Context-dependent Computation by Recurrent Dynamics in Prefrontal Cortex. Mante, Sussillo, Shenoy, Newsome. Nature 2013.

  1. Study behavior of macaque behavior in the face of noisy stimulus
  2. Although individual behavior is complex and difficult to tie down, at a population level analysis is more fruitful and can be related to a simulated recurrent NN
  3. “This mechanism indicates that selection and integration are two aspects of a single dynamical process unfolding within the same prefrontal circuits, and potentially provides a novel, general framework for understanding context-dependent computations.”
  4. Behavior can change drastically and rapidly according to changing context, related to gating
    1. “These observations have led to the hypothesis that early selection may account for the larger effect of relevant as compared to irrelevant sensory information on contextually sensitive behaviour.”
  5. In the task studied here, context-dependent behavior is required based on visual stimulus
  6. Neural responses recorded “… in and around the frontal eye field … an area of PFC involved in the selection and execution of saccadic eye movements, the control of visuo-spatial attention, and the integration of information towards visuomotor decisions.
  7. “We found no evidence that irrelevant sensory inputs are gated, or filtered out, before the integration stage in PFC, as would be expected from early selection mechanisms.  Instead, the relevant input seems to be selected late, by the same PFC circuitry that integrates sensory evidence towards a choice.  Selection within PFC without previous gating is possible because the representations of the inputs, and of the upcoming choice, are separable at the population level, even though they are deeply entwined at the single neuron level.”
  8. Monkeys were able to act appropriately based on cue and ignore irrelevant information
  9. “As is common in PFC, the recorded responses of single neurons appeared to represent several different task-related signals at once, including the monkey’s upcoming choice, the context, and the strength of motion and color evidence (…).”
    1. So instead they attempt to understand activity at population level
  10. Most units were recorded during separate sessions <Hm…>
  11. The task involved stimulus with colored moving dots.  In some cases the color was more important for decision making and in other cases the motion was
  12. Population responses are represented as trajectories. “We focused our analyses on responses in a specific low-dimensional subspace that captures across-trial variance due to the choice of the monkey …” also based on motion, color information and which of those was relevant for decision making in that context
    1. Uses PCA and then a projection onto independent axes of choice, motion, color, and context
  13. “This population analysis yields highly reliable average response trajectories (…) that capture both the temporal dynamics and the relationships among the task variables represented in PFC.”
  14. For example, trajectories go in opposite directions based on saccade direction
    1. Likewise, patterns of activation of PFC “… are very different from those corresponding to either choice…”
  15. “Indeed, the population response does not follow straight paths along the choice axis, but instead forms prominent arcs away from it (…).  The magnitude of each arc along the axes of motion or colour reflects the strength of the corresponding sensory evidence (…), whereas its direction (up or down) reflects the sign of the evidence.”
  16. “… the signals along axes of motion and colour are transient — the arcs return to points near the choice axis by the time of dots offset.”  Because they come and go, they are called momentary evidence
  17. “Third, context seems to have no substantial effect on the direction of the axes of choice, motion and color, and only weak effects on the strength of the signals represented along these axes.”  So 3 axes are sufficient to represent the data
  18. Although direction of trajectories are invariant to context, where those lie in the state space are clearly separated spatially when considering context
  19. <Tried to upload graph of  results but WP won’t cooperate right now… anyway> Depending on what the context is (selection by color or motion), projecting the data one way or the other will either allow the data to be trivially linearly separable (with a line running through the origin), or will not be linearly separable
  20. There is however, fair separability even on items that have no behavioral effect
    1. This is unsupported by the early selection model, so there must be something else going on
  21. Then they made a recurrent network and trained it to carry out the task
  22. “As in PFC, the contextual input does not affect the strength of the sensory inputs–selection occurs within the same network that integrates evidence towards a decision.”
  23. “After training, the model qualitatively reproduces the monkeys’ behaviour, confirming that the model solves the selection problem at the ‘behavioral’ level (…).”
  24. Analysis of NN population has same characteristics as those in monitored neurons
    1. “… integration of evidence corresponds to gradual movement of the population response along the choice axis.”
    2. “… momentary motion and color evidence ‘push’ the population away from the choice axis, resulting in trajectories that are parametrically ordered along the motion and color axes.”
    3. “… direction of the axes of choice, motion and color are largely invariant with context, as are the strength of the motion and color inputs, as these are not gated before entering the network.”
    4. “… trajectories during motion and colour contexts are separated along the axis of context (…).”
  25. The main difference between neural and NN responses is that “… signals along the input axes are transient in the physiology, but not in the model, yielding PFC trajectories that curve back to the choice axis before the end of the viewing interval (…).   This difference suggests that the sensory inputs to PFC are attenuated after a decision is reached.”
  26.  “Notably, the rich dynamics of PFC [and ANN] responses during selection and integration of inputs can be characterized and understood with just two features of a dynamical system–the line attractor and the selection vector, which are defined only at the level of the neural population.  <I don’t know what either of those two things are>  This parsimonious account of cortical dynamics contrasts markedly with the complexity of single neuron responses typically observed in PFC and other integrative structures, which reveal multiplexed representation of many task-relevant and choice-related singals.  In light of our results, these mixtures of signals can be interpreted as separable representations at the level of the neural population.  A fundamental function of PFC may be to generate such separable representations, and to flexibly link them through appropriate recurrent dynamics to generate the desired behavioral outputs.”

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: