The model and task.
a) Overview of the decoding approach: Given a simulated trajectory with coordinates x, the output states of the network are decoded in terms of their spatial center locations µ, which in turn are used to decode an estimate of the current location . The network is trained to minimize the squared difference between true and decoded positions. b) Illustration of the proposed decoding procedure. For a single unit, the center location is estimated as the average location, weighted by the unit activity along a trajectory. By iterating this procedure, every unit can be assigned a center location. A location can then be estimated as the average center location, weighted by the activity of the corresponding unit at a particular time. Repeating this for every timestep, full trajectories can be reconstructed. c) The investigated geometries, each with an example simulated trajectory. Each environment is labelled by its context signal (one-hot vector). d) Illustration of the network architecture and inputs. g features recurrently connected units, while p receives densely connected feedforward input from g. When moved between environments, the state of the RNN is maintained (gprev). The input v denotes Cartesian velocities along simulated trajectories, while c is a constant (in time and space) context signal.