIST dataset using a CNN with 4 clients, the Representative method achieved 97.59% accuracy, surpassing FedAvg’s 95.97%.
  • Superior Convergence: In head-to-head tests, the proposed method converged faster and with significantly fewer fluctuations than FedAvg (see Figure 2 in the paper).
  • The Necessity of Residuals: The inclusion of the residual error term ($\tau$) was shown to be critical. Without it, the model's accuracy plateaued much earlier and at a lower level.

Limitations and Future Work

While promising, the paper notes a few areas for further exploration:

  • Computational Overhead: Generating a representative requires local iterations ($H$) to solve the optimization problem. The authors suggest a future study on the trade-off between this local computation and the resulting communication efficiency.
  • Scaling Human Oversight: While the representatives are interpretable, the paper doesn't yet define a formal protocol for how a human should intervene if they spot a "bad" representative.
  • Broader Architectures: The current study focuses on MLPs and CNNs. Testing this on Transformers or Large Language Models (LLMs) remains an open challenge.

By turning abstract gradients into tangible "representatives," this research offers a path toward distributed AI that is not just powerful, but also transparent enough for humans to trust.

Summary updated

Abstract

This paper introduces a representative-based approach for distributed learning that transforms multiple raw data points into a virtual representation. Unlike traditional distributed learning methods such as Federated Learning, which do not offer human interpretability, our method makes complex machine learning processes accessible and comprehensible. It achieves this by condensing extensive datasets into digestible formats, thus fostering intuitive human-machine interactions. Additionally, this approach maintains privacy and communication efficiency, and it matches the training performance of models using raw data. Simulation results show that our approach is competitive with or outperforms traditional Federated Learning in accuracy and convergence, especially in scenarios with complex models and a higher number of clients. This framework marks a step forward in integrating human intuition with machine intelligence, which potentially enhances human-machine learning interfaces and collaborative efforts.
BibTeX
@inproceedings{fan2024interpretable,
  author = {Mengchen Fan and Baocheng Geng and Keren Li and Xueqian Wang and Pramod K. Varshney},
  title = {{Interpretable Data Fusion for Distributed Learning: A Representative Approach via Gradient Matching}},
  booktitle = {FUSION},
  year = {2024},
  doi = {10.23919/fusion59988.2024.10706324},
}

Loading related content...

Figure