Dynamic trajectory cues drive sequenced integration in approach detectors

TL;DR: Change nothing about an object's size or position — just modulate its brightness with a slow-rising sawtooth — and both humans and fruit flies perceive it as approaching. The fly's classic "loom detector" neurons (LPLC1, LPLC2) turn out to mediate this luminance-driven percept too, making them general approach detectors rather than expansion detectors. And the punchline: these neurons combine luminance and expansion cues supralinearly only when luminance change comes first — which, by simple geometry, is the order these cues arrive during any real approach. This is a clean example of a circuit tuned not to a feature but to the temporal script of a natural event.

The problem: "looming" is only one page of the script

Every sighted animal needs to detect things coming at it — predators, collisions — and forty years of neuroscience has converged on a canonical answer for how: loom detection. An approaching object grows in angular size on the retina, and neurons selective for expansion (locust LGMD, fly LC4/LPLC2, mouse superior colliculus, zebrafish tectum, pigeon nucleus rotundus) trigger escape. Expansion has so thoroughly dominated the field that "loom detector" and "approach detector" are used interchangeably.

But approach is a physical process, and physics hands the observer a whole bundle of correlated cues, of which expansion is just one. As an object closes in, it also progressively occludes the background and replaces it with its own luminance — so the total light arriving from that patch of the visual field changes systematically. Human psychophysics has long hinted this matters (brighter and higher-contrast things look nearer), but nobody had asked whether luminance change alone can drive an approach percept, or how a real circuit combines it with expansion.

There's also a resolution problem the expansion-only story glosses over. A distant object may subtend less than the eye's spatial resolution (~5° for a fly ommatidium). Its edges can't drive motion detectors at all — expansion is literally invisible early in the approach. Integrated luminance over a receptive field, however, changes detectably even when the object is a sub-resolution speck. So if you only listen to expansion, you find out late.

An illusion that both species fall for

The authors adapted a stimulus from Weiss et al. (2004): a 6×6 grid of circles whose sizes and positions never change. Each circle's luminance follows a sawtooth wave (0.5 s period) with a random starting phase, so the display averages to uniform gray with no net motion. The only choice is the sawtooth's shape: slow rise, fast reset ("ramp-up") or slow fall, fast reset ("ramp-down").

In a two-alternative forced-choice task, humans reliably reported ramp-up as approaching and ramp-down as retreating. The authors then built a continuum parameterized by \Phi, the fraction of each cycle spent slowly brightening (\Phi{=}1 is pure ramp-up, \Phi{=}0 pure ramp-down, \Phi{=}0.5 a symmetric triangle wave). Approach reports rose sigmoidally with \Phi and sat at chance for the symmetric wave — the percept lives entirely in the asymmetry of luminance increments vs. decrements. Note that ramp-up and ramp-down have luminance derivatives of equal magnitude and opposite sign averaging to zero, so any linear derivative-based mechanism is blind to the difference; the percept must come from asymmetric ON/OFF processing, the same kind of imbalance behind the peripheral-drift motion illusion.

Then the same stimuli went to tethered walking flies on an air-supported ball. Flies treat expanding discs as threats: they turn away from lateral looms and freeze to frontal ones, while contracting discs draw them in. The luminance grids reproduced this pattern exactly — ramp-up drove turning-away and robust freezing like an expanding disc; ramp-down drove turning-toward and minimal freezing like a contracting one. Sweeping \Phi produced a fly behavioral curve closely matching the human psychometric curve, with near-zero net turning at \Phi = 0.5. Two visual systems with radically different architectures, separated by ~600 million years, fall for the same illusion in the same graded way.

From percept to circuit

Behavioral mimicry could in principle reflect some arbitrary spatiotemporal preference rather than an approach percept. The fly lets you test this causally. Using shibire^ts to reversibly block synaptic output, the authors silenced LPLC1 and LPLC2 — two well-characterized loom-selective visual projection neuron types — and, as a control, LC11 (a small-object detector that ignores looms).

Silencing LPLC1 or LPLC2 reduced freezing to expanding discs (expected) and to ramp-up grids (the new result), relative to genetic controls. Silencing LC11 affected nothing. Turning showed a consistent picture: LPLC1 silencing reduced turning-away from ramp-up and, interestingly, enhanced turning-toward ramp-down; LPLC2 silencing selectively impaired turning-away from ramp-up. So the luminance illusion isn't routed through some separate luminance pathway — it flows through the very neurons the field calls loom detectors. Two-photon GCaMP6f imaging closed the loop: both cell types responded more strongly to ramp-up than ramp-down grids across all sawtooth periods tested, mirroring behavior. These neurons are approach detectors, not expansion detectors, and "loom detector" has been a description of the stimuli we happened to test them with.

Geometry says luminance comes first

Here is the paper's conceptual core. Take a disc of radius R approaching at constant speed v, at distance D(t). Its angular half-size is \theta \approx R/D, so the expansion rate is \dot\theta \approx Rv/D^2. The irradiance at the eye from the disc's luminance goes as the solid angle, \propto R^2/D^2, so its rate of change goes as R^2 v/D^3. Total motion signal — expansion rate summed around the circumference, which grows like 1/D — also scales as R^2 v/D^3. So asymptotically the two cues grow at the same rate.

