Detail

Learning Compatible Representations

Representation learning is a fundamental aspect of deep learning and underpins core tasks such as image retrieval. Early research in these areas has largely considered models that do not change over time. Less attention has been given to scenarios where the availability of new training data prompts model updates to improve performance, which can lead to completely different representations. In these cases, recomputing the feature vectors for all images in the gallery set becomes essential, but it can be prohibitively expensive for galleries containing vast amount of data, or even infeasible if the original data was no more available due to privacy concerns or storage restrictions. As this issue has gained more attention, learning compatible representations has become a central focus, i.e. learning a new representation model without needing to recalculate gallery features using the updated model. Our talk will explore the foundational challenges of compatible representation learning and the key role of representation stationarity in this process, along with novel training techniques for developing compatible feature representations through stationarity.