Recent advances in artificial intelligence have allowed algorithms to perform classification tasks with ever-improving performance. However, most clustering algorithms are unable to extract key characteristics from their clusters on-the-fly. The literature has described some techniques to detect differences between groups, but they tend to be post-processing steps and to rely on probabilistic models. In this report, we introduce an online algorithm capable of one-shot learning and vector-based classification of both objects and their differences, the latter we call contrasts. We then describe a few clustering and contrast experiments on test datasets to show the potential of this contrast learning algorithm. We expect this exploratory study will enable further research to use the concept of contrasts for more elaborate tasks related to cognitive processes, such as producing relevant descriptions and zero-shot learning.
Launch the jupyter notebook third_draft.ipynb
and run all cells in order.
For more information, read the following report.