PaperlayerView exact revision on arXiv

Source-linked AI summary

Rapid online learning and robust recall in a neuromorphic olfactory circuit

Nabil Imam, Thomas A. Cleland

arXiv:1906.07067v3cs.NEq-bio.NC

TL;DR

The paper addresses odor learning and identification under noise, including the need for feature combinations and continued learning despite catastrophic forgetting. It proposes a spike-timing framework with attractor dynamics, plasticity, packet-based communication, neuromodulation, and adult neurogenesis, demonstrating signal restoration for odor inputs strongly occluded by contaminants.

  • Problem

    Odor learning and identification under noise require feature combinations, while catastrophic forgetting limits deep networks and lifelong learning requires a steady source of undifferentiated GCs.

  • Method

    The approach instantiates autoassociative attractor dynamics within a spike-timing framework using synaptic plasticity, packet-based neural communication, active neuromodulation, and adult neurogenesis.

  • Results

    The system demonstrates powerful signal restoration for odor inputs strongly occluded by contaminants and reveals computational features of the early olfactory network.

  • Takeaways & Limitations

    Spike-timing mechanisms, neuromodulation, and adult neurogenesis provide functional roles for algorithmic approaches to robust signal identification.

  • Takeaways & Limitations

    The learning algorithm irreversibly consumes GCs, so lifelong learning requires a steady source of undifferentiated GCs.

Abstract

from arXiv · show

We present a neural algorithm for the rapid online learning and identification of odorant samples under noise, based on the architecture of the mammalian olfactory bulb and implemented on the Intel Loihi neuromorphic system. As with biological olfaction, the spike timing-based algorithm utilizes distributed, event-driven computations and rapid (one-shot) online learning. Spike timing-dependent plasticity rules operate iteratively over sequential gamma-frequency packets to construct odor representations from the activity of chemosensor arrays mounted in a wind tunnel. Learned odorants then are reliably identified despite strong destructive interference. Noise resistance is further enhanced by neuromodulation and contextual priming. Lifelong learning capabilities are enabled by adult neurogenesis. The algorithm is applicable to any signal identification problem in which high-dimensional signals are embedded in unknown backgrounds.

Results

The model implements mammalian olfactory-bulb principles as an event-driven spiking network, using gamma-timed activity and plasticity to build odor representations. Iterative processing and learned inhibitory structure produce evolving representations for classification.

  • Temporal coding: Gamma-constrained mitral-cell spikes encode sensor activation through a hybrid channel-and-phase code across successive processing cycles.Sensory integration occurs during permissive epochs, while inhibitory epochs reset mitral-cell activation; stronger excitation produces earlier spikes.
  • Iterative processing: Five gamma cycles embedded within each sniff allow feedback inhibition to iteratively modify spike timing and generate evolving odor representations.In the Loihi implementation, each gamma cycle requires 0.4 ms.
  • Plasticity: Heterosynaptic STDP makes granule cells selective for specific combinations of mitral-cell spikes, effectively implementing a k-winners-take-all learning rule.Synapses preceding a postsynaptic spike are strengthened while other incoming synapses are weakened, causing nonselected inputs to decay toward zero.
  • Plasticity: Granule-cell receptive fields learn higher-order stimulus correlations and deliver local inhibition to mitral cells in subsequent gamma cycles.Granule-cell spikes inhibit their cocolumnar mitral cell, while distinct learned combinations produce diverse receptive fields.

Inhibitory plasticity denoises MC representations

Inhibitory plasticity learns odor-specific timing relationships, allowing the network to transform corrupted inputs toward stored representations. This produces timing-based attractor memories with greater capacity and rapid, interference-resistant classification.

  • Inhibitory plasticity: Inhibitory plasticity learns timing relationships among mitral-cell spikes, thereby encoding odor-specific ratiometric activation patterns.Relative spike times represent mitral-cell activation levels, so learned inhibition captures relationships among those levels.
  • Representation capacity: Spike-timing representations distinguish odors sharing the same active mitral cells but differing in relative activation levels.The spatiotemporal code provides greater memory capacity than spatial patterning alone.
  • Attractor denoising: Timing-based relational encoding and odor-specific feedback inhibition create attractor states that classify inputs despite destructive interference.Incoming stimuli are classified by their similarity to learned representations.
  • Attractor denoising: After one-shot training on toluene, the network attracted occluded toluene activity toward the learned representation across five gamma cycles.Suppressing inhibitory plasticity prevented the trained network from denoising the mitral-cell representation.
  • Capacity under noise: Increasing undifferentiated granule-cell numbers improved signal restoration by increasing similarity between occluded inputs and learned representations.The simulations used five granule cells per trained odorant and five gamma cycles per sniff to avoid ceiling effects.
  • Capacity under noise: The algorithm irreversibly consumes granule cells, so lifelong learning requires a steady source of undifferentiated granule cells.Successive odor learning becomes increasingly handicapped as the undifferentiated-cell pool declines.

