INME, a circuit using only four cardinal-direction cells, reproduces eleven motion phenomena while million-parameter optical-flow networks miss most
Synopsis
Motion direction estimation has lacked a unified model spanning species and brain regions, since existing models tend to fit one animal class, one brain area or one phenomenon. INME uses only cells selective to the four cardinal directions, embedding dendritic nonlinearity and a back-propagating action potential in one circuit with a post-summation readout stage. Across eleven synthetic-stimulus experiments it qualitatively reproduces phenomena from contrast saturation and reverse-Phi to the breaking of time-reversal symmetry, while five deep optical-flow networks with millions of parameters reproduce only a few of them.
Interpretation
A neural circuit model relying only on cells selective to the four cardinal directions can qualitatively cover motion phenomena ranging from basic to previously unexplained. Earlier motion-direction models typically applied to a specific animal, brain region or phenomenon: a primate-oriented model does not transfer directly to insects, a model developed for MT fails in V1 and the superior colliculus, and second-order motion models do not predict first-order motion. Across eleven synthetic-stimulus experiments the model reproduced the missing-fundamental illusion, reverse-Phi, contrast saturation, motion masking, first- and second-order motion, direction for gratings and Type I and Type II plaids, interaction among scales, the breaking of time-reversal symmetry and three static motion illusions; five deep optical-flow networks reproduced only a few phenomena under the same stimuli and none reproduced interaction among scales, time-reversal symmetry breaking or the three static illusions.
The model has three stages: linear spatio-temporal filtering with pointwise nonlinearities at the first and last stages, and a highly nonlinear motion sensor in the middle. The middle stage extends an existing one-dimensional motion sensor model by letting the nonlinearity take linear and nonlinear transforms of the stimulus as arguments. In the mechanistic reading, the linear spatial terms correspond to passive dendrites, the middle-stage nonlinearity to active-dendrite responses, and its subtracting term to a back-propagating action potential; outputs are projections onto the four cardinal directions, combined in pairs into horizontal and vertical motion-opponent signals from which direction and magnitude are read out.
Estimated motion direction tracks the true direction closely for gratings and plaids. Previously such direction readout required changing model structure or parameters between phenomena. For drifting sinusoidal gratings the mean absolute error between predicted and true direction is 1.2 degrees; for Type I plaids it is 0.2 degrees; for Type II plaids, where vector average and intersection-of-constraints directions diverge, the model follows the intersection-of-constraints direction with a mean absolute error of 2.3 degrees.
Perspective
The results rest on eleven synthetic visual-stimulus experiments, compared against five deep optical-flow networks run with their official public implementations and pretrained weights without fine-tuning; the authors state the aim is qualitative reproduction of phenomena rather than precise quantitative matching.
The authors do not claim the circuit covers every aspect of motion direction estimation, and other neural computations may contribute or produce similar outcomes through a different mechanism; mammals have two parallel motion pathways from retina to cortex, and whether direction selectivity computed de novo is combined with retinal direction selectivity at some stage remains to be established; in interaction among scales the model's qualitative response is completely flat when there is no inversion, which does not match the experimental data; some basic phenomena share one parameter set, while more challenging experiments required different parameter values.