Dreams, mnemonics, and tuning for criticality

Barak A. Pearlmutter and Conor J. Houghton

Open peer commentary on Such stuff as dreams are made on? Elaborative encoding, the ancient art of memory, and the hippocampus by Sue Llewellyn

Abstract: According to the tuning-for-criticality theory, the essential role of sleep is to protect the brain from super-critical behaviour. Here we argue that this protective role determines the content of dreams and any apparent relationship to the art of memory is secondary to this.

It is widely believed that memory consolidation is the purpose of sleep. However, as detailed in the target article, the description of this process that emerges from experiment is confusing and complex. Mindful that complex phenomenology is commonly the attribute of a secondary purpose, we proposed (Pearlmutter & Houghton 2009) that sleep has a different primary purpose - tuning for criticality - and that the link with memory consolidation is secondary. Memory consolidation may occur during sleep, it may even occur preferentially during sleep, but it is not the essential purpose of sleep.

The obvious goals of learning, rapid responses to stimuli, and prolonged retention of short-term memories are, from a network dynamics point of view, attributes of near-critical systems. Thus, it is likely that, during learning, neuronal circuits become increasingly critical and approach super-critical behaviour - behaviour that would involve runaway oscillations and constitute a pathological disruption of normal brain function. Thus, the optimisation of behaviour during learning requires a mechanism for preventing the brain from straying into a parameter region where it could be stimulated into pathological oscillations, and learning can occur only if there is a margin of safety around the current state of the brain.

In the tuning-for-criticality theory the purpose of sleep is to establish this margin of error. Thus, during sleep there is a synaptic plasticity regime which is different from the one which supports learning during wakefulness, and this works to move the brain away from criticality. In this way the sleep-work cycle tunes the brain so that it is close to criticality and optimised to respond to likely stimuli, but safe from the uncontrolled behaviour associated with super-criticality.

According to this proposal, the role of dreams is to stimulate the brain in a manner that mimics awake cognition as part of a search for self-reinforcing loops. Dreams are therefore a guess, based on recent and distant memories, of possible future stimuli. The target article describes the attributes of dreams. Dreams are emotionally salient and sometimes shocking or disgusting, they are associated with pontine-geniculate-occipital waves, they are narrative but the narrative is fragmentary, they contain incongruities, and the identities of people in them are often fused or indeterminate. Dreams are largely visual, with snatches of auditory sensation but almost no olfactory, gustatory, or tactile content. They are often characterised by the illusion of movement and spatial navigation. The target article points to this as a congruence between the form of the classic mnemonic and dreams, though it does not elaborate on what cognitive or biochemical mechanism related to this congruence would act to improve the efficacy of the mnemonic. We contend, however, that these are all attributes which dreams might be expected to have if they are a mechanism the brain uses to cast around for neuronal circuits in which potential stimuli could evoke runaway oscillations.

Pontine-geniculate-occipital waves arise only in wakefulness in response to unexpected events and are believed to focus attention, heightening responses. Emotionally salient, disgusting, and shocking dream content is likely to provoke stronger neuronal responses. Because the brain responds to temporally integrated stimuli, the exploratory simulated environment produced by dreams needs a narrative structure, but for this purpose, the narrative may be fragmented and discontinuous. The dream world need not be wholly specific; fused and indeterminate aspects of dream content seem unlikely to reduce its capacity to evoke responses indicative of potentially harmful self-reinforcing loops. The visual, auditory, and ambulatory character of dreams reflects the timescales and stimulus types most pertinent to everyday learning and common threats. In short, the nature of dream content is consistent with a tuning-for-criticality function of sleep. The pertinence of dream content means that similar content is likely to be memorable and this might explain the utility of this type of content in mnemonics.

Learning for behavioural optimisation is a key objective of cognition and content salient to this objective is likely to be both memorable and provocative. It may be this, rather than a direct functional link, that relates the art of memory and the nature of dreams. Alternatively, we might speculate that the random activity involved in dreaming first evolved to avoid criticality and that the availability of random narratives based partly on recent memories encouraged the development of learning strategies which make strong use of these narratives through a sort of hashing algorithm. In this light, we would view mnemonic consolidation as a particularly apt secondary purpose of sleep, in that it makes use not only of physical quiescence but also exploits the computational processes already present in tuning for criticality.

Mnemonic consolidation may avoid waste by utilizing the dream content that was generated, at some metabolic expense, to locate potentially super-critical neuronal circuits. However, we believe that there is no evidence to suggest that memory processes like consolidation and mnemonic association can occur only during sleep. Sleep is such an extreme and hazardous adaptation it seems unlikely it evolved to serve a function, like mnemonic association, that could also be performed while awake.

Reference

Pearlmutter, B. A. & Houghton, C. J. (2009) A new hypothesis for sleep: Tuning for criticality (2009). Neural Computation 21(6):1622-41.