Most galaxies in future sky surveys will get their distance estimate, or redshift, from a few broadband brightness measurements rather than from an expensive spectrum. The machine-learning models that make these estimates are trained on galaxies that do have spectra, and those galaxies are brighter and nearer than the ones the models will be used on. The faint, distant galaxies differ in two ways. Their measurements are noisier, and some filters record nothing at all because the galaxy's light has dropped out of that band. Every neural model used so far for this problem treats each unit as a "point neuron": one weighted sum of all inputs passed through one nonlinearity. In that design a single bad band can push the whole sum.
We propose replacing point neurons with a simplified model of a neuron with dendrites, the branching input structures of real neurons. Each branch looks at a few adjacent filters, saturates instead of growing without bound, and is divided down by that filter's measurement uncertainty. This copies a mechanism called shunting inhibition, in which an inhibitory input sitting between an excitatory input and the cell body scales the excitatory signal down instead of subtracting from it. The claim we test is narrow. These units should not beat tree ensembles on galaxies like the training set. They should degrade more gracefully when the redshift range is pushed beyond the training data into fainter, noisier, partly missing photometry.
Core Claim
Two mechanisms from dendritic neuroscience map onto the failure modes of extending photometric redshifts to fainter and more distant galaxies.
The first is saturating, sublinear integration on each branch. In detailed simulations of cerebellar Purkinje cells, Implementing feature binding through dendritic networks of a single neuron finds up to 49% sublinear summation on thin distal branches. As a result, clustered input on one branch is suppressed, and the cell needs evidence spread across several branches before it fires. That paper uses the effect to explain single-cell feature binding. We use it as a structural guard: a galaxy's redshift estimate should not be dominated by one corrupted band.
The second is the reduced shunting model of Nonlinear multiplicative dendritic integration in neuron and network models, which collapses a three-compartment neuron into a single unit whose output contains a multiplicative excitation-times-inhibition term. That paper uses the term to stabilize recurrent spiking networks at realistic firing rates. We repurpose it so that the "inhibitory" input to each branch is the measured flux uncertainty in that branch's bands.
The resulting unit, applied to branch $b$ with input band subset $S_b$, is
$$d_b = g\!\left(\frac{w_b^\top x_{S_b}}{1+\kappa_b\, u_b^\top \sigma_{S_b}}\right), \qquad y = \sum_b a_b\, d_b,$$
where the symbols are defined as follows. The vector $x_{S_b}$ holds the fluxes or colors of the contiguous filters assigned to branch $b$, and $\sigma_{S_b}$ holds their reported uncertainties. The weights $w_b$ and $u_b$ are learned, with $u_b$ constrained to be nonnegative. The shunting strength $\kappa_b \ge 0$ is also learned. The function $g$ is a bounded saturating nonlinearity, here a scaled tanh, and $a_b$ are the weights that combine branch outputs at the "soma," the cell body. Layers of such units feed a conditional density head that outputs a binned redshift distribution, because high-redshift color degeneracies make Gaussian outputs inappropriate (Improving Generalization and Uncertainty Quantification of Photometric Redshift Models).
The unit differs from the closest existing dendritic network in three ways. Improved Expressivity Through Dendritic Neural Networks (DENN) partitions inputs into random sparse branches combined by a max operation, and evaluates them only on generic benchmarks such as 121 UCI datasets. Our branches are tied to adjacent filters in wavelength order rather than drawn at random. They saturate instead of taking a max. And they are gated by measurement uncertainty, which DENN has no slot for. To our knowledge, no work in the research file applies any dendritic model to astronomical data.
Research Program
Unit and training. We will implement the unit as dense block tensors. Because branch subsets are fixed and contiguous, the computation reduces to batched small matrix multiplies, in the spirit of DENN's intragroup sparsity. The network will have two or three dendritic layers. Training will use Adam, since DENN reports that sparse branch structure trains poorly with plain stochastic gradient descent. The loss is cross-entropy over redshift bins.
Datasets and shift splits. We will run on three datasets, each with an explicit "range extension" split.
The first is the tabular SDSS DR12 magnitude catalog used by Benchmarking and scalability of machine-learning methods for photometric redshift estimation. We train on the brighter half in $r$-band magnitude and the lower redshift range, then test on the fainter, higher-redshift tail.
The second is the GalaxiesML spectroscopic set (HSC-PDR2) for training, with transfer to the fainter TransferZ set (COSMOS2020) for testing, following Improving Generalization and Uncertainty Quantification of Photometric Redshift Models. That paper warns that TransferZ labels may carry errors 1.5–2 times larger than catalogued, so we report label-error-aware metrics as well.
The third is the 9-band GALEX/PanSTARRS/unWISE spectroscopic sample of Mantis Shrimp, using catalog fluxes and errors rather than image cutouts. We train on galaxies below a redshift cut and test on higher-redshift spectroscopic targets, including DESI early-data objects. In this split the 4000 Å break has moved into the reddest optical and infrared bands, and UV dropouts are common.
