Accurate Path Integration in Continuous Attractor Network Models of Grid Cells

If your brain integrates velocity signals to track your position in space, and if that integration runs for ten minutes without a landmark to correct it, the accumulated error should be enormous — the math demands it. Small biases compound. Noise drifts. Any realistic neural signal has variability. And yet, rats navigate home after long, winding foraging trips without stopping to check a map. Something in the brain is doing dead reckoning, and doing it well. Burak and Fiete showed that a network of a few thousand neurons, wired with the right geometry, can keep position error small enough to navigate a football field without a single external cue to reset it. That number — one hundred meters and ten minutes — is the engine of the whole lecture. Grid cells are the neurons at the center of this story. They live in the dorsolateral band of the medial entorhinal cortex, and their defining property is almost absurdly geometric: each cell fires whenever the rat occupies any vertex of a regular triangular lattice that tiles the entire environment. It's a hexagonal grid, stamped invisibly across the floor, and the cell lights up at every intersection. The hypothesis that follows almost naturally is that this spatial periodicity reflects an internal path integration system — the network receives velocity and heading inputs, integrates them over time, and uses that running sum to maintain a continuously updated position estimate, even in the dark and even without landmarks. The behavioral evidence for this capacity is real. Rats can compute straight return paths after random foraging trajectories of one to three meters without any external cues. Physiological recordings show that grid cell tuning remains sharp and stable over trajectories lasting tens of minutes and hundreds of meters. That kind of stability places enormous demands on any model that claims to explain it. Existing models haven't met those demands. Oscillatory interference models, where single cell dynamics create periodic spatial responses through interference between theta frequency oscillations, are acutely vulnerable to phase noise. The phase of theta in the entorhinal cortex typically slips by half a cycle in less than ten cycles, or about one second. At one meter per second, that's roughly one meter of travel before the code for position degrades. Earlier continuous attractor network models did better conceptually, but still failed in practice: pattern rotations and nonlinear velocity responses caused errors to compound quickly, and the models needed frequent resets by sensory cues to stay coherent. That's not dead reckoning — that's dead reckoning with training wheels. A continuous attractor network is a recurrent network whose connectivity creates a stable, low-dimensional surface of possible activity states. Think of it as a landscape with a long, flat valley: the system can slide along the valley freely, but it resists being pushed off to the sides. In Burak and Fiete's model, that valley corresponds to translations of a population activity pattern — a triangular lattice of activity bumps distributed across a two-dimensional sheet of neurons. The lattice is produced by center-surround recurrent connectivity: each neuron inhibits activity in a surrounding ring while supporting a central peak. With the right inhibitory weight profile, this is enough to spontaneously generate a spatially periodic, triangular pattern across the network, and individual neurons inherit grid cell-like firing fields directly from that population pattern. Path integration is implemented by coupling velocity-sensitive feedforward inputs to this attractor. Each neuron is assigned a preferred heading — north, south, east, or west — and outgoing weights are shifted slightly in that preferred direction. When the animal moves, feedforward inputs carrying velocity projected along each neuron's preferred heading bias the recurrent interactions, causing the entire lattice pattern to flow across the sheet in the direction and at the speed of actual movement. The key parameter is the product of the weight shift and the velocity gain: get that right, and the pattern translates smoothly. The neural time constant in the model is ten milliseconds, and networks in the reported simulations typically contained around sixteen thousand neurons — within the biologically plausible range of ten thousand to one hundred thousand neurons estimated for the relevant entorhinal circuits. One architectural choice turns out to matter enormously: boundary conditions. In a periodic network — effectively a torus, where the edges wrap around — the attractor manifold supports pure translations and resists rotations. The lattice slides cleanly. In an aperiodic network, which has actual edges, boundary effects distort the lattice and introduce rotational modes that can destabilize the pattern. Burak and Fiete found that smoothly tapering feedforward inputs near the edges helps, but aperiodic networks remain more fragile and generally require larger sizes to perform comparably to periodic ones. Now for the results that make this more than a theoretical exercise. Burak and Fiete ran a deterministic periodic network of 128 by 128 neurons — about sixteen thousand neurons — on a recorded rat trajectory spanning 260 meters and 20 minutes. The network's position estimate differed from the rat's true position by about 15 centimeters total, against a grid period of 48 centimeters. That's an average integration error of less than a tenth of a centimeter per meter traveled. A much smaller periodic network, 40 by 40 neurons, produced coherent single neuron grids for the same trajectory. The conclusion is that periodic boundary conditions make accurate integration dramatically easier to achieve, and even modest network sizes can do the job. Aperiodic networks can succeed too, but with caveats. The most significant failure mode is pinning: at low speeds, below roughly ten centimeters per second in some of their examples, the lattice pattern simply fails to translate. The energy landscape isn't flat enough. Increasing network size reduces this problem, but they found they couldn't build an aperiodic network substantially smaller than about ten thousand neurons that performed comparably to the periodic case. Spiking variability is the other major constraint. In their stochastic simulations, noise produces diffusive drift of the population's phase and orientation. The characteristic coherence time — the time over which the network maintains a coherent grid — scales with the grid period squared, with network size, and inversely with the coefficient of variation of spiking squared. For a periodic network with roughly ten thousand neurons and a coefficient of variation of one — that is, Poisson-like spiking — the estimated coherence time is around 400 seconds. For aperiodic networks, rotational drift is the dominant decoherer, with estimated decoherence times ranging from about 85 seconds at a coefficient of variation of one, to about 680 seconds at a coefficient of variation of approximately one over the square root of eight. These aren't failures — they're honest upper bounds that identify exactly which biological parameters need to fall within what ranges for the mechanism to work. The paper closes by turning simulations into experimental predictions sharp enough to actually decide the question. The logic is this: if grid cells reflect a continuous attractor, then the population pattern is constrained to a low-dimensional manifold, and perturbations that push it off that manifold should be corrected. Phase relationships between cells should be conserved across allowed translations, and any correlated distortions — a shared stretch, a shared rotation, a shared lattice defect — should appear identically across every cell in the network, up to a global phase shift. The most decisive manipulation Burak and Fiete propose is selective head direction stimulation. In an attractor network, stimulating a subset of head direction inputs should produce a rigid translation of the entire population pattern — every grid cell shifts by the same phase, whether or not it received direct stimulation. Independent cell models predict phase shifts only in the stimulated cells. That's a clean prediction, testable with current multielectrode recording technology. Pharmacology provides a second probe. Local blockade of spiking in the medial entorhinal cortex should abolish spatial periodicity even in subthreshold activity if periodicity arises from recurrent network dynamics. Models that compute grids within single cells would retain spatially periodic subthreshold structure under the same manipulation. Environmental rescaling is a third test. When grid fields stretch or rotate with the enclosure geometry, the attractor view predicts that phase relationships across the population are preserved — and that the amplitude of velocity modulation in head direction inputs should decrease along the stretched dimension by exactly the percentage of observed field stretching. If instead the population pattern itself deforms, phase relationships change, and that amplitude signature won't appear. What Burak and Fiete have built is a proof of concept that continuous attractor dynamics can, under biologically plausible conditions, implement accurate dead reckoning in the rat brain. The numbers are striking: sixteen thousand neurons, one hundred meters, ten minutes, and fifteen centimeters of error. But the deeper contribution is turning an abstract theoretical mechanism into a set of concrete, differentiating predictions. The question of whether a sheet of neurons in a running rat's entorhinal cortex is doing accurate dead reckoning geometry is no longer just philosophical. It's testable. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.