The asymmetry is a resolution threshold, not a scaling law. Expansion is measured by motion detectors that compare neighboring photoreceptors ~5° apart; until the object subtends several ommatidia, there simply are no edges to move across detectors. Luminance integrated over a receptive field has no such floor — it changes detectably while the object is still an unresolvable dot, as long as its contrast \Delta L against the background is large enough. The authors confirm this with a simulation: projecting an approaching object through a minimal synaptic model of the fly's T4/T5 motion detectors, expansion crosses detection threshold only once the object is substantially larger than one ommatidium, whereas the integrated luminance signal is informative throughout.

This ordering isn't a species-specific adaptation — it's a geometric inevitability for any eye with finite resolution. Which sets up the natural hypothesis: if LPLC1/2 are genuinely tuned to approach as an event, they should care about cue order.

Sequenced integration: order-dependent synergy

To test this, the authors needed the cues decoupled. For pure expansion, they built an isoluminant expanding annulus — a checkerboard-patterned ring (inner/outer radii 5°/15° growing to 20°/30° at 30°/s) whose average luminance never changes. For pure luminance, a 5° dot at the neuron's receptive-field center stepping gray→black or gray→white. Both cell types respond to each cue alone (with the familiar bias toward decrements — dark things approaching matter more).

Then the joint presentations, with the luminance step offset from expansion onset anywhere from −500 ms (before) to +500 ms (after):

• Luminance before expansion: the response exceeded the linear sum of the two isolated responses — supralinear, synergistic integration. This held for both gray→black and gray→white steps, in both LPLC1 and LPLC2.
• Luminance after expansion: the joint response was sublinear or merely additive — sometimes actively suppressed.

A crucial control is built into the data: if the supralinearity were an artifact of GCaMP's nonlinearity, you'd expect a consistent distortion regardless of cue order. Observing supralinear summation in one temporal order and sublinear in the other, in the same neurons, argues for genuine stimulus-dependent amplification and suppression.

The behavioral echo: flies shown the isoluminant annulus plus a luminance step at ±250 ms offsets turned away and slowed significantly more when the luminance step preceded expansion than for expansion alone or the reversed order. The order-sensitivity isn't just a calcium-imaging curiosity; it reaches motor output.

How strong is this, really?

Strong points. The percept transfers across species with matched parametric behavior (the \Phi curve). The silencing experiments are genuinely causal for the claim that LPLC1/2 mediate luminance-grid responses, with a sensible negative control (LC11) and the standard empty-Gal4 genetic controls. The imaging, silencing, and behavior triangulate consistently. And the geometric argument (worked out in a short, readable appendix) turns what could have been a phenomenological temporal-order effect into a prediction derived from first principles.

Caveats worth holding onto. This is a preprint, not yet peer-reviewed, and the extracted text reports statistics only by reference to figures — effect sizes for the synergy aren't quotable here. The "sequenced integration" claim at the neural level is an imaging result: nobody has yet silenced LPLC1/2 during the joint-cue behavioral task to show the behavioral order-effect flows specifically through these cells, so the neural synergy and the behavioral synergy are linked by plausibility rather than direct causation. GCaMP6f at ~8.5 Hz frame rates is slow relative to the ±250–500 ms offsets in play, so the fine temporal structure of the interaction is coarse-grained. The mechanism of the amplification — presynaptic gain change, dendritic nonlinearity, disinhibition — is entirely open. And the human side is purely psychophysical; the claim of a shared circuit strategy across phyla is an inference, not a measurement.

One more subtlety the authors handle honestly: the grid illusion depends on luminance asymmetry (increments vs. decrements), yet in the joint-cue experiments both polarities of luminance step enhanced expansion responses. The illusion likely arises from ON/OFF-imbalanced spatiotemporal integration upstream, which is related to, but not identical to, the synergy mechanism in LPLC1/2. The paper links the percept to the neurons causally, but the full computation producing the illusion isn't solved.

Why this matters beyond flies

The framing shift is the real contribution: stop thinking of approach detectors as feature detectors (for expansion) and start thinking of them as event detectors tuned to the stereotyped temporal script that physics writes for approaching objects. The authors draw the right parallels — audiovisual integration windows that mirror light preceding sound, Reichardt-style motion detectors that amplify one spatiotemporal sequence and suppress its reverse. Sequenced cue integration may be a general motif: wherever nature imposes a fixed cue order, circuits can buy sensitivity and specificity simultaneously by gating one cue's gain on another's recent arrival.

For an ML reader, there's a design lesson here. A system trained or built to detect events from video shouldn't treat correlated cues as an unordered feature bag; the order of cue onset carries discriminative information that a physics-aware prior can exploit, especially near sensory resolution limits where the strongest cue hasn't arrived yet. The fly gets early warning from a cheap, low-resolution signal (integrated brightness) and uses it to prime a more specific, expensive one (expansion) — a cascade with a temporal validity check baked in.

If you read one part of the paper, make it the "Natural sequencing enhances responses" section with Figure 6, then the two-page Appendix — the 1/D^2 vs. 1/D^3 derivation is where the whole logic of the paper clicks into place.