Analysis of what the branches learn. For each trained model we will measure the learned shunting strengths against signal-to-noise. We will also measure how much the predicted distribution moves when a single band is perturbed, compared with a coherent shift across several bands. This directly tests whether the scatter-sensitivity reported for Purkinje cells carries over to the trained network.
Baselines
Naive reference point. The do-nothing reference is the standard point-neuron multilayer perceptron on magnitudes and colors, trained only on the bright split, as benchmarked by Henghes et al. (2021). Its in-distribution accuracy is reported only as worse than MSE 0.0046 on SDSS DR12. It anchors how badly an ordinary network degrades when pushed beyond its training range.
Strongest in-distribution competitor. This is Random Forest, together with gradient-boosted trees. Henghes et al. (2021) report Random Forest as the best tabular method on SDSS DR12, with MSE = 0.0042, and kNN, boosted trees and extremely randomized trees all below 0.0046. No reported result was found on Lacuna for LightGBM or XGBoost photo-z. We will run boosted trees anyway, because a reviewer would. All tree models get the same inputs, including flux uncertainties as features, the same splits, and an equal hyperparameter search budget of the same number of trials on the same validation fold. Since trees predict piecewise-constant values bounded by the training targets, we expect them to win in distribution and lose on the extension split. Beating them means lower outlier fraction (share of galaxies with $|\Delta z|/(1+z)>0.15$) and lower bias on the extension split, with the improvement larger than the spread across five seeds.
Strongest existing domain-shift method. This is NN-Combo, the composite-training network of Improving Generalization and Uncertainty Quantification of Photometric Redshift Models. It reports 4.5 times lower bias and 1.1 times lower scatter over 0.3 < z < 1.5 than a spectroscopic-only model, and meets one of three LSST requirements on TransferZ. Composite training changes the data rather than the architecture, so the two approaches can be stacked. The fair comparison is therefore our unit against their point-neuron network under both spectroscopic-only and composite training. A win means our unit improves TransferZ bias or outlier rate under both training regimes.
Image-based reference. The image-based reference is the early-fusion ConvNeXt of Mantis Shrimp. On its own test set it reports NMAD 0.0244, bias 0.0115 and outlier rate 4.51%. It sees pixels, which our model does not, so we treat it as an upper reference rather than a competitor. Approaching it from catalog fluxes alone, without image cutouts, would be a meaningful practical result.
Generic dendritic baseline. DENN with random branches and max integration is matched in parameter count to our model.
Ablations of the key ingredients. We run four ablations at matched parameter count. The first removes shunting and feeds uncertainties as ordinary additive inputs. The second makes branches linear, with no saturation. The third replaces wavelength-contiguous branches with random ones. The fourth applies the shunting gate to a dense layer with no branch structure.
Why not just…?
Why not just use gradient-boosted trees, which already win on tabular photometry? On galaxies resembling the training set, they probably will win, and we expect to report that. The proposal's claim lives only on the extension split. There, trees cannot predict redshifts outside their training targets, and their splits on a band that becomes noise at faint magnitudes have no built-in down-weighting. Henghes et al. also report Random Forest becoming the slowest method to train at a million galaxies. If boosted trees match us on the extension split, the claim fails, and the head-to-head comparison above settles it.
Why not just give a standard network the uncertainties and train with noise and band-dropout augmentation? This is the most serious threat. Such a network could learn uncertainty-aware down-weighting from data alone. We will run it as a dedicated baseline: the same multilayer perceptron, with $\sigma$ inputs, augmentation matched to the faint-split noise distribution, and the same tuning budget. If it matches the dendritic unit, then the structure is unnecessary and augmentation is the real lesson. We would report exactly that. The no-shunting ablation run with augmentation separates the two effects.
Why not just use a generic gated or multiplicative layer, since the shunting term is just a product? Algebraically the gate is a learned division, and a gated layer can represent it. The question is whether tying the gate to measurement uncertainty, within wavelength-local saturating branches, helps when data near the redshift frontier are scarce. The dense-gated ablation tests precisely this. If it matches the full model, the branch structure contributes nothing and the dendritic framing should be dropped.
Decisive Checks
Two checks decide whether the idea holds. In the first, every model is trained on the Mantis Shrimp catalog fluxes below the redshift cut. On higher-redshift spectroscopic galaxies, the full dendritic model must beat both the augmented multilayer perceptron and boosted trees on outlier rate and bias, by more than the variation across seeds. Removing either saturation or uncertainty shunting must erase most of that gap.
In the second, we corrupt one band of test galaxies with noise at the faint-split level. The full model's predicted distribution should shift measurably less than the point-neuron baselines'. When all bands are shifted coherently, as a real change in redshift would do, its predictions should still change normally. Failing either check would mean the Purkinje-cell mechanism does not carry over to this problem.
References
- Implementing feature binding through dendritic networks of a single neuron
- Nonlinear multiplicative dendritic integration in neuron and network models
- Improving Generalization and Uncertainty Quantification of Photometric Redshift Models
- Improved Expressivity Through Dendritic Neural Networks
- Benchmarking and scalability of machine-learning methods for photometric redshift estimation
- Mantis Shrimp: Exploring Photometric Band Utilization in Computer Vision Networks for Photometric Redshift Estimation