Online learning of multiple representations

Sequential learning lets the network acquire multiple odors without disrupting previously learned memories. Neuromodulatory trajectories and contextual priming further improve identification when sensory inputs are heavily corrupted.

  • Online learning: Sequentially trained odors remain recognizable because subsequent odor training does not disrupt previously learned memories.The network therefore supports robust online learning and resists catastrophic forgetting.
  • Odor identification: Strongly occluded samples from all ten odors were attracted toward their corresponding learned representations and identified within five gamma cycles.Impulse noise replaced 60% of sensory inputs in this evaluation.
  • Odor identification: Classification performance remained strong through occlusion level P = 0.6, after which increasing occlusion gradually impaired performance.The occlusion was generated by replacing a proportion of sensory inputs with random values.
  • Neuromodulation: Traversing multiple neuromodulatory states approximately doubled classification performance at impulse-noise level P = 0.8.Different states performed best for different odors and noise instantiations, motivating a trajectory rather than a single state.
  • Contextual priming: Contextual priming dramatically improved classification at P = 0.9, with gains proportional to the fraction of odor-responsive granule cells primed.Priming lowered the spike thresholds of granule cells normally activated by the presented odor.

Sample variance arising from plume dynamics

Plume dynamics introduce time-varying sample differences, but they are a smaller variance source than impulse noise and only moderately reduce EPL performance. The network still converges toward the correct learned odor representations across plume variation and combined contamination.

  • Samples of the same odorant differed over time because of evolving plume dynamics, while different odorants also differed in sensor-array analyte sensitivities.
  • After one-shot training on all ten odors, test-sample activity was attracted over five gamma cycles toward the corresponding learned representation.
  • Plume dynamics alone constituted a relatively minor source of variance compared with impulse noise.
  • With plume variation and impulse noise at P = 0.4, activity again converged toward the correct representation across contaminated, time-varying samples.
  • Adding plume-based variability moderately reduced performance, whereas introducing noise correlations did not affect network performance.
  • The EPL substantially outperformed MF, TVF, and PCA, and exceeded a one-sample-trained DAE when testing noise was unseen during training.

Summary

The study develops a simplified mammalian-olfactory-bulb network on Loihi using sensor-array inputs and deliberately general, minimally dataset-specific processing. It targets rapid, robust signal identification in interfering backgrounds and broader deployment settings.

  • The simplified olfactory-bulb model was instantiated on Loihi to support rapid online learning and restoration of strongly occluded odor inputs.
  • The framework applies to high-dimensional signal-identification problems lacking meaningful lower-dimensional structure and embedded in highly interfering backgrounds.
  • Inputs came from 72 metal-oxide sensors and were represented as discretized, sparsified 72-element activity vectors.
  • Sensors were sampled naively without cross-referencing duplicate sensor types or mitigating plume-based variance, preserving algorithmic generality over dataset-specific optimization.
  • The model used ten gas-phase odorants sampled from a wind-tunnel dataset, with recordings taken at a specified sensor location, wind speed, and heater voltage.
  • The Loihi implementation mapped each sensor to an olfactory-bulb column containing one principal neuron and inhibitory granule-cell circuitry.

Intrinsic gamma and theta dynamics

The model separates fast gamma-timescale processing from slower theta-like sampling, using spike phase and repeated gamma cycles to iteratively transform sensory input. This supports continuous processing faster than sensor sampling.

  • Gamma-phase spike timing is the model’s informative output, while repeated oscillations let the network approach a learned state from stationary sensory input.
  • Lateral inhibitory interactions iteratively modify information exported from the olfactory-bulb model through reciprocal effects on mitral and granule-cell spike timing.
  • Each gamma cycle alternated a permissive epoch for sensory integration and spiking with an inhibitory epoch that prevented spike generation.
  • The implementation used 16 timesteps for the permissive epoch and 24 for inhibition, totaling 40 timesteps per gamma cycle.
  • Each sniff presented one steady-state sample across five gamma cycles, allowing multiple iterative computations before the next sample.
  • Around 2 ms were required to process one sniff on Loihi, five times faster than the sensors’ 10-ms sampling interval.

Mitral cells

Mitral cells integrate sensor inputs in apical dendrites and use gamma-timed inhibition to regulate somatic spike propagation. Their spike timing provides the classifier representation, while plasticity produces selective higher-order odor features.

  • Mitral cells: Each mitral cell was modeled with apical dendrite and soma compartments that separately handle sensory excitation and lateral inhibition.Sensor activation enters the apical dendrite; the soma integrates dendritic excitation with granule-cell inhibition.
  • Mitral cells: Stronger sensory inputs initiate earlier phase-leading mitral-cell spikes, whereas presynaptic granule-cell inhibition can delay their propagation.Inhibition primarily modulates spike timing relative to the gamma cycle.
  • Mitral cells: Mitral cells can generate at most one spike per permissive gamma epoch before apical-dendrite and soma compartments reset during inhibition.This event-driven cycle gates subsequent spike initiation and propagation.
  • Mitral cells: Granule cells require convergent mitral-cell excitation, spike at most once per gamma cycle, and provide lateral inhibition to mitral-cell somata.Their excitatory synapses are weighted and their activity is delayed into the inhibitory gamma epoch.
  • Mitral cells: Repeated spike-timing-dependent plasticity potentiates synapses from coactive mitral cells and depresses others, creating sparse, selective higher-order granule-cell receptive fields.Learned fields capture correlations among sensor-vector components and become unresponsive to other combinations.
  • Mitral cells: Using mitral-cell spikes for classification reduces communication bandwidth and energy consumption because the network contains fewer mitral cells than granule cells.The granule-cell output is instead used to denoise the spike-timing-based mitral-cell representation.

Adult neurogenesis

Granule-cell differentiation consumes available interneurons as odors are learned. Periodic addition of new undifferentiated granule cells maintains the network’s capacity for lifelong odor learning.

  • Adult neurogenesis: Granule-cell differentiation permanently depletes interneurons available for recruitment into new odor representations.This creates a capacity constraint as the number of learned odors increases.
  • Adult neurogenesis: The network periodically adds new, undifferentiated granule cells on a slower timescale than synaptic plasticity.This process is modeled on adult neurogenesis in the olfactory bulb.
  • Adult neurogenesis: After each newly learned odor, five additional undifferentiated granule cells are configured in every column.The network initially contains five granule cells per column, and new cells receive probabilistic excitatory and inhibitory connectivity.

Inhibitory synaptic plasticity

Inhibitory synaptic plasticity learns the timing relationship between granule-cell and mitral-cell spikes. The resulting inhibitory weight matrix forms odor-specific attractors that counter destructive interference and distinguish relative activation patterns.

  • Inhibitory synaptic plasticity: Granule-cell synapses onto mitral-cell somata transition among inactive, blocking, and release states.Blocking applies unit inhibition for a learned duration, while release briefly restores excitation.
  • Inhibitory synaptic plasticity: The learned blocking period governs mitral-cell spike latency and functions as the inhibitory synaptic weight.The synapse releases inhibition after a duration ΔB determined during training.
  • Inhibitory synaptic plasticity: The rule adjusts inhibition when granule-cell spikes are followed by mitral-cell spike initiation, aligning inhibitory release with the mitral-cell event.If no following mitral-cell event occurs, inhibition grows until it spans the entire gamma cycle.
  • Inhibitory synaptic plasticity: The inhibitory rule trains the network to learn granule-cell–mitral-cell spike timing relationships associated with an odor.Multiple local granule-cell inputs to a common mitral cell are modified independently.
  • Inhibitory synaptic plasticity: Training constructs a fixed-point attractor around each learned odor representation, counteracting destructive interference during testing.The rule also learns ratiometric activation patterns, allowing odors with the same active mitral-cell population to be distinguished.

Testing procedures

Testing evaluates learned odor recognition across sensor occlusion, impulse noise, and variation in odor-plume sampling time. The network is tested after one-shot learning with plasticity disabled, while few-shot training is also examined for corrupted samples.

  • Testing procedures: Occluded test samples replace a fraction P of the 72 sensor-vector elements with uniformly sampled values from 0 to 15.Occluded positions and replacement values are redrawn for each sample used to estimate average performance.
  • Testing procedures: Destructive impulse noise models competing background odorants that alter or block receptors and disrupt odor-specific ratiometric activation patterns.Occlusion is implemented by replacing selected sensor activity values with random operating-range values.
  • Testing procedures: Odor-plume variance is tested by sampling 30 points per plume at 5-second intervals from 30 to 180 seconds.Samples are drawn from 180-second datastreams at different timepoints.
  • Testing procedures: After one-shot training on a single sample, the network is tested on other samples from the odor plume.Testing includes conditions with and without impulse noise, with plasticity disabled throughout evaluation.
  • Testing procedures: In few-shot mode, the network gradually adapts to training-sample statistics and learns robust representations from impulse-noise-corrupted samples.This effect is illustrated in supplementary Figure S2.

