Signal Conditioning for Learning in the Wild

Abstract

The mammalian olfactory system learns rapidly from very few examples, presented in unpredictable online sequences, and then recognizes these learned odors under conditions of substantial interference without exhibiting catastrophic forgetting. We have developed a brain-mimetic algorithm that replicates these properties, provided that sensory inputs adhere to a common statistical structure. However, in natural, unregulated environments, this constraint cannot be assured. We here present a series of signal conditioning steps, inspired by the mammalian olfactory system, that transform diverse sensory inputs into a regularized statistical structure to which the learning network can be tuned. This preprocessing enables a single instantiated network to be applied to widely diverse classification tasks and datasets -here including gas sensor data, remote sensing from spectral characteristics, and multilabel hierarchical identification of wild species -without adjusting network hyperparameters.

A Systematic Framework for Olfactory Bulb Signal Transformations

“…Some aspects of learning in the wild simply require the inclusion of established glomerular-layer computations affording concentration tolerance and regulated contrast enhancement (Imam et al, 2012 ; Borthakur and Cleland, 2019a , b ). Enabling the explicit representation of similarity further requires that we loosen the strict controls over the allocation of adult-born neurons used in the Loihi model (Imam and Cleland, 2020 ), thereby enabling granule cell interneurons to belong to multiple ensembles according to the similarity profiles among learned odor representations.…”

“…Heterogeneity in GC spike thresholds, for example, generates a controllable range in the order of their post-differentiation receptive fields, rendering some highly selective for specific known odorants and others more broadly tuned. Similarly, the duplication of MTCs in each MOB column, coupled with heterogeneity in their excitability properties, facilitates the statistical regularization of sensory input data, thereby enabling the network to respond effectively to a wider diversity of sensory inputs (Borthakur and Cleland, 2019a ). In neuromorphic systems, these properties enable single, parameterized networks to function effectively when presented with widely disparate datasets—from chemosensor arrays with years of accumulated drift and decay (Borthakur and Cleland, 2019b ) to non-olfactory datasets with diverse input statistics (Borthakur and Cleland, 2019a ).…”

Sapinet: A sparse event-based spatiotemporal oscillator for learning in the wild

“…As SNN parameters are hard to tune for use with data of different dimensions, we also here provide a model scaling technique for the systematic tuning of SNNs with respect to sensor array size -an essential requirement of learning in the wild. Moreover, as shown for the signal conditioning preprocessor layer [3,4], we demonstrate that the Sapinet architecture can utilize rapid online learning without catastrophic forgetting as a solution for mitigating sensor drift.…”

“…To assess the consistency of MC and GC spike counts, the goodness of preprocessing as described in Borthakur & Cleland [4] is used with additional constraint -spike counts of MCs / GCs cannot be greater then a set threshold ( set to 0.9 here. )…”

“…We modeled the MCsoma to GC connection pattern as sparse random excitatory synapses with local learning (asymmetric STDP, see Methods). Similar to previous works [3,4], these synapses were modeled as double exponentials. The effect of this plasticity model was that, during training, granule cells (GCs) formed higher order receptive fields, learning to respond only when particular sets of MCs fired together, hence becoming selective to odor-diagnostic patterns among specific groups of co-activated MCs.…”