Rapid online learning and robust recall in a neuromorphic olfactory circuit

Model architecture

• The Loihi chip model instantiated hypothesized mammalian olfactory-bulb principles, including a dynamically acquired, learning-dependent EPL lateral-inhibitory topology and the importance of gamma-discretized, spike-timing-based EPL computation (12, 14, 15, 17, 35).Source: Model architecture · pp. 5–6
• The architecture incorporates a minority of principal neurons in gamma dynamics per stimulus, permanent granule-cell differentiation requiring adult neurogenesis, and neuromodulation as an optimization trajectory.Mitral cells excite granule cells irrespective of distance, whereas granule cells inhibit mitral cells only locally and through granule-cell spiking.Source: Model architecture · p. 6
• Each implicit network column contains one mitral-cell principal neuron and up to 50 granule-cell interneurons, with sparse intercolumnar excitation and local inhibition.Intercolumnar connection probability is 0.2.Source: Model architecture · p. 6
• A 72-column network received samples from 72 metal-oxide gas sensors distributed across a 1.2-m-wide wind tunnel, with sniffs drawn from 180-s odorant presentations.Sensor-amplitude patterns reflected odor quality and concentration variance caused by plume dynamics.Source: Model architecture · p. 6
• During permissive epochs of an ongoing gamma oscillation, stronger sensor excitation evoked earlier mitral-cell spikes, creating odor-specific spatiotemporal precedence codes across the population.Inhibitory epochs reset and held mitral-cell activation at zero; the spike pattern repeated across successive gamma cycles in the absence of learning with stationary input.Source: Model architecture · p. 7
• The dynamical architecture enables multiple iterative processing cycles per sample by exploiting the timescale gap between sampling and processing.The implementation embedded five gamma cycles, each taking 0.4 ms, within every training and testing sniff.Source: Model architecture · p. 7
• Five 0.4-ms gamma cycles per sniff let feedback inhibition iteratively alter mitral-cell spike timing and the precedence code, producing evolving representations for classification.Each representation consists of spikes identified by active mitral cell and latency within its gamma cycle; representations are classified by similarity to those known by the network.Source: Model architecture · pp. 7–8

Excitatory plasticity determines GC receptive fields

• Distributed presynaptic MCs drive GC inputs; when a GC reaches threshold, it spikes and inhibits its cocolumnar MC in the next gamma cycle.Source: Excitatory plasticity determines GC… · p. 8
• A gamma-timescale STDP rule strengthens synapses whose MC spikes immediately precede a GC spike and weakens other incoming synapses, including inactive ones.Source: Excitatory plasticity determines GC… · p. 8
• After training, each odorant evokes distributed GC ensembles tuned to distinct higher-order correlations, with each GC retaining a synchronous set of k MC inputs as other weights decay to zero.Source: Excitatory plasticity determines GC… · pp. 8–9

Inhibitory plasticity denoises MC representations

• GC-mediated inhibition delays local MC spikes, and inhibitory windows are iteratively tuned until release coincides with an MC apical-dendrite threshold crossing.Source: Inhibitory plasticity denoises MC… · p. 9
• Inhibitory plasticity learns odor-specific MC timing relationships, encoding relative activation levels so odors engaging the same MC population at different levels remain distinguishable.Source: Inhibitory plasticity denoises MC… · pp. 9–10
• This spatiotemporal encoding supports greater memory capacity than spatial patterning alone and uses fewer spikes with faster readout than rate coding.Source: Inhibitory plasticity denoises MC… · p. 10
• Learned odor memories act as attractors that correctly classify incoming stimuli despite surprisingly high destructive interference.Source: Inhibitory plasticity denoises MC… · p. 10

