EPOC: Efficient Perception via Optimal Communication

Masoumeh Heidari Kapourchali, Bonny BanerjeeAAAI 2020 · DOI

Introduction

In multi-agent or multi-sensor systems, the standard "full communication" paradigm—where every agent broadcasts to every other agent—is inherently unscalable. It leads to bandwidth congestion and the propagation of noise from unreliable or uninformative sensors. EPOC (Efficient Perception via Optimal Communication) addresses this by enabling autonomous agents to learn whom to query to resolve environmental uncertainty in an online, unsupervised manner.

The EPOC Framework

EPOC shifts away from traditional reinforcement learning, instead utilizing Predictive Coding (PC). The core objective is the minimization of Variational Free Energy (VFE), a brain-inspired metric representing the "surprise" or discrepancy between an agent's internal model and its actual observations.

The EPOC framework flowchart illustrating the cycle of feature extraction, state estimation, communication selection, and model updating.

In this architecture, communication is treated as an "active perception" action. An agent calculates which peer's data would most effectively reduce its internal VFE. This process is decentralized; each agent performs its own inference and determines its own communication path, making the system robust to central points of failure.

Experimental Setup

The framework was evaluated using two primary benchmarks for human activity recognition, where individual sensors or skeletal joints were treated as independent agents:

  • KARD (Kinect Activity Recognition Dataset): 18 activities performed by 10 individuals, using 15 3-D skeletal joints.
  • UTD-MHAD (Multimodal Human Action Dataset): A complex dataset with 27 activities, including skeletal joints, inertial sensors, and depth video.

Key Insights and Results

EPOC treats communication as a tool for targeted uncertainty reduction. For instance, if a "head" agent is uncertain whether a subject is lunging or bowling, it learns to prioritize data from the wrist or ankle agents—the most discriminative joints for those specific actions.

Skeletal joint agents for walking and waving activities. The varying joint markers illustrate the selective communication policy where specific agents are queried based on their informational value.

Performance Highlights

  • Efficiency through Sparsity: Agents achieved high accuracy while communicating only 61-63% of the time, successfully ignoring redundant or uninformative peers.
  • Natural Robustness: The system adapts to sensor failure automatically. If a peer provides noisy data, the agent's prediction error increases, leading it to dynamically switch to more reliable partners.
  • Lightweight Architecture: EPOC achieved 85.11% accuracy on UTD-MHAD with only 67,000 parameters. This is a massive reduction compared to standard Deep Learning models (like ConvNets) which often require ~60 million parameters, making EPOC ideal for edge devices.

Limitations and Future Work

Currently, the authors utilize linear generative functions to maintain simplicity and differentiability. Future research will explore integrating non-linear functions (such as neural networks) to further boost accuracy. Additionally, the researchers aim to test the scalability of EPOC's selection logic in massive swarms containing hundreds or thousands of potential communicators.

Summary updated

BibTeX
@inproceedings{kapourchali2020epoc,
  author = {Masoumeh Heidari Kapourchali and Bonny Banerjee},
  title = {{EPOC: Efficient Perception via Optimal Communication}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i04.5830},
}
Figure