Sample classification

The network converts sequential gamma-cycle MC spike patterns into odor representations and classifies them with similarity thresholding. Compared with a deep autoencoder, it learns odors with far fewer samples and avoids sequential-learning interference.

  • Representation: Each odor representation records MC spike identities and latencies across five successive gamma cycles.The spike pattern in each cycle is represented using 16 discrete timestep bins.
  • Classification: Samples are assigned to the most similar known odor when final-cycle similarity exceeds 0.75; otherwise, they are classified as unknown.When multiple odorants exceed threshold, the odorant with the greatest similarity across five cycles is selected.
  • Comparisons: The EPL network was compared with median filtering, total variation filtering, PCA, and a seven-layer deep autoencoder using a shared nearest-neighbor classifier.Outputs were normalized vectors, and similarity was computed from Manhattan distance.
  • Data efficiency: 3000 training samples per odorant were required by the DAE to match EPL performance achieved with 1 training sample per odorant.The authors report that EPL was 3000 times more data efficient than the DAE.
  • Sequential learning: Sequential acetone training progressively degraded the DAE’s toluene representation until toluene classification failed.EPL training on acetone did not interfere with the preexisting toluene representation while learning all ten odorants sequentially.

Implementation on the Loihi neuromorphic system

The olfactory network was mapped onto Intel Loihi’s distributed neuromorphic cores, using configurable spiking, plasticity, routing, and neuromodulatory mechanisms. Its inference was fast and energy-efficient, with fine-grained parallelism supporting scalability.

  • Neuromorphic architecture: Each Loihi core uses leaky-integrate-and-fire units that integrate filtered presynaptic spikes and emit thresholded postsynaptic spikes.Cores support configurable spike-timing-dependent learning, conduction delays, and neuromodulatory effects.
  • Network mapping: The implementation assigned each model column to one core, using 72 cores on a single Loihi chip.Cocolumnar interactions remained within cores, while global MC projections used the intercore routing mesh.
  • Inference cost: 2.75 ms and 0.43 mJ were required for one 72-core inference cycle consisting of five gamma cycles and 200 timesteps.Dynamic energy accounted for 0.12 mJ.
  • Scalability: Inference time was not significantly affected by problem scale, owing to Loihi’s fine-grained parallelism.Energy consumption increased only modestly as network size increased.

Acetone

The neuromorphic olfactory model encodes chemosensor activity through mitral-cell spiking and granule-cell interactions, using plasticity to form odorant-specific representations. Trained networks denoise strongly occluded samples over successive gamma cycles and retain odor recognition during sequential learning.

  • Model structure: Sensor activation determines the timing of mitral-cell spikes, which propagate through lateral dendrites to granule cells and back onto cocolumnar mitral cells.The architecture uses sensor input at mitral-cell apical dendrites and granule-cell-mediated inhibition.
  • Plasticity: Repeated coincident mitral-cell spikes strengthen selected excitatory granule-cell synapses while weakening other inputs, making granule cells selective for higher-order odor features.Inhibitory plasticity also adjusts the duration of granule-cell inhibition until it aligns with mitral-cell spike initiation.
  • Iterative denoising: Occluded test representations are iteratively drawn toward learned odor representations across five successive gamma cycles through granule-cell-mediated inhibition.The Jaccard similarity to toluene increased over five cycles, reaching the classification criterion of similarity > 0.8.
  • Recognition: Inhibitory plasticity was required for denoising, whereas disabling it left the network unable to denoise mitral-cell representations during testing.The trained network increased similarity to toluene over five gamma cycles; untrained and inhibition-disabled networks did not show this systematic increase.
  • Recognition: The trained network reliably recognized all ten odorants from substantially occluded examples at P = 0.6.Toluene-tuned granule cells were progressively recruited while granule cells tuned to the other nine odorants were negligibly recruited.
  • Lifelong learning: Sequential training preserved previously learned odor recognition in the EPL network, unlike the DAE, which forgot toluene after learning acetone.Both networks initially recognized toluene with 100% accuracy; the DAE required 3000 occluded training samples for comparison.
Loading 1906.07067v3…