Odor learning enables identification of occluded stimuli

• Under impulse noise (P = 0.6), activity evoked by occluded toluene moved toward its learned representation and enabled identification, unlike a naïve network’s uninformative response; suppressing inhibitory plasticity blocked denoising.The toluene network was trained in one shot over five gamma cycles before testing.Source: Odor learning enables identification… · pp. 10–11
• Odor learning permanently differentiates granule cells into selective higher-order feature detectors, while increasing undifferentiated granule cells per column improved signal-restoration similarity after five gamma cycles.Simulations used five granule cells per trained odorant and five gamma cycles per sniff to avoid ceiling effects.Source: Odor learning enables identification… · p. 11

Adult neurogenesis enables lifelong learning

• Because odor memories irreversibly consume differentiated granule cells, a declining undifferentiated-cell pool would hinder later learning, so lifelong learning requires a continuing source of new cells.The authors propose adult neurogenesis as the biological source of this resource.Source: Adult neurogenesis enables lifelong… · pp. 11–12
• The model simulated adult neurogenesis by adding five new granule cells per column after each odor; training 72 columns on ten odors yielded 3,600 differentiated granule cells.New granule cells received initial connections from a random 20% of mitral cells across the network and inhibited their cocolumnar mitral cell.Source: Adult neurogenesis enables lifelong… · p. 12

Online learning of multiple representations

• Sequential one-shot training produced similar results regardless of odor order, and later training did not disrupt previously learned odor memories, demonstrating resistance to catastrophic forgetting.A new set of undifferentiated granule cells was added after each odor was learned.Source: Online learning of multiple… · p. 12
• With impulse-noise occlusion P = 0.6, the network rapidly identified occluded instances of all ten odors within five gamma cycles; under standard conditions, performance was strong up to sample occlusion P = 0.6.An odor was considered identified when its evoked spiking pattern exceeded a Jaccard similarity of 0.75 to one learned representation.Source: Online learning of multiple… · p. 13

Neuromodulation and cortical priming improve classification performance

• The model treats neuromodulation as a dynamic search trajectory rather than a single state, allowing transient states across successive sniffs to govern stimulus identification.Neuromodulator concentration was modeled as rising around effector neurons and synapses during active olfactory sampling.Source: Neuromodulation and cortical priming… · p. 13
• Neuromodulation was implemented as a gradual reduction in GC spiking threshold over five sniffs, and classification used the greatest similarity measured in each sniff’s last gamma cycle.The changing threshold represented increasing neuromodulator efficacy.Source: Neuromodulation and cortical priming… · pp. 13–14
• Neuromodulatory states approximately doubled classification performance at P = 0.8, with different states performing best for different test odors and noise instantiations.The strategy evaluated the greatest similarity value measured in the last gamma cycle of each sniff across five progressively changing states.Source: Neuromodulation and cortical priming… · p. 14
• At P = 0.9, priming odor-responsive GCs by lowering their spike thresholds improved classification in proportion to the primed fraction, allowing a weak prior to retrieve occluded odors.The unprimed network largely failed to classify samples under this destructive interference.Source: Neuromodulation and cortical priming… · p. 14

Sample variance arising from plume dynamics

• Plume dynamics alone were a relatively minor source of variance compared with impulse noise, while test activity moved toward the corresponding learned representation over five gamma cycles.The samples were drawn at different times to capture evolving odor plumes after one-shot training on all ten odors.Source: Sample variance arising from plume… · p. 15
• With plume variation and impulse noise at P = 0.4, activity was again attracted over five gamma cycles toward the correct learned representation.Each odor was sampled at widely dispersed times, with independent impulse-noise instantiations across sniffs.Source: Sample variance arising from plume… · p. 15
• Across impulse-noise levels from P = 0.0 to P = 1.0, plume variability moderately reduced performance, while introducing noise correlations did not affect it.Performance was compared with the green curve in Figure 4e.Source: Sample variance arising from plume… · p. 15

