Limitations in odour recognition and generalization in a neuromorphic olfactory circuit

Neuromorphic computing is one of the few current approaches that have the potential to substantially reduce power consumption in machine learning and artificial intelligence, and has drawn vast inspiration from considerations of biological systems and circuits. In their work, Imam and Cleland1 presented a neuromorphic odour-learning algorithm that is inspired by mammalian olfactory bulb circuitry, which they assessed by considering its performance in ‘rapid online learning and identification’ and ‘broad generalization beyond experience’ of gaseous odourants and odourless gases (in short, ‘gases’) using a set of gas-sensor recordings of different odour presentations and corrupting them by impulse noise. We replicated parts of the study and discovered limitations thereof, which are (1) that the dataset used suffers from sensor drift and a non-randomized measurement protocol that render it of limited use for odour identification benchmarks, and (2) that the model is restricted in its ability to generalize over repeated presentations of the same gas. Therefore, a validation of the model that goes beyond restoring a previously learned data sample remains to be shown, in particular its coherence with the attributed capabilities of robustness and broad generalization beyond experience, as well as its suitability to realistic odour identification tasks.