Neural Network Representation

Description

Neural Network Representation (NNR) is a Rel-19 concept specified in 3GPP TR 26.847 and TR 26.927 that defines a standardized internal representation for neural network models within 3GPP network functions. Unlike NNEF, which is designed for exchange between different systems, NNR focuses on how a neural network is represented and manipulated inside a single network entity, such as a core network function, RAN node, or UE. NNR provides a common data model that supports model training, inference, compression, and lifecycle management.

The architecture of NNR is based on a layered data model. At the top level, a neural network is represented as a computational graph, where nodes correspond to operations and edges correspond to tensors. Each operation has a type, attributes, and input/output tensors. Tensors are described by their shape, data type, and memory layout. The model also includes metadata such as model identifier, version, training status, and performance metrics. This layered approach allows NNR to be both expressive and efficient, supporting a wide range of neural network architectures from simple feedforward networks to complex transformers and graph neural networks.

Key components of NNR include the graph schema, the parameter store, and the execution context. The graph schema defines the structure of the computational graph, including operation types, connectivity, and attributes. The parameter store holds the weights, biases, and other learnable parameters, often in a compressed or quantized form. The execution context defines how the model is executed, including device capabilities, memory constraints, and parallelism settings. NNR also supports dynamic graphs, allowing the model structure to change at runtime based on input data or network conditions.

In the 3GPP context, NNR plays a central role in enabling AI/ML-based features. For example, a RAN node may use NNR to represent a beam management model that is trained locally and executed in real time. A core network function may use NNR to represent a network slicing optimization model that is trained centrally and distributed to edge nodes. A UE may use NNR to represent an on-device inference model for media processing or positioning. By providing a common internal representation, NNR simplifies the integration of AI/ML into 3GPP systems and enables consistent behavior across different vendors and deployments.

NNR is closely related to NNEF and NNC. NNEF defines the exchange format used to move models between systems, while NNC defines the compression and bitstream syntax used to transport models efficiently. NNR defines the internal representation used within a system. Together, these three concepts form a complete framework for representing, exchanging, and coding neural networks in 3GPP networks. The specifications TR 26.847 and TR 26.927 provide the technical details, including data model definitions, operation semantics, and performance considerations.

Purpose & Motivation

NNR was created to provide a standardized internal representation for neural network models within 3GPP network functions. Before Rel-19, each vendor or framework used its own internal representation, which led to fragmentation, integration challenges, and inefficiencies in model training, inference, and lifecycle management. NNR solves this by defining a common data model that any 3GPP entity can adopt, enabling consistent processing and interoperability.

The motivation for NNR comes from the need to manage AI/ML models throughout their lifecycle within 3GPP networks. Models are trained, validated, compressed, distributed, executed, and updated. Each of these stages requires a clear, unambiguous representation of the model. Without a standardized internal representation, each stage would require custom adapters and conversions, increasing complexity and error rates. NNR provides a single, consistent representation that supports all stages, reducing integration effort and improving reliability.

Historically, previous approaches to internal model representation were either framework-specific or not optimized for telecom networks. For example, TensorFlow uses a graph-based representation, while PyTorch uses a dynamic computation graph. These representations are powerful but not standardized across vendors and not designed for the constraints of 3GPP networks, such as limited memory, real-time execution, and integration with network management. NNR builds on existing industry practices but adds 3GPP-specific requirements and constraints, ensuring that models can be efficiently processed and executed within 5G-Advanced networks.