In biological and artificial noses, similar odorants activate substantially overlapping populations of sensors. Increasing the discriminability of such odorants through learning is challenging, as learning-related plasticity in neurons activated by one such odorant then many inappropriately alter the representations of similar, overlapping odorants. We here describe a transfer learning algorithm, based on mammalian olfactory bulb architecture and asymmetric STDP, in which interneurons construct experience-dependent higher-order receptive fields (HORFs) that become diagnostic of more complex odorant signatures (covariance patterns). Consequently, even highly similar odorants can evoke nonoverlapping ensembles of interneurons, and thereby independently regulate principal neuron spike timing patterns and prevent the cross-contamination of plasticity. HORF construction is regulated by parameters such as maximum synaptic weight, initial connection probabilities, and the interneuron membrane time constant and spike threshold. This core algorithm can in principle allocate inhibition so as to regulate generalization and discrimination among specific odorant representations. © 2017 IEEE.