Neural Network Exchange Format

Description

Neural Network Exchange Format (NNEF) is a Rel-19 concept specified in 3GPP TR 26.847 and TR 26.927 that defines a standardized, vendor-neutral format for representing and exchanging neural network models. In the context of 3GPP, NNEF serves as the interchange format that allows AI/ML models to be moved between different training frameworks (e.g., TensorFlow, PyTorch), inference engines, and network functions without loss of information or semantic ambiguity. It is analogous to ONNX in the broader AI community, but tailored to the requirements and constraints of 3GPP networks.

The architecture of NNEF is based on a graph-based representation. A neural network is modeled as a directed acyclic graph (DAG) where nodes represent operations (e.g., convolution, pooling, activation) and edges represent tensors (multi-dimensional arrays) flowing between operations. The format includes a schema for describing the graph topology, operation types, tensor shapes, data types, and parameter values. It also supports metadata such as model version, author, training dataset information, and quantization parameters. This structured representation enables tools to parse, validate, transform, and execute models consistently across different platforms.

Key components of NNEF include the model description language, the binary parameter container, and the operation set. The model description language is a human-readable, text-based format (often using a syntax similar to JSON or YAML) that defines the graph structure and operation attributes. The binary parameter container stores the actual weights and biases in an efficient binary format, often compressed using NNC. The operation set defines a standardized list of neural network operations, ensuring that a model trained in one framework can be correctly interpreted in another. NNEF also supports extensions for custom operations, allowing vendors to add proprietary layers while maintaining backward compatibility.

In the 3GPP context, NNEF plays a critical role in enabling multi-vendor AI/ML interoperability. For example, a model trained by one vendor's core network function can be exported in NNEF and imported by another vendor's RAN node for inference. Similarly, a UE can receive an NNEF model from the network and execute it using its local inference engine. NNEF also facilitates model lifecycle management, including versioning, updates, and rollback. By providing a common exchange format, NNEF reduces integration effort and accelerates the deployment of AI/ML features in 5G-Advanced networks.

NNEF is closely related to NNR (Neural Network Representation) and NNC (Neural Network Coding). While NNR defines the internal representation used within a network function, NNEF defines the external exchange format used between network functions. NNC provides the compression and bitstream syntax for transporting NNEF models efficiently. Together, these three concepts form a complete pipeline: NNR for internal processing, NNEF for interchange, and NNC for compression. The specifications TR 26.847 and TR 26.927 provide the technical details, including use cases, requirements, format syntax, and performance considerations.

Purpose & Motivation

NNEF was created to address the lack of a standardized, vendor-neutral format for exchanging neural network models in 3GPP networks. Before Rel-19, AI/ML models were typically exchanged using proprietary formats or framework-specific serialization (e.g., TensorFlow SavedModel, PyTorch pickle), which led to interoperability issues, vendor lock-in, and integration complexity. NNEF solves this by providing a common format that any vendor or framework can adopt.

The motivation for NNEF comes from the multi-vendor nature of 3GPP networks. In a typical deployment, the core network, RAN, and UE may be supplied by different vendors, each with its own AI/ML toolchain. Without a standardized exchange format, integrating AI/ML features across these domains would require custom adapters and extensive testing. NNEF eliminates this overhead by defining a clear, unambiguous representation of neural networks that can be parsed and executed by any compliant system. It also enables model portability, allowing models to be trained in one environment and deployed in another.

Historically, previous approaches to model exchange were either framework-specific or not optimized for telecom networks. For example, ONNX is widely used in the AI community but does not address 3GPP-specific requirements such as integration with NNC, support for network slicing, or compliance with 3GPP security and management frameworks. NNEF builds on existing industry formats but adds 3GPP-specific extensions and constraints, ensuring that models can be efficiently coded, transmitted, and executed within 5G-Advanced networks.