Classification performance of the neuromorphic model

• With one-sample learning, EPL substantially outperformed DAE and the median-filter, total-variation-filter, and PCA baselines.The one-sample DAE received one training sample from each odorant over 1000 epochs per odorant, without information about the impulse-noise distribution arising during testing.Source: Classification performance of the… · p. 16
• 3000 samples per odor let the DAE match EPL's classification performance from 1 sample per odor; with further training, DAE performance exceeded EPL's.The DAE was trained on samples independently occluded with impulse noise drawn from P = [0.2, 0.8], including information about testing variance.Source: Classification performance of the… · p. 16
• After sequential training on toluene and then acetone, EPL retained high-fidelity recognition of both odors, whereas DAE forgot toluene.Both networks initially performed well on toluene; the loss appeared after subsequent acetone training.Source: Classification performance of the… · pp. 16–17
• EPL was competitive overall but excelled at rapid online learning and generalization under unpredictable variance, while a converged DAE exceeded it when trained on expected variance.The authors describe EPL as suited to rapid training and novel environments with unpredictable variance, while the evaluated DAE performed best when trained to convergence on data drawn from expected variance.Source: Classification performance of the… · p. 17

Discussion

• EPL combines autoassociative attractor dynamics with spike-timing computation, proposing roles for plasticity, packet communication, neuromodulation, and adult neurogenesis.The authors distinguish the work by rapid learning, identification of strongly occluded signals, and a field-deployable Loihi implementation.Source: Discussion · p. 18

Summary

• The Loihi-implemented model supports rapid online learning and signal restoration for odor inputs strongly occluded by contaminants.The authors describe the framework as based on mammalian olfactory-bulb architecture and dynamics.Source: Summary · p. 19
• The framework is presented as applicable to steady-state identification of high-dimensional patterns lacking meaningful lower-dimensional structure in highly interfering backgrounds.The authors also frame the results as relevant to understanding mammalian olfaction and improving artificial chemosensory systems.Source: Summary · p. 19

Funding

• The work was supported by NIDCD R01 grants DC014701 and DC014367 to TAC.Source: Funding · p. 19

Dataset and odorant sampling

• The study used a 72-sensor metal-oxide wind-tunnel dataset, selecting recordings from location L4 at 0.21 m s-1 and 500 V.The source dataset is attributed to Vergara and colleagues 4.Source: Dataset and odorant sampling · p. 20
• Ten gas-phase odorants were sampled, with 10–20 presentations per odorant lasting 180 seconds and one randomly chosen plume used for each odorant.The odorants were acetone, acetaldehyde, ammonia, butanol, ethylene, methane, methanol, carbon monoxide, benzene, and toluene.Source: Dataset and odorant sampling · pp. 20–21
• For plume-variance studies, test samples for each odorant were drawn from different plume time points, spanning 30-180 seconds after presentation at 5-second intervals, and tested with and without added impulse noise.Source: Dataset and odorant sampling · p. 21

Intrinsic gamma and theta dynamics

• MC spike phases relative to gamma oscillations are the model’s most informative output, and repeated cycles let the network iteratively approach a learned state under stationary sensory input.Each oscillation provides a distinct, interpretable representation.Source: Intrinsic gamma and theta dynamics · p. 23
• GC inhibition can alter MC spike timing on gamma timescales, while MC spikes alter GC responsivity, enabling lateral interactions to iteratively modify information exported from the olfactory bulb.Source: Intrinsic gamma and theta dynamics · p. 23
• Each MC alternates between a permissive epoch for sensory integration and spiking and an inhibitory epoch with excitation held at zero, forming a 40-timestep gamma cycle.The permissive epoch lasts 16 timesteps and the inhibitory epoch 24.Source: Intrinsic gamma and theta dynamics · p. 23
• The permissive epoch sets the number of discrete sensory-input levels encoded by the spike-timing metric, and that range can expand at the cost of greater time and energy.Source: Intrinsic gamma and theta dynamics · p. 23
• Each sniff-like sampling cycle presents one steady sensory sample across five gamma cycles, with the number of cycles adjustable to regulate iterative processing per sample.The implementation held the count at five.Source: Intrinsic gamma and theta dynamics · p. 24
• 200 timesteps (five gamma cycles) take around 2 ms per sniff on Loihi, five times faster than 100-Hz sensor sampling, allowing each sample to be iteratively processed before the next sensor update.The implementation uses 200 timesteps to process one sniff; the sensor state does not update during those five processing cycles.Source: Intrinsic gamma and theta dynamics · p. 24

Mitral cells

• Apical dendrites integrate sensor input during each permissive gamma epoch, triggering a soma-directed event at threshold; stronger inputs reach threshold earlier.Each apical dendrite can initiate only once per permissive epoch.Source: Mitral cells · p. 25
• The soma propagates the event as an action potential when granule-cell inhibition releases; this inhibition modulates mitral-cell spike timing relative to the gamma cycle.In the first gamma cycle after odor presentation, before granule-cell inhibition is active, the soma propagates the spike immediately.Source: Mitral cells · p. 25
• Mitral-cell spikes serve as network output to the classifier and are also delivered to postsynaptic granule cells; both compartments reset during the inhibitory epoch.After spiking, the soma cannot spike again during that permissive epoch.Source: Mitral cells · pp. 25–26

Granule cells

• Single-compartment granule cells summed weighted mitral-cell excitation, spiked above threshold, reset their excitation, and were refractory for 20 timesteps, so convergent input was generally needed and each cell spiked at most once per gamma cycle.Source: Granule cells · p. 26
• Synapse-specific excitation delays were randomly assigned and stable, reflecting propagation-delay heterogeneity and timing granule-cell spikes within gamma’s inhibitory epoch.Source: Granule cells · p. 26

Excitatory synaptic plasticity

• MC-to-GC weights initialized at we; during training after GC spikes, asymmetric additive STDP potentiated synapses whose MC spike preceded the GC spike by 1 timestep by 0.05we, depressing all others by 0.2we.Source: Excitatory synaptic plasticity · pp. 26–27
• Synaptic weights were capped at 1.25we, keeping differentiated granule cells sensitive to coincident activity in a particular ensemble of mitral cells, whose size determines receptive-field order.Source: Excitatory synaptic plasticity · p. 27
• Odor learning transforms broad initial granule-cell receptive fields into highly selective fields of order M, tuned to learned odor signatures and unresponsive to other input combinations.These higher-order fields reflect correlations between components of individual sensor vectors—the higher-order signatures of learned odors.Source: Excitatory synaptic plasticity · p. 27
• Granule-cell output denoises the spike-timing-based mitral-cell representation, which mitral cells can communicate for classification or further processing with reduced bandwidth and energy.There are many fewer mitral cells than granule cells.Source: Excitatory synaptic plasticity · p. 27

Adult neurogenesis

• Because GC differentiation permanently depletes interneurons available for new odor representations, the network periodically adds undifferentiated GCs more slowly than synaptic plasticity operates.It starts with five GCs per column and adds five per column after learning each odor.Source: Adult neurogenesis · p. 28

Inhibitory synaptic plasticity

• GC-to-cocolumnar-MC inhibitory synapses default to inactive, then enter a blocking state after a GC spike, inhibiting the MC soma at unit strength for learned duration ΔB.Blocking inhibits somatic spike propagation, and ΔB functions as the inhibitory synaptic weight.Source: Inhibitory synaptic plasticity · p. 28
• After blocking, each synapse enters a one-timestep release state that provides unit excitation to the MC soma before returning to inactive.New GCs' inhibitory weights start at ΔB = 0 ts.Source: Inhibitory synaptic plasticity · pp. 28–29
• An MC soma spikes only when its combined dendritic excitation and inhibition from presynaptic GCs is positive, and cannot spike again during that gamma cycle.During training, inhibition's effects on MC somata are suppressed.Source: Inhibitory synaptic plasticity · p. 29
• During training, a cocolumnar MC spike in the permissive epoch after a GC spike modifies ΔB so inhibition release aligns with MC spike initiation.The update uses the MC dendritic spike-initiation time, inhibition-release time, and learning rate η.Source: Inhibitory synaptic plasticity · p. 29
• If a GC spike is not followed by MC spike initiation in the permissive epoch, that synapse's ΔB grows until inhibition spans the entire gamma cycle.Inputs from multiple local GCs onto a common MC are modified independently.Source: Inhibitory synaptic plasticity · p. 29
• Inhibitory plasticity trains a fixed-point attractor around each learned odor representation, countering destructive interference during testing.The rule learns timing relationships between GC spikes and cocolumnar MC spikes associated with an odor stimulus.Source: Inhibitory synaptic plasticity · pp. 29–30
• The rule learns odor-specific relative MC activation levels, distinguishing odors that activate the same MC population at different levels.This captures ratiometric activation patterns among MCs.Source: Inhibitory synaptic plasticity · p. 30

