Feature Compression for Machines

Description

Feature Compression for Machines (FCM) is a novel concept introduced in 3GPP Release 19 as part of the ongoing evolution of 5G and future 6G networks. It addresses the growing need to transport machine learning (ML) features—intermediate representations of data generated by neural networks—between different network entities, such as between a User Equipment (UE) and an edge application server, or between distributed network functions. FCM defines mechanisms to compress these features efficiently, minimizing bandwidth consumption and transmission latency while preserving the information necessary for downstream ML tasks.

In modern AI/ML architectures, it is common to split a neural network across multiple devices or network nodes. For example, a UE might perform initial feature extraction (e.g., from a camera or sensor) and then transmit the extracted features to an edge server for further processing or inference. These features, often represented as multi-dimensional tensors, can be large, especially for high-resolution inputs. Transmitting them uncompressed over the network would consume significant bandwidth and introduce latency, which is unacceptable for real-time applications like autonomous driving, augmented reality, or industrial automation. FCM provides a standardized framework for compressing these features, enabling efficient split computing.

FCM operates by defining a set of compression algorithms and protocols that can be applied to feature data. These algorithms may include quantization, pruning, entropy coding, or learned compression models. The choice of algorithm depends on the characteristics of the features and the requirements of the application. FCM also specifies signaling mechanisms to negotiate compression parameters between the sender and receiver, ensuring interoperability. The compression process is typically lossy, but the goal is to minimize the impact on the accuracy of the downstream ML task. This requires careful design to balance compression ratio and task performance.

Architecturally, FCM can be implemented at various layers of the protocol stack. At the application layer, it may be integrated into ML frameworks and APIs. At the transport layer, it can leverage existing protocols like QUIC or TCP with appropriate extensions. At the network layer, it may involve 5G core functions such as the User Plane Function (UPF) or the Network Exposure Function (NEF) to facilitate feature transport. The 3GPP specifications 26.847 and 26.927 provide the technical details for FCM, including the compression formats, signaling procedures, and performance requirements.

The introduction of FCM in Release 19 reflects the broader trend of integrating AI/ML into 3GPP networks. It enables new use cases such as distributed inference, federated learning, and collaborative perception, where multiple devices share features to improve overall model accuracy. By standardizing feature compression, FCM ensures that these use cases can be deployed across multi-vendor networks and devices, fostering an open ecosystem. As networks evolve towards 6G, FCM is expected to play a critical role in enabling efficient and scalable AI/ML services.

Purpose & Motivation

FCM was created to solve the problem of efficiently transporting machine learning features across 3GPP networks. With the rise of AI/ML applications, there is an increasing need to split neural network computations between devices and edge/cloud servers. However, the intermediate features generated by these networks can be large, and transmitting them uncompressed would strain network bandwidth and increase latency, hindering real-time performance. FCM provides a standardized way to compress these features, reducing the data volume while maintaining the fidelity required for accurate ML inference.

Prior to FCM, there was no standardized approach for feature compression in 3GPP networks. Each vendor or application might use proprietary compression methods, leading to interoperability issues and inefficiencies. FCM addresses this by defining common compression algorithms and signaling protocols, enabling seamless interworking between different network elements and devices. This is essential for widespread adoption of split computing and distributed AI/ML services.

Moreover, FCM is motivated by the need to support emerging use cases such as extended reality (XR), autonomous vehicles, and smart factories, where low latency and high reliability are paramount. In these scenarios, features must be transmitted quickly and reliably, often over wireless links with limited capacity. FCM ensures that the feature data is compressed to fit within the available bandwidth while preserving the essential information for the task. This enables real-time decision-making and enhances the overall user experience. By standardizing FCM, 3GPP aims to facilitate a vibrant ecosystem of AI/ML applications that can leverage the network efficiently.