receiving, via a data interface, a training dataset of image samples in a source domain;

receiving, at a first training iteration of the one or more neural network prediction models, a first target image query from a streaming sequence of target image queries in a target domain;

generating, by a first neural network prediction model implemented on one or more hardware processors, a first prediction relating to at least one item in the first target image query;

generating, by a second neural network prediction model implemented on one or more hardware processors, a second prediction relating to the at least one item in the first target image query;

generating, by the first neural network prediction model and the second neural network prediction model, a first output and a second output each identifying a respective probability that the at least one item is in an input image sample from the training dataset, respectively;

in response to determining that the second output is greater than a pre-defined threshold, selecting the second prediction as a first pseudo label;

training at least the first neural network prediction model based on a weighted sum of (i) a first loss objective computed as a cross-entropy between the first output and a corresponding label from the training dataset, and (ii) a second loss objective computed as a cross-entropy between the first prediction and the selected second prediction as the first pseudo label;

erasing, from a memory, the first target query at an end of the first training iteration of the one or more neural network prediction models; and

performing, using the trained first neural network prediction model, image recognition of a testing image.