Rapid learning of predictive maps with STDP and theta phase precession

  1. Tom M George
  2. William de Cothi
  3. Kimberly L Stachenfeld
  4. Caswell Barry  Is a corresponding author
  1. Sainsbury Wellcome Centre for Neural Circuits and Behaviour, University College London, United Kingdom
  2. Research Department of Cell and Developmental Biology, University College London, United Kingdom
  3. DeepMind, United Kingdom
5 figures and 1 additional file

Figures

STDP between phase precessing place cells produces successor representation-like weight matrices.

(a) Schematic of an animal running left-to-right along a track. 50 cells phase precess, generating theta sweeps (e.g. grey box) that compress spatial behaviour into theta timescales (10 Hz). (b) We …

Figure 2 with 4 supplements
Successor matrices are rapidly approximated by STDP applied to spike trains of phase precessing place cells.

(a) Agents traversed a 5 m circular track in one direction (left-to-right) with 50 evenly distributed CA3 spatial basis features (example thresholded Gaussian place field shown in blue, radius σ=1

Figure 2—figure supplement 1
STDP and phase precession combine to make a good approximation of the SR independent of place cell size and running speed statistics.

(a) Figure 2 panels a-e have been repeated (additional 30 min simulation carried out) for ease of comparison. (b) We repeat the experiment with non-uniform running speed. Here, running seed is …

Figure 2—figure supplement 2
The STDP and phase precession model learns predictive maps irrespective of the weight initialisation and the weight updating schedule.

In the original model weights are set to the identity before learning and kept (‘anchored’) there, only updated on aggregate after learning. In these panels, we explore variations to this set-up. (a)…

Figure 2—figure supplement 3
A hyperparameter sweep over STDP and phase precession parameters shows that biological parameters are suffice, and are near-optimal for approximating the successor features.

(a) A table showing all parameters used in this paper and the ranges over which the hyperparameter sweep was performed. For each parameter setting, we estimate performance metrics to judge whether …

Figure 2—figure supplement 4
Biological phase precession parameters are optimal for learning the SR.

(a) We model phase precession as a von Mises centred at a preferred theta phase which precesses in time. This factor modulates the spatial firing field. It is parameterised by κ (von Mises width …

Place cells (aka. successor features) in our STDP model show behaviourally biased skewing resembling experimental observations and successor representation predictions.

(a) In the loop maze (motion left-to-right), STDP place cells skew and shift backwards, and strongly resemble place cells obtained via temporal difference learning. This is not the case when theta …

Multiscale successor representations are stored by place cells with multi-sized place fields but only when sizes are segregated along the dorso-ventral axis.

(a) An agent explores a 1D loop maze with 150 places cells of different sizes (50 small, 50 medium, and 50 large) evenly distributed along the track. (b) In rodent hippocampus, place cells are …

Author response image 1
In a fully spiking model, CA1 neurons inherit phase precession from multiple upstream CA3 neurons.

Top, model schematic – each CA1 neuron receives input from multiple CA3 neurons with contiguous place fields. Bottom, position vs phase plot for an indicative CA1 neuron, showing strong phase …

Additional files

Download links