Branching dendritic trees and motoneuron membrane resistivity

TL;DR. In 1959 Wilfrid Rall used cable theory to show that a motoneuron's dendritic tree, not its soma, dominates how the cell integrates input — turning the neuron from a point summator into a spatially extended analog computer, and exposing a systematic error in the era's most authoritative electrophysiology. He could not simulate realistic trees (no computers up to it) and nobody could record from a dendrite to check him. Both barriers are gone: dendritic patch clamp and imaging have revealed local nonlinear "subunits" inside single cells (NMDA spikes, XOR in human dendrites; one cortical neuron ≈ a 5–8 layer temporal CNN), and dendrite-inspired architectures with local nonlinearities and local credit assignment are one of the more plausible unexploited routes to parameter- and energy-efficient learning. This is the founding paper of that whole line.

The neuron Rall was arguing against

The state of the art in 1959 was Eccles' lab in Canberra: intracellular sharp-electrode recordings from cat spinal motoneurons, the first quantitative biophysics of a mammalian central neuron. Coombs, Eccles, and Fatt had measured whole-cell input resistance (order 1 MΩ) and, treating the neuron essentially as an isopotential sphere — the soma, with dendrites as passive appendages that barely mattered electrically — derived a membrane resistivity of very roughly 500 Ω·cm² and short synaptic-potential time courses.

Rall's 1959 paper is, on its surface, a correction of that estimate. He worked out the steady-state input conductance of a soma with realistic branching dendrites attached, using anatomical measurements of motoneuron trees, and showed the dendrites are not appendages: they are the majority of the membrane and they sink most of the injected current. His dendritic-to-somatic conductance ratio came out well above 1 (severalfold — I won't pretend to remember his exact figure), which meant the soma-only analysis had underestimated membrane resistivity several-fold and, downstream, mischaracterized synaptic integration. When Rall and Eccles disagreed about the decay of the EPSP in the early 1960s, Rall turned out to be right.

But the correction is the small idea. The big idea is the method.

The idea: a neuron is a cable network you can actually solve

Passive neural membrane obeys the one-dimensional cable equation. For a cylinder of diameter d, voltage attenuates with a space constant

where R_m is specific membrane resistivity (Ω·cm²) and R_i is axial resistivity (Ω·cm). Fine: but a dendritic tree is dozens of coupled cylinders with branch points, and solving that by hand in 1959 was hopeless.

Rall's move was to find the condition under which the whole tree collapses into a single equivalent cylinder. At a branch point, if the daughter branches satisfy

then the impedance looking into the daughters matches the impedance of a continued parent cable, and — applied recursively — the entire tree becomes analytically equivalent to one cable of finite electrotonic length. The exponent 3/2 falls out of how input conductance of a semi-infinite cable scales with diameter (G \propto d^{3/2}: cross-section gives d^2, membrane leak per length gives d, geometric mean). Remarkably, real motoneuron trees approximately obey this rule, which is why the trick had teeth rather than being a mathematician's convenience.

Once you can solve the tree, the conceptual conclusion is unavoidable: where a synapse sits matters. Distal and proximal inputs are filtered differently, inputs interact sublinearly on the same branch, and the temporal order of inputs along a dendrite shapes the somatic response. The neuron is not a summing junction; it is a spatially distributed device whose geometry is part of its computation. That framing — implicit in 1959, explicit in Rall's 1960s work — is the paper's real payload.

Why it could not be pushed further in 1959

Compute. The equivalent cylinder is exact only under the symmetry assumptions; real questions (asymmetric trees, nonuniform membrane, synapses at specific loci, and above all active dendritic conductances) require compartmental simulation — discretize the tree into hundreds of coupled ODEs and integrate. Rall himself pioneered compartmental modeling around 1964, but on the machines of the day it was agony. An IBM 704/7090-class machine delivered roughly 10^4–10^5 operations per second; a realistic Hodgkin–Huxley-style tree with ~1,000 compartments needs on the order of 10^6–10^7 operations per simulated millisecond. A modern GPU does 10^{14}–10^{15} FLOP/s — roughly ten orders of magnitude more — and simulates thousands of morphologically detailed neurons in real time.

Measurement. More fatally, the theory was untestable at its point of action. Sharp microelectrodes could impale a 30–70 µm soma; a dendrite 1–3 µm thick was untouchable. There was no way to record voltage in a dendrite, no calcium or voltage imaging, no way to place a synaptic input at a known dendritic location. Every prediction about dendritic integration had to be inferred through the soma, which is exactly the low-pass-filtered, geometry-confounded viewpoint Rall was warning about. Even the anatomy was shaky: dendritic diameters came from Golgi and Nissl material with real shrinkage and sampling artifacts.

So the paper landed as a technical correction to Eccles and a piece of applied mathematics, and the "neurons compute in their dendrites" reading had to wait decades for instruments.

What changed

Dendrites became recordable. Stuart and Sakmann's dendritic patch recordings (1994) showed action potentials actively backpropagating into dendrites — dendrites are not passive cable at all. Schiller and colleagues (2000) found NMDA spikes: local, regenerative, branch-confined plateau events, meaning each thin basal branch is a sigmoidal nonlinear subunit. Larkum's work established calcium spikes in the apical tuft as a coincidence-detection mechanism coupling feedforward and feedback streams. Gidon et al. (2020, Science) reported a calcium-mediated dendritic event in human layer 2/3 pyramidal cells whose graded, non-monotonic activation lets a single dendrite compute XOR — the canonical function a point neuron cannot.

The point-neuron abstraction was quantified as wrong. Poirazi and Mel modeled a pyramidal neuron as a two-layer network of sigmoidal branch subunits (2003). Beniaguev, Segev, and London (2021) went further: fitting the full I/O of a detailed layer-5 pyramidal cell model required a temporal convolutional network 5–8 layers deep. A "neuron" in the connectionist sense underestimates a biological neuron by a small network's worth of depth — and most of that depth comes from NMDA-dependent dendritic nonlinearity, i.e., from the physics Rall put on the table.

Dendrites became an ML design pattern. Two-compartment neuron models (Urbanczik & Senn 2014; Sacramento et al. 2018; Guerguiev et al. 2017; Payeur et al. 2021) use the apical/basal segregation to carry error or feedback signals separately from feedforward drive, giving biologically plausible, local approximations to backprop. Numenta-style "active dendrite" networks use dendritic gating to reduce catastrophic interference in multitask learning. Chavlis and Poirazi (2025) showed dendritic ANNs — sparse, structured receptive-field-like connectivity into nonlinear "dendrite" units — match conventional MLPs with substantially fewer trainable parameters and better robustness. Neuromorphic hardware (Intel's Loihi supports multi-compartment neurons; several academic chips target dendritic plateaus explicitly) can implement these at synapse-local energy costs.

What a serious 2026 revival looks like

Keep from Rall: the commitment that placement is computation — that a tree of coupled nonlinear compartments is a different (and cheaper) function class than a wide flat sum — and the discipline of deriving what the geometry buys you rather than gesturing at it.

Replace: passive membrane with the modern two-nonlinearity template (branch-local NMDA-like sigmoids plus an apical plateau/gating event), and analytic collapse with learned structure.

Concretely, three experiments worth running:

1. Scaling laws for dendritic ANNs. Take dendritic units (k sparse branches → saturating nonlinearity → soma) as a drop-in replacement for MLP blocks in a transformer, and measure loss-versus-parameters and loss-versus-energy scaling against standard blocks up to the ~1B scale. The Poirazi-lab results are at small scale; nobody has shown whether the parameter-efficiency advantage survives scaling or is absorbed by optimization. Either outcome is informative.

2. Plateau-gated local learning on neuromorphic hardware. Two-compartment neurons where an apical "plateau" event gates a three-factor synapse-local plasticity rule (the burst-dependent scheme of Payeur et al. is the natural candidate), implemented on Loihi-class silicon. Benchmark: online continual learning at fixed energy budget against backprop on GPU. This is the setting where locality is not a handicap but the whole point.

3. Closing Beniaguev's loop. If one neuron ≈ an 8-layer TCN, then a network of N dendritic neurons should match a point-neuron network of \sim kN units for some k \gg 1 on temporally structured tasks. Measure k on speech and event-camera streams, where the dendritic delay-line and coincidence machinery Rall analyzed (his later work on synaptic sequence discrimination) should matter most.

Status: vindicated, descendants, still open

The biophysics is fully vindicated — Rall's framework is modern dendritic modeling; NEURON and Arbor are industrial-strength descendants of his compartmental method, and the equivalent-cylinder analysis is textbook material (Segev, Rinzel & Shepherd collected the papers in The Theoretical Foundation of Dendritic Function). The computational reading is empirically supported at the single-cell level (NMDA spikes, human dendritic XOR, deep-network-equivalence) and increasingly in vivo (dendritic plateaus correlate with place-field formation and perceptual detection).

What is open is the part your field cares about: whether dendritic structure is a load-bearing efficiency mechanism at scale, or a biological workaround that gradient descent on homogeneous hardware simply doesn't need. The parameter-efficiency and continual-learning results are encouraging but small; no dendritic architecture has yet won at frontier scale, and no neuromorphic dendritic chip has beaten GPUs on a task anyone was forced to care about. That is exactly the shape of an underexploited idea rather than a refuted one.

Where to read it

The paper: doi.org/10.1016/0014-4886(59)90046-9 (bibliographic details verified; my characterization of specific numerical values inside it is from memory of the secondary literature and should be checked against the text). Read alongside: Rall's 1964 compartmental-model chapter and his 1967 sequence-discrimination paper; Stuart, Spruston & Häusser's Dendrites for the experimental era; Poirazi & Mel (2003), Gidon et al. (2020), and Beniaguev et al. (2021) for the single-neuron-as-network results; and Sacramento et al. (2018) plus Chavlis & Poirazi (2025) for what the revival is already starting to look like.