Brief manifesto on control and neuroscience

I've recently joined the "Control Lab" at the university of Cambridge, but I'll still be researching neuroscience. This is a fanciful attempt to summarize the current dogma in neuroscience about how things like "free will" and "homeostasis" might relate to technical concepts like "optimal control". None of this is new; it is mostly motivated by old ideas of cybernetics, as best explored by Todorov and colleagues. 

Computation in the brain employs predictive negative feedback. This inverts (i.e. controls) a system by cancelling a prediction error (residual). Learning rules employ this on slow timescales, consolidating these inverse representations into feed-forward circuits. Recurrent dynamics complement this, by generating the spiking output required to cancel any error not cancelled by the forward network. In turn, these recurrent dynamics refine the forward weights and reduce the amount of feedback needed in the future (assuming the environment is stationary). This then bring us back to the general idea of learning, perception, and control. Todorov's papers are a good treatment of this. 

The brain has access to two sources of error: primary homeostatic drives (thirst, hunger), and sensorimotor prediction errors. Classically, these are divided into supervised and unsupervised learning, respectively. (The machine learning community has recently appreciated that so-called unsupervised learning can really be phrased in terms of exchanging sensory prediction mismatch signals, an interpretation that they have called self-supervised.)

These two error signals divide cognition into two components. There is a  "free will" that is driven by the homeostatic needs of the organism and species. This drives the active component of neural activity: the spontaneous production of spiking patterns that constitute thought and action. This system can be viewed as motor control in the most general sense. This is the part of the brain implementing optimal control.

There is also a passive observer that learns the latent causal and correlational structure of reality. This observer builds internal models which mimic the external world. This system is a statistical estimator. This is the part of the brain implementing optimal statistical estimation, which is synonymous with predictive coding.

Both systems act on fast and slow timescales, through neural activity and learning, respectively. The "free will" uses internal feedback to cancel homeostatic errors, and is responsible for us doing anything at all. These homeostatic errors are a motive force: all will or action reflects (if distantly) a complex transformation of these needs. The observer uses active internal feedback to cancel its own prediction errors. We feel this when encountering a new environment or modality. Perception, and our memories of the things perceived, feels effortful. We misunderstand, or forget, and feel tired at the end of the day even if we have not moved or acted.

There is also an interesting bridge between the observer and the will. This is a source of motivation akin curiosity, but has its roots in more basic, almost reflexive movements, like the searching eye movements we make to take in a scene. This is active sensing: when we act in order to refine our internal sensory models. In active sensing, internal prediction residuals themselves trigger movements. This arises because error representations in the sensory (observer) system can couple to error representations in the "Will" system. One might think of this as a higher-order sensory drive: evolution has determined that perceptual uncertainty itself hurts survival, and so provides an inbuilt mechanism to actively reduce this uncertainty. (Ignorance should feel uncomfortable).

 All actions are goal directed. I remember in a lecture Koch produced the quote "We don't do what we want, we want what we do". The original is in German, and I can't track down the attribution at the moment. Like all control systems, the purpose of the Will is to generate causal rules that shape the correlation of states in time, in the body and in the external world. This control can be viewed at canceling external perturbations that are hostile for survival.

(… In conclusion, the brain is a very complicated thermostat, metabolism is a slow burn, and life is a fire that would prefer not to die. We exist only so long as we serve to accelerate the heat death of the universe.)