Table 1 Descriptions of new architectural elements of the RA

Processing

Processing applied to in situ data

Inference

Making predictions by applying the trained model [53]

Model compression

Compression applied in model training [53] (e.g. compression of gradients [8] in neural network [53] training)

Modeling buffer

Short-term storage of modeling data

Model synchronization

Synchronization between modeling processes for updating of model’s parameters [53]. For example, a parameter server [53] may be used for synchronization in a distributed network

Serving

Serving functionality for interfacing and visualisation (e.g. a server or a proxy)

Model experimentation

Execution of experiments with a model [53]

Model packaging

Packaging model(s) into executable/loadable file(s)

Model loading

Loading a model into memory for inference

Model distribution

Transfer and deployment of model(s) into node(s)

Model distribution scheduling

Scheduling of model distribution