CTC and Attention

Communication through coherence and a three-population example of selective attention.

Created

Updated

Collections

Notes
CTC Attention

Abstract

Background

Communication through coherence (CTC) proposes that rhythmic synchronisation regulates communication between neuronal populations without changing their anatomical connections. In Fries’s account, gamma rhythms create brief excitatory windows followed by longer inhibition; entrainment aligns incoming volleys with these windows despite transmission delays, without requiring zero phase difference. Representations may remain encoded in spatial firing-rate patterns, while synchronisation determines their downstream influence. Attention is proposed to bias this competition through gamma strength and frequency, with alpha–beta feedback providing control and theta rhythms periodically reopening selection [1].

Mechanism

Let A and B represent competing stimuli and both send excitatory projections to C. Each population contains excitatory and inhibitory neurons capable of gamma dynamics. A and B have no direct connections to each other.

  1. Attention biases A upstream. A control signal strengthens A’s gamma synchronisation and/or makes its rhythm slightly faster. B remains active and continues sending input to C.
  2. A gains an arrival-time advantage. Following a shared phase reset, A’s faster rhythm can produce a synchronised volley that reaches C before B’s. This advantage depends on the projection delays and circuit dynamics.
  3. A activates C, then recruits C’s inhibition. A’s volley excites C’s excitatory neurons, which recruit local inhibitory neurons. Fries describes excitation preceding inhibition by roughly 3 milliseconds. B’s later volley encounters inhibition and therefore has less influence.
  4. C becomes entrained to A. Subsequent A volleys arrive as C’s inhibition decays, repeatedly finding an excitable window. C’s output preferentially reflects A’s representation, while B’s input is suppressed.
  5. Selection can be reopened. A theta-related reset may terminate the existing timing relationship and allow renewed competition. Selection can therefore persist across several gamma cycles without becoming permanent.[1]

Proposed Experiment

We propose training the A/B → C network to classify one of two simultaneous digits according to an attention cue. This adapts Bosman’s competing-input task to supervised classification; it is a proposed experiment, not a replication or completed result [2].

  1. Sensory inputs. Independently sample two 8×8 scikit-learn digits and encode each through 64 spike-input channels, one image into A and the other into B. Both images remain present during classification.
  2. Cue and target. Cued control spikes favour A or B through the corresponding Attention input. The target is the cued image’s digit class; changing the cue leaves both sensory inputs unchanged.
  3. Classification. A ten-class readout receives activity only from C. Neither the images, A/B activity nor the cue bypass C to reach the readout. Train against the attended digit’s label.
  4. Evaluation. Split source images into training, validation and test sets before constructing pairs. Balance cue choices and digit classes. Test each held-out pair with both cues and identical sensory spike trains; for different-class pairs, the predicted class should switch with the cue.
  5. CTC tests. After training, freeze the weights and apply the three tests below to held-out image pairs across training seeds.

Circuit

Sensory and attention inputs drive A and B, whose excitatory populations target E and I in C. Background C independently drives C's E and I cells without digit or cue information. Darker paths emphasise inter-area projections; lighter paths show local recurrence. Optional feedback from C to A and B appears in a separate inset with repeated population symbols.

Figure 1. Illustrative A/B → C circuit. E and I denote excitatory and inhibitory neurons; subscripts identify the circuit. A/B represent separate V1 sites and C a shared V4 receiver. Black/grey arrows denote excitation and red bars inhibition. Inter-area projections are darker and heavier than local recurrence; this visual emphasis does not encode synaptic strength. Solid connections are kept and dashed connections are optional. The feedback inset repeats the same populations, not additional neurons. X_A/X_B denote sensory inputs, X_C task-independent background input to C’s E and I cells, and U attention inputs. W labels weights with the receiving population first and the source second. Select the diagram to enlarge it.

This circuit implements Bosman’s competing-input motif [2]. The statuses below describe our model design; the recordings do not establish individual E/I synapses. All optional connections are included in the diagram and illustrative bundle. Omitted pathways are absent.

Modelling choices

Circuit choiceWhy
Keep
Excitatory input from A/B to CMatches V1 inputs converging on a shared V4 receiver.
A/B recruit C's inhibitory cellsImplements inhibition invoked in Bosman's timing account.
Local excitation–inhibition loopsOur mechanism for gamma; not specified by Bosman.
Separate sensory and cue inputsProvide two stimuli and a controllable attention bias.
Background input to CSets baseline excitability; independence, strength and E/I targets are modelling choices.
Optional
Recurrence within E and I populationsCompatible with the paper, but not required by it.
Excitatory feedback from C to A/BBosman discusses feedback; its E/I targets remain our choice.
Omission
Inhibitory projections between areasNot required by the paper; omitted for simplicity.
Direct A–B connectionsKeeps competition in C; does not imply V1 populations are isolated.
Separate cortical layersCombines layers for simplicity, including Bosman's proposed deep-layer feedback source.
Table 1.

Circuit choices relative to Bosman et al. [2]. Parameters are illustrative. Background C carries no digit or cue information; keep its statistics fixed across cue conditions. Disable optional feedback for the A→C baseline tasks.

Testing CTC

Test whether gamma timing contributes to selective routing:

  1. Cue-dependent coupling. With identical sensory inputs, changing the cue should switch both C’s preferred gamma coupling and its classification. Compare with trial-shuffled controls, matching firing rates and gamma power.
  2. Phase-dependent transmission. Identical weak probes should evoke different responses at different phases of C’s gamma cycle. The cued sender’s volleys should arrive preferentially at high-gain phases.
  3. Timing disruption and rescue. Disrupt attended-input timing while preserving spike counts and within-volley synchrony, then restore it. Selection should deteriorate and recover beyond effects seen with sham, distractor and rate-matched nonoscillatory controls.

Limitations

Classification or coherence alone cannot establish CTC. Rate-based solutions, shared inputs and perturbation side effects require controls. Positive results would support CTC in this model, not establish the brain’s attention mechanism.

Plan

Proceed through six tasks, using mean-field models that retain E/I gamma oscillations before testing the corresponding spiking circuits.

  1. Mean-field coupling. Map stable gamma activity and entrainment across coupling strengths and measured uncoupled frequency differences, using A→C only with feedback disabled. Entrainment means C locks to A’s frequency with a stable phase difference; their peaks need not coincide.

    exp153Mean-field CouplingUpdated

  2. Spiking coupling. Repeat the same analysis using a spiking model.

    exp133Lowet (2015): Spiking CouplingUpdated

  3. Mean-field competition. Keep sensory drives fixed and vary the cue-related drive. Map selective locking, cue-driven switching, persistent selection and regimes where neither sender dominates. Apply matched weak probes to each sender and compare C’s responses, accounting for firing-rate differences, to test effective transmission as well as synchronisation.

  4. Spiking competition. Verify selective coupling, preferential transmission and reliable cue-driven switching in the corresponding circuits.

  5. Spiking digit classification. Establish classification through either pathway alone, then train with simultaneous digits and balanced cues. Evaluate held-out image pairs under both cues using a readout receiving only C’s activity.

  6. CTC validation. Freeze the trained weights and apply the three tests in Testing CTC across training seeds, controlling for firing rates, gamma power and perturbation effects.

References

  1. P. Fries. “Rhythms for Cognition: Communication through Coherence.” Neuron 88(1), 220–235 (2015). doi:10.1016/j.neuron.2015.09.034
  2. C. A. Bosman et al. “Attentional Stimulus Selection through Selective Synchronization between Monkey Visual Areas.” Neuron 75(5), 875–888 (2012). doi:10.1016/j.neuron.2012.06.037