Features Cover story

Dream power
The fictions we conjure while we sleep may do
something far more profound than reinforcing
learning, says neuroscientist Erik Hoel

I
F ALIENS ever visited Earth, they bookstore, and as a novelist, this question of
might notice something strange. the importance of fictions is especially dear
Nearly everyone, everywhere, spends a to me. I think the imaginary aliens are in the
significant part of their day paying attention same position as a scