Does Odor Mixture Decoding Need Recurrence? A Fully Trained Circuit Search Space

Picture a nose smelling coffee. Hundreds of chemicals arrive at once, and each one activates many of the same receptor types. Neuroscience has two opposing accounts of how a brain picks out one smell from such a blend. One says a simple weighted sum over a few receptor channels is enough. The other says the brain needs feedback loops that "explain away" overlapping inputs, a kind of internal inference. Each side has built one hand-designed circuit and shown that it works. Nobody has searched the space of circuits in between to find where the answer changes.

We propose to build that space and train every circuit in it. The search space is a bounded family of small decoding circuits running from sparse linear readouts to deep recurrent inhibitory loops. We train and score every member on simulated and recorded mouse odor-mixture responses, across tasks, mixture sizes and noise levels. The resulting table is a NAS-Bench-style benchmark (a lookup of true performance for every architecture) for olfaction. It serves two purposes. Scientifically, it shows which circuit motifs win in which regime. Methodologically, it gives ground truth for testing whether cheap architecture-ranking shortcuts still work on recurrent circuits fed small, noisy neural data. That test decides whether such search can be afforded on high-density recordings with thousands of neurons over many sessions, where training every candidate is too expensive.

Core Claim

We claim that whether a recurrent "demixing" circuit beats a feedforward readout depends on the task, not on how complex the mixture is. For detecting whether one target odor is present, feedforward readouts trained on mixtures should sit at or near the top of the fully trained table at every mixture size. For reporting the full composition of a mixture, and for decoders trained only on single odors, circuits with input normalization or targeted recurrent inhibition should move into the top decile as the number of components grows. Our second claim concerns the shortcuts: zero-cost proxies, which score an untrained network from one pass of data, will rank these circuits worse than they rank image classifiers. Budgeted short training will keep its rankings, provided the budget covers the number of recurrent steps rather than the number of epochs. Each claim can fail against the table.

How This Relates to Prior Work

Some prior work motivates the question by showing the conflict. Reading out olfactory receptors decodes mouse glomerular calcium images and finds that a linear readout of about 15 glomeruli matches mouse detection accuracy with 14 distractors. The same paper finds that the readout fails when trained on single odors and tested on mixtures. A probabilistic approach to demixing odors takes the other side and derives a recurrent bulb–cortex inference circuit with targeted divisive inhibition. Neural Circuits for Fast Poisson Compressed Sensing in the Olfactory Bulb derives recurrent mitral–granule dynamics that recover tens of odors from a panel of 1,000. All three papers fix one architecture and test it. We place each of their circuits as a point inside a larger space and ask where they rank.

Other work constrains the simulator we build. Antagonism in olfactory receptor neurons models receptors in which odorants compete for binding sites, and this competition normalizes responses before any neural circuit acts. Antagonistic odor interactions in olfactory sensory neurons confirms in vivo that about 80.5% of responding units respond to mixtures below the linear sum. Sparse input representations explain odor discrimination shows that mouse detection is limited by intrinsic neural noise rather than by background complexity. Together these mean that a simulator with linear summation or no noise model would bias the table in favor of recurrent circuits, so ours uses competitive binding and noise matched to the behavior.

The method comes from architecture-search work. EZNAS evolves zero-cost scoring formulas and reports weak discrimination among the best 10% of models. Rethinking Performance Estimation in NAS shows that budgeted training preserves rankings only when the right training settings are pruned. Both were tested on image-classification spaces. Neither covers recurrent circuits, neural inputs, or datasets of a few hundred examples.

Two papers bring neural-recording benchmarks. FALCON and Spikebench evaluate hand-picked decoders on neural recordings, and neither includes odor stimuli. POMMix hand-designs a deep model of mixture perception from chemical structure. It works from molecular graphs, not from neural responses, so it answers a different question.

Research Program

The search space. It is a small grammar of circuits. The input is a glomerular response vector or a vector of binned spike counts. The first choice is input normalization: none, subtractive, or divisive. The second is zero to two feedforward hidden layers, each with a choice of width. An optional sparse expansion layer with winner-take-most sparsity mimics the insect mushroom body. An optional recurrent inhibitory loop comes in global, random or learned-targeted form and runs for 1, 3 or 10 steps. The readout is linear, logistic or multi-label. The linear readout of Mathis et al. and discretized versions of the Grabska‑Barwińska and compressed-sensing circuits are all members of this space. The space should contain a few thousand distinct circuits. That is small enough to train completely and large enough to hold motifs nobody has proposed.

The simulator. We generate synthetic responses from the competitive-binding receptor model. Receptor affinities are drawn from the gamma fit to the 228-glomerulus imaging data used in the compressed-sensing paper. Noise is calibrated to the target-dominated regime of the sparse-input study.

