Machine Learning
Also known as: ML
Definition
Machine Learning is a method within artificial intelligence in which algorithms are trained on historical or streaming data to identify patterns and make predictions or classifications on new data, with performance improving as more relevant data is processed, without requiring manual reprogramming for each new scenario.
Key Points
- ML is a subset of artificial intelligence, not a synonym for it — AI is the broader goal, ML is the dominant method used to achieve it today.
- ML models require two distinct network-relevant phases: training (data-intensive, often batch, tolerant of latency) and inference (may be latency-sensitive depending on the application).
- Training data volume and location determine architecture: centralized cloud training needs high-capacity backhaul; on-site or federated training reduces data-transfer burden but requires distributed compute.
- Model drift — degradation in accuracy as real-world conditions diverge from training data — creates an ongoing connectivity dependency for periodic retraining and redeployment.
- In industrial settings (mining, energy, maritime), ML models are frequently trained centrally on aggregated historical data but deployed at the edge for real-time inference, decoupling the two network demands.
Concept
A machine learning pipeline has two operationally distinct stages. Training ingests historical data — sensor logs, telemetry archives, image sets — to build a statistical model, and is typically compute-intensive and tolerant of network latency since it runs on batched, non-time-critical data. Inference applies the trained model to new, live data to produce a prediction or decision, and its latency tolerance depends entirely on the use case: a quarterly demand-forecasting model can tolerate minutes of delay, while a real-time anomaly-detection model screening sensor streams for a safety trip may need sub-second inference. This split allows network architects to size connectivity independently for each phase — heavy, infrequent uplink for training data, and localized, possibly connectivity-independent, compute for inference.
Explainer
For ConnectedEarth's cross-industry audience, machine learning's operational footprint is best understood through its data-movement pattern rather than its algorithmic detail. An energy utility running ML-based fault prediction across a distributed grid needs to move sensor telemetry from thousands of remote points to a central training environment — a sustained, aggregate bandwidth requirement best served by fixed or cellular backhaul rather than expensive satellite capacity. A maritime fleet running ML-based engine anomaly detection, by contrast, cannot depend on continuous connectivity at sea, so the trained model is pushed to onboard edge compute and only summary results or flagged anomalies are transmitted back over VSAT — inverting the data flow from continuous-upload to occasional-alert. Understanding which phase (training vs. inference) is running where, and what that implies for uplink, downlink, and retraining cadence, is the practical lens ConnectedEarth applies to any ML deployment story.