Abstract
Learning to detect content-independent transformations from data is one of the central problems in biological and artificial intelligence. An example of such problem is unsupervised learning of a visual motion detector from pairs of consecutive video frames. Rao and Ruderman formulated this problem in terms of learning infinitesimal transformation operators (Lie group generators) via minimizing image reconstruction error. Unfortunately, it is difficult to map their model onto a biologically plausible neural network (NN) with local learning rules. Here we propose a biologically plausible model of motion detection. We also adopt the transformation-operator approach but, instead of reconstruction-error minimization, start with a similarity-preserving objective function. An online algorithm that optimizes such an objective function naturally maps onto an NN with biologically plausible learning rules. The trained NN recapitulates major features of the well-studied motion detector in the fly. In particular, it is consistent with the experimental observation that local motion detectors combine information from at least three adjacent pixels, something that contradicts the celebrated Hassenstein-Reichardt model.
Introduction
A fundamental challenge in both artificial and biological intelligence is learning "content-independent transformations"—the ability to detect motion regardless of the specific object being moved. Previous computational models, notably by Rao and Ruderman, framed this as an unsupervised learning problem of finding mathematical "generators" (Lie group operators) to explain changes between consecutive frames. However, these models relied on reconstruction-error minimization, which typically requires "non-local" learning rules. In biological systems, synapses only have access to local information, making reconstruction-based models biologically implausible.
Methods: Similarity Matching and Local Rules
This study proposes a shift from reconstruction to similarity preservation. The core objective is Similarity Matching (SM): if two pairs of image frames represent similar motions, the network's output should also be similar.
Optimizing this similarity-preserving objective leads to an online algorithm that maps onto a single-layer neural network using biologically plausible Hebbian and anti-Hebbian learning rules. The authors also explore Non-negative Similarity Matching (NSM) to satisfy the biological constraint of non-negative neural firing rates.
The network identifies that motion detection hinges on the outer product of the change in pixel intensity ($\Delta x$) and the intensity itself ($x$), effectively performing a correlation between temporal change and spatial context. Validation was performed using 1D and 2D synthetic datasets, as well as periodic gratings common in fly vision experiments.
Results: Recapitulating the Fly Visual Circuit
The trained network naturally evolves into a circuit that mimics the fly's visual system. A key finding is the emergence of a "three-pixel" detector. While the classic Hassenstein-Reichardt model uses two pixels, recent connectomics data from Drosophila shows that T4/T5 motion-sensing neurons sample from at least three points in space.
The model demonstrates that while local 3-pixel detectors differ from the 2-pixel model, they are mathematically equivalent when summed across the entire visual field (global motion). This explains why previous behavioral experiments could not distinguish between them. Furthermore, the NSM version successfully separates opposite directions of motion into distinct neural channels, mirroring the ON/OFF pathways in biological vision.
Discussion
The proposed framework suggests that complex motion detection can occur within a single neuron's dendritic tree using mechanisms like shunting inhibition. While the model currently relies on infinitesimal transformation operators (Lie algebra)—making it best suited for smooth, small displacements—it provides a principled path for reverse-engineering how the brain performs advanced tasks like object recognition across varying viewpoints.
Summary updated