A New Theory of Dreaming by Hobson and Friston
Dreaming is a generative model of the world.
Posted October 17, 2012
Two highly respected neuroscientists, J. Allan Hobson and Karl Friston, have recently published a new theory of dreams (J.A. Hobson and K.J. Friston; Waking and dreaming consciousness: Neurobiological and functional considerations. Prog Neurobiol. 2012 July; 98(1): 82–98. doi: 10.1016/j.pneurobio.2012.05.003; PMCID: PMC3389346).
The theory formalizes a conception of the dreaming brain as a simulation machine or a virtual reality generator that seeks to optimally model and predict its waking environment and needs REM sleep processes, (particularly PGO waves) to do so. The basic idea is that the brain comes genetically equipped with a neuronal system that generates a virtual reality of the waking world during REM sleep because REM sleep processes are essential to optimize this generative model.
Treatments of the mind/brain as a virtual reality machine or a prediction-error device or a “Helmholtz machine” (all roughly the same thing) are rife throughout the cognitive and neurosciences, and it makes a lot of sense to consider dreaming along these lines as well. A dream, after all, is experienced as a fully realized “world” that appears to be generated internally without benefit of current sensory input (as sensory input is blocked during REM).
Hobson and Friston suggest that sensory data are sampled during wakefulness to build-up a complex model of the world that can guide behavior and reduces prediction error and surprises. Then the model is taken offline during sleep and is subjected to an optimization procedure that prunes redundancy and reduces complexity, thus improving the model’s fit to the world.
During waking life changes in the model’s parameters (experienced subjectively as percepts) are driven by the need to explain unpredicted visual input. During dreaming however, there is no visual or other sensory input so dreaming percepts are driven by the need to explain unpredicted oculomotor input. Dream content therefore is the brain's attempt to find plausible explanations for fictive visual searches triggered by oculomotor input (via rapid eye movements and PGO waves presumably) and by the pruning of synaptic connections that is part of the complexity reduction optimization process.
Why is it necessary to go offline to optimize the simulation machine? The authors in my view never adequately answer this question. The optimization procedure gives us a better model that can better guide behavior. That is fine and good, but it does not explain why optimization has to happen offline. After all, optimization of the model proceeds during waking life and arguably does so much more efficiently given the sensory feedback available to the waking brain.
The authors suggest that using the offline option was especially acute for the complex brains of mammals (and birds) that exhibit REM sleep. But REM sleep measures are not correlated with brain size or complexity. There are many animals (e.g. marsupials) with a lot of REM sleep and not very complex brains.
The authors also suggest that their theory throws some light on the lapse in thermoregulatory reflexes that are characteristic of REM. The reversion during REM to a poilkothermic state has long been one of the many biological mysteries associated with REM. Why does Mother Nature subject the animal to a dangerous lapse in thermoregulation during sleep? The authors argue that among other functions, the simulation machine generates predictions concerning thermal needs and conditions of the organism. But if the machine is taken offline, “the brain will be impervious to fluctuations in temperature and will not respond to suppress thermal prediction errors, resulting in a suspension of homeothermy." But this is tantamount to saying that thermal regulatory processes cannot proceed because thermal regulatory reflexes are inhibited as part of the REM state. But what we want to know is why these reflexes are inhibited in the first place.
Perhaps the authors want to argue that sensory reflexes and input in general are inhibited as part of the REM state because the optimization procedure cannot work unless all sensory input is gated. The authors argue that optimization can proceed with sensory gating, but they do not establish that it must proceed with gating; i.e., that gating is required.
The fact that optimization can occur during waking life argues against the idea that gating is absolutely necessary. Note that the benefits of offline optimization would have to outweigh its risks, including increasing vulnerability to predation, the lapse in thermoregulation and so forth.
To bolster the argument that offline optimization is necessary, the authors argue that without periodic offline repair (pruning), the model will become overly complex and dysfunctional, thus reviving the old idea of Francis Crick that REM dreaming represents a purging or pruning of superfluous associations and complexity in the cognitive system. “In short, taking the brain off-line to prune exuberant associations established during wakefulness may be a necessary price we pay for having a sophisticated cognitive system that can distil complex and subtle associations from sensory samples.” But again, there are many species of animals without complex cognitive systems that nonetheless have abundant REM and vice versa—there are species with complex brains with little or no REM (e.g. some sea mammals).
Does viewing the brain as a virtual reality or generative model of the world help us understand dream content? To answer this question the authors urge appropriate caution: “ finding order in the real world may not be the same as finding order in the virtual world.” A virtual world that is undergoing a process of model-fitting or optimization is likely to generate all kinds of unpredictable content it seems to me. That is why I think the Hobson–Friston theory needs major emendation to work for dream content.
Dreams are not all that unpredictable. Thousands of dream content studies have now clearly established regularities in dream content. Such regularities are broadly consistent with the dream as virtual reality machine theory but the theory needs to take dream content regularities seriously if it hopes to obtain a good fit with the data. To obtain that fit the authors suggests that more than one modeling process must come into play as part of the optimization procedure. An optimal balance between rehearsing what has already been learned about the world and exploring new hypotheses and possibilities that could be experienced needs to be struck. This suggestion makes sense to me.