Recorded data. Three datasets carry the real-data evaluation. The Rokni/Mathis glomerular set has 16 odorants and about 500 mixtures. The Reddy set has mixtures of up to 12 components. The Berners-Lee posterior piriform tetrode recordings, from Experience-dependent evolution of odor mixture representations, are deposited on Dryad.

Conditions. There are three tasks: target detection, full composition as a multi-label problem, and concentration estimation. Mixture size runs from 1 to 16 components. There are two training regimes, mixtures versus single odors only, and three seeds.

The full table. We train every circuit in every condition. We then measure how well the zero-cost proxies (Synflow, NASWOT, and EZNAS-A transferred as-is) and budgeted training (fewer epochs, fewer trials, fewer neurons, fewer recurrent steps) reproduce the true rankings. We measure both overall rank agreement and recall of the true top decile.

Extension to high-density recordings. The estimator that best preserves rankings is then used on time-resolved piriform spike data. Because the only olfactory spike data available to us is from tetrodes, this stage also motivates a Neuropixels recording of mixture discrimination. Across sessions, drift makes training every candidate prohibitively expensive.

Baselines

The first reference point is a chance-level decoder. Next to it is a readout trained only on single odors, which Mathis et al. report "fails" on mixtures. This anchors the training-regime effect.

The strongest existing baseline is the sparse linear readout: the Optimal Linear Estimator and L1-regularized logistic regression of Reading out olfactory receptors. The OLE is reported at about 91.6% detection accuracy with 14 distractors, against about 90% for mice. It is the strongest competitor because it already matches behavior at almost no cost. We rerun it on identical splits with the same hyperparameter budget given to each searched circuit. On detection, a searched circuit beats it only with a statistically reliable gain across seeds, or equal accuracy using fewer input channels. On composition and single-odor-trained generalization, the bar is a significant gain in multi-label F1 at mixture sizes of 8 or more.

The second baseline is the pair of hand-derived recurrent circuits: the Grabska‑Barwińska variational circuit and the compressed-sensing MAP circuit. Both are run in their published form and also as members of our space, trained on the same data. If a newly found motif outranks both on composition, that counts as the positive result.

The third baseline is tsfresh features fed to XGBoost, following Spikebench. Spikebench reports about 89% balanced accuracy for this method on retinal stimulus classification, though the file has no reported result for it on odor data. We apply it to the time-resolved piriform spikes with a matched tuning budget.

For the shortcuts, the reference is the fully trained table itself. The baseline proxy is EZNAS-A, which reports Kendall τ = 0.65 on NASBench-201 CIFAR-10. We have found no reported result on Lacuna for any proxy on recurrent circuits or neural data.

The key ablation removes the recurrent-loop and normalization options from the space and reruns everything. If the feedforward-only subspace matches the best circuit in the full space on composition, our first claim is false.

Why not just…?

Why not just compare the three hand-built circuits? For the detection task this may well be enough, and we expect the fully trained table to agree with it. The trouble is that three points cannot show where recurrence starts to pay off along mixture size and noise. Three points also cannot reveal a cheaper motif, such as divisive input normalization alone, that matches recurrence at the same point. The ablation is the direct test. If no circuit outside the three hand-built ones ever enters the top decile, we will report that, and the paper becomes a confirmation rather than a discovery.

Why not just use a sparse linear readout trained on mixtures? Mathis et al. show that it matches mice on detection. That result says nothing about full composition, and it fails under single-odor training. Our claim is precisely that the advantage of more complex circuits is confined to those regimes. If the linear readout stays on top there too, the recurrent-demixing view loses support, and that outcome is informative in itself.

Why not skip the ranking shortcuts and train every candidate, since the models are tiny? For glomerular vectors this objection is correct, and that is why we train the full space rather than search it. The shortcuts matter only for time-resolved, multi-session high-density data. There, FALCON shows drops of up to 0.28 R² across sessions, which forces a search per session. The table tells us whether shortcuts are trustworthy before they are relied on. If none keeps top-decile recall, we will recommend full training on smaller spaces instead.

Why not tsfresh with XGBoost on the spike data? It is a real threat on small datasets. It cannot answer the circuit-motif question, but if it outperforms every circuit on piriform decoding, we will say that circuit structure is not what limits decoding accuracy in that data.

Decisive Experiments

The first check is whether the detection task shows a flat feedforward advantage across mixture size while the composition task shows a crossover toward normalized or recurrent circuits. This must hold in both the simulator and the Rokni/Mathis data. The second check repeats the analysis with a linear-summation simulator. If the crossover appears only in that version, the recurrence result is an artifact of ignoring receptor antagonism. The third check reports Kendall τ and top-decile recall for each shortcut against the table. It tests whether step-matched budgeted training beats every zero-cost proxy, which we predict. The fourth check applies the best shortcut to the Berners-Lee piriform tetrode data held out by session. It asks whether the circuits it selects keep their rank order across sessions better than LDA, the decoder used in the original study.