Testing procedures

• Occlusion replaced a fraction P of the 72 sensor values with independent uniform random values from 0 to 15, with occluded elements and values redrawn across test samples.All testing was performed with network plasticity disabled.Source: Testing procedures · p. 30
• After one-shot training on one plume sample, the network was tested on 30 samples drawn every 5 seconds from 30 to 180 seconds, with and without impulse noise.Summary figures averaged across 30 timepoint samples and, where applicable, 100 independent impulse-noise instantiations per training odor.Source: Testing procedures · pp. 30–31, Sample classification · p. 32
• In few-shot mode, the network gradually adapted to training-sample statistics and learned robust representations even when training samples were corrupted by impulse noise.The paper illustrates this effect in supplementary Figure S2.Source: Testing procedures · p. 31

Sample classification

• Each sniff was represented by five successive gamma-cycle spike sets, with each spike specified by its mitral-cell identity and latency from the permissive epoch’s onset.The epoch contained 16 discrete timesteps used for Jaccard calculations.Source: Sample classification · p. 31
• For occluded samples, similarities were computed between each of five unknown-sample representations and learned odor representations using the Jaccard index of spike-set intersection over union.Descriptive figures used the Jaccard index; comparisons with other methods did not.Source: Sample classification · p. 31
• Samples were classified as known odors when similarity exceeded 0.75 in the fifth gamma cycle; if multiple odors crossed the threshold, the highest similarity across five cycles won, otherwise the sample was unknown.This combined nearest-neighbor classification with thresholding to allow a “none of the above” outcome.Source: Sample classification · pp. 31–32

Benchmarks

• The seven-layer autoencoder outperformed the tested six-, five-, four-, and three-layer networks on the same training set.The seven-layer network was used for the benchmark comparison with EPL.Source: Benchmarks · p. 33
• The DAE required 3000 training samples per odorant to achieve the classification performance the EPL model achieved with 1 training sample per odorant, making EPL 3000 times more data efficient.Source: Benchmarks · p. 34
• After acetone training, the DAE’s toluene representation declined until it could no longer correctly classify toluene.The DAE initially classified toluene test presentations with high fidelity, but its similarity to the learned toluene representation progressively fell during acetone training.Source: Benchmarks · p. 34
• EPL training on acetone did not interfere with its existing toluene representation, which remained unaffected as EPL learned all ten odorants in sequence.The passage reports no change in similarity between toluene test samples and the learned toluene representation during sequential learning.Source: Benchmarks · p. 34

• Loihi cores integrate filtered presynaptic spike trains and emit threshold-triggered spikes, with configurable spike-timing-dependent learning, conduction delays, and neuromodulation.Source: Implementation on the Loihi… · p. 35
• Each model column occupied one neuromorphic core, using 72 cores on one chip; cocolumnar interactions stayed within cores, while global MC somatic spikes used the intercore mesh.Source: Implementation on the Loihi… · p. 35
• A 72-core inference cycle (one sniff, five gamma cycles, 200 timesteps) took 2.75 ms and consumed 0.43 mJ, including 0.12 mJ of dynamic energy.Source: Implementation on the Loihi… · p. 36
• Inference time was not significantly affected by problem scale, while energy consumption increased only modestly as network size grew, aided by colocalized memory and compute and sparse spiking communication.Source: Implementation on the Loihi… · p. 36