standard ml
On standard digital time-series data, the input signal can be chunked as “windows” which helps give the model the most complete information possible at a given time point from the recent past, in order to achieve an accurate prediction.
Usually, ML engineers do not spend a lot of time on “feature engineering” i.e. trying to identify the best features to characterize the target data. It is easier to employ a large neural network, feed in data, tune it, and let the black box gods identify the target.
ML is ubiquitous and applications employing standard ML models can be deployed on any hardware that is capable of providing digital data for the model to operate on.
analog ml
In the analog domain, the input data stream is operated on directly by signal processing blocks without keeping a running memory of the input. Hence, knowing the target signal well is crucial.
For making an AnalogML model, it is crucial to spend time on feature engineering to be able to identify desirable features in the data that can identify the target well even before employing a neural network.
20X power-saving AnalogML models are deployed on special hardware engineered by Aspinity’s RAMP technology which is capable of performing ML directly on analog input without converting it to digital data.
