InfoCom: Kilobyte-Scale Communication-Efficient Collaborative Perception with Information Bottleneck¶
🕒 Published (v1): 2025-12-11 05:51 UTC · Source: Arxiv · Venue: AAAI · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
TL;DR¶
InfoCom is a collaborative perception framework for autonomous driving that reduces inter-agent communication from megabytes to kilobytes (~7.9 KB) by applying an extended Information Bottleneck principle to extract minimal sufficient task-critical representations, rather than compressing spatial feature maps. It achieves 440× and 90× bandwidth reduction over Where2comm and ERMVP respectively while maintaining near-lossless 3D object detection accuracy across three datasets.
Problem¶
Existing communication-efficient collaborative perception methods assume MB-scale bandwidth (3.4 MB–34 MB per agent per collaboration cycle), which is unreliable under real 5G vehicular conditions (rates can drop below 0.4 MB/s). Furthermore, prior feature-selection and feature-compression approaches lack theoretical guarantees on the communication-performance trade-off and operate on high-dimensional BEV feature redundancy rather than purifying task-critical information directly.
Method¶
InfoCom replaces the communication layer of existing collaborative perception backbones (plug-and-play) with three co-designed modules:
-
Information-Aware Encoding (IAE): Extends the IB Markov chain from
Y→X→ZtoY→X→Z→(E,M), mapping intermediate BEV featuresZ ∈ R^{C×H×W}into a low-dimensional stochastic embeddingE ∈ R^D(D ≪ C×H×W) via a variational encoder outputting (μ, σ), with KL regularizationKL(p(E|Z) ‖ N(0,I))enforcing the IB objective. -
Sparse Mask Generation (SMG): Generates a spatial importance mask
M ∈ R^{H×W}via multi-scale convolutions (3×3, 5×5, 7×7), then applies top-k filtering (α=0.1, retaining 10% of positions) and 4-bit uniform quantization. Gradients through the non-differentiable operations use straight-through estimation. -
Multi-Scale Decoding (MSD): At the receiver, expands
Eto a lower-resolution feature map via FC+transposed conv, applies mask-guided modulation (element-wise product), then progressively upsamples via cascaded decoder blocks with multi-scale mask guidance until recovering the fullC×H×Wresolution. The resulting decoded features are fused via the unchanged backbone fusion network.
Training loss: L = L_detect + β·KL(p(E|Z) ‖ N(0,I)).
Key Contributions¶
- First theoretically grounded communication-efficient collaborative perception framework, with formal proofs of bandwidth reduction, task-information preservation, and noise suppression via Lemma 1 (noise suppression bound) and Proposition 1.
- Information purification paradigm operating in representation space rather than spatial feature space, eliminating explicit feature-level alignment constraints.
- Kilobyte-scale transmission (~7.9 KB on OPV2V/V2XSet, ~5.9 KB on DAIR-V2X) comparable to Late Collaboration overhead, while delivering intermediate collaboration accuracy.
- Plug-and-play design: only the communication layer is replaced; validated across CoAlign, AttFuse, and MKD-Cooper backbones.
Results¶
- OPV2V (7.875 KB): AP@50 = 0.9650 vs. Where2comm 0.9463 (3.44 MB) and ERMVP 0.9557 (0.74 MB); matches Standard Collaboration (0.9653 at 34.4 MB).
- V2XSet (7.875 KB): AP@50 = 0.9273 vs. Where2comm 0.8604; ERMVP OOM. Exceeds Standard Collaboration AP@30 (0.9360 vs. 0.9317).
- DAIR-V2X (5.922 KB): AP@50 = 0.7789 vs. Where2comm 0.7539 and ERMVP 0.7791; near-matches Standard Collaboration (0.7843).
- Communication volume: 440× reduction vs. Where2comm, 90× vs. ERMVP, 4000× vs. Standard Collaboration.
- Efficiency metric: 1.8×10⁻² average AP gain/KB vs. 3.2×10⁻⁵ (Where2comm) and 1.7×10⁻⁴ (ERMVP).
- On weaker backbones (MKD-Cooper): InfoCom improves mean AP by +1.27% while reducing bandwidth from 34.3 MB to 2.7 KB.
- Ablation: removing SMG drops mean AP by 1.39%; removing mask from MSD drops 6.79%; removing IAE drops 1.98%.
Limitations¶
- Computation time is approximately 2× that of Where2comm (measured on RTX 3090); total system latency gains depend on network conditions offsetting this overhead.
- Retaining only 10% of spatial cues (α=0.1) and 4-bit quantization are empirically chosen; performance at very high IoU thresholds (AP@70) still slightly lags Standard Collaboration (e.g., OPV2V: 0.9202 vs. 0.9229).
- Evaluated on vehicle-centric and V2X scenarios only; generalization to other multi-agent settings (e.g., drone swarms, robot teams) is unvalidated.
- The IB regularization weight β requires tuning per deployment.
Relevance to Agentic AI / LLM Agents¶
Though grounded in autonomous driving, InfoCom directly addresses the bandwidth-constrained multi-agent communication problem that is equally central to agentic AI systems where agents must exchange intermediate representations under network or token-budget constraints. The Information Bottleneck formulation provides a principled framework for identifying what information is minimally sufficient for a shared task — a general design principle applicable to any multi-agent system that needs to compress inter-agent messages. The plug-and-play modularity (replace only the communication layer) mirrors how agentic architectures modularize perception, planning, and communication components. The theoretical treatment of noise suppression and task-relevance preservation is directly transferable to settings where LLM agents must selectively share context without overwhelming each other's inputs.