US 12,423,960 B1
Concept shift detection and correction using probabilistic models and learned feature representations
Lukas Stefan Balles, Berlin (DE); Giovanni Zappella, Berlin (DE); and Cedric Philippe Archambeau, Berlin (DE)
Assigned to Amazon Technologies, Inc., Seattle, WA (US)
Filed by Amazon Technologies, Inc., Seattle, WA (US)
Filed on Mar. 29, 2022, as Appl. No. 17/707,004.
Int. Cl. G06V 10/77 (2022.01); G06F 11/07 (2006.01); G06N 20/20 (2019.01); G06V 10/772 (2022.01); G06V 10/774 (2022.01)
CPC G06V 10/7747 (2022.01) [G06F 11/0769 (2013.01); G06N 20/20 (2019.01); G06V 10/772 (2022.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method comprising:
retraining a primary machine learning (ML) model based at least in part on use of a first batch of data elements in a sliding window of a memory;
obtaining a second batch of data elements to be used as part of retraining the primary ML model, wherein each of the second batch of data elements is associated with a corresponding target value;
training a gaussian process (GP) ML model based at least in part on use of representations of the first batch of data elements generated at least in part via use of the primary ML model;
generating predictions for the second batch of data elements, via use of the GP model, based on use of representations of the second batch of data elements generated at least in part via use of the primary ML model;
determining a likelihood of concept shift evidenced within the second batch of data elements based at least in part on an analysis of the predictions and target values associated with the second batch of data elements;
determining that the likelihood of concept shift satisfies a criteria; and
transmitting a message identifying the likelihood of concept shift.