AI Glossary / Governance, ethics and safety / Advanced

Machine learning underlay network

Definition

A machine learning underlay network is the telecommunications network, together with its network functions, that interacts with an ML overlay. It is the network-side environment from which relevant observations can be obtained and to which ML-supported results may be applied.

What it means

The overlay organizes ML capabilities and their integration; the underlay is the telecommunications network those capabilities support. This distinction helps describe responsibilities without assuming that every network function contains its own independently designed learning system.

Underlay does not mean only cables, radio hardware or the lowest protocol layer. In this ITU context it includes related network functions, which may have software implementations. It should also not be confused with the use of underlay in generic packet-overlay discussions without checking the architectural meaning.

How it works

Network functions expose relevant information through defined interfaces and data-handling arrangements. An ML application uses suitable observations to produce forecasts, classifications or other outputs. Network-side functions may then consume those outputs within the operator's management and control processes.

The integration must respect the underlay's capabilities and constraints. Available measurements, update frequency, compute resources, communication delays and permitted actions can all affect whether an ML use case is feasible.

Keep the network's actual state distinct from a model's estimate of it. If measurements are stale or incomplete, the ML overlay may form an inaccurate picture. Changes in network configuration or behavior can also change the meaning of previously reliable inputs, making coordination and monitoring necessary.

Three practical examples

• Traffic forecasting: Network counters provide observations for a model that estimates future demand. The counters' definitions and collection intervals are part of the integration.
• Fault support: A model analyses network events and returns classifications to an operational support function, which handles them according to its workflow.
• Resource management: An ML recommendation is evaluated against the underlay's available capacity and permitted configuration changes before any action is taken.

Think of it this way

Think of a transport system receiving advice from a planning service. The roads, vehicles and operating processes are the environment being supported; the advisory service is organized separately. The analogy is limited because the underlay includes precise telecommunications functions and interfaces, not simply physical infrastructure.

Worked example

Imagine a fictional network use case where an ML overlay expects measurements every second and returns a resource recommendation valid for five seconds. The relevant underlay function currently exports measurements only once per minute.

The model may be technically deployable, but the available observations do not match its assumed timing. Sending the same minute-old value repeatedly would not create fresh one-second measurements.

The team must revise the data provision, choose a model suited to the available cadence or change the use case. This example shows that underlay capabilities constrain application feasibility. It does not prescribe a standard sampling interval or guarantee that faster collection is always worthwhile.

Illustrative case study

Fictional case: CoreBridge Lab connects an ML overlay to a simulated telecommunications underlay. A network-function upgrade changes a utilization counter from an interval average to a peak value while keeping a similar field name.

Forecast quality changes because the model now receives a different measurement meaning. The team versions the data contract and evaluates the revised integration rather than blaming the model alone. This invented case illustrates coordinated change management, not a real network incident.

Common misconceptions and mistakes

An ML underlay network is not the neural network inside a model. It is the telecommunications environment that interfaces with ML capabilities. Nor is it necessarily a passive source of training data: it may also be the recipient of operational outputs.

Do not assume the overlay can command every function simply because connectivity exists. Permissions and operational constraints remain separate. A simulation of an underlay is also distinct from the actual network and may omit important behavior. Finally, logical separation does not eliminate dependencies on timing, semantics or resource availability.

Business use

For operators, the concept helps organize integration between ML specialists and network engineering teams. It makes clear that successful deployment depends on what the network can observe, expose and act upon, not only on model accuracy.

Document the relevant functions, data contracts and action boundaries for each use case. Include network changes in model-impact reviews and model changes in operational reviews where appropriate. Evaluate the complete loop, including delays and fallback behavior, before making broad claims about autonomous network performance.

People also ask

Does underlay mean only physical hardware?

No. In this context it includes the telecommunications network and its related network functions.

Is it the neural network used by the ML model?

No. The terms refer to different kinds of network.

Can a change in the underlay affect model results?

Yes. Measurement definitions, timing, topology or operating conditions can change the inputs and consequences of predictions.

Check your understanding

An ML model assumes fresh measurements every second, but the network supplies one update per minute. Is interface connectivity alone enough?

Answer: No. The underlay's data cadence does not meet the model's timing assumptions, so the use case needs reassessment.

Related terms

• Machine learning overlay
• Data Contract
• network function
• telemetry
• model serving
• network orchestration
• simulation environment
• operational control

Learn next

For a network ML use case, list the underlay functions that provide data and consume results. Record units, update frequency, permissions and fallback behavior. Identify a network change that would invalidate a model assumption, then design an integration check that would detect it.

More in Governance, ethics and safety

• Machine learning model metadata - Machine learning model metadata is information describing a model and its relevant characteristics.
• Machine learning profile - A machine learning profile, in ITU-T Y.3187, is the interpreted content of an ML intent together with relevant runtime updates.
• Machine learning sandbox - A machine learning sandbox is a controlled environment for training or testing ML models and evaluating their effects.
• Malware - Malware is software or related executable logic deliberately designed to perform harmful or unauthorized activity.
• Measurement Bias - Measurement bias is systematic distortion in how a quantity is recorded or represented.
• Membership inference attack - A membership inference attack attempts to determine whether a particular record was included in a machine-learning model's training data.

Sources

ITU-T Y.3172 (06/2019), definition 3.2.7 and Section 8.3, consulted for the underlay and its source/sink interfaces: https://www.itu.int/rec/T-REC-Y.3172-201906-I/en. ITU-T Y Supplement 97, term 3.1.83, also consulted: https://www.itu.int/rec/T-REC-Y.Sup97-202511-I/en. Timing and counter-definition examples are original engineering illustrations, not mandatory sampling requirements.

Previous: Machine learning system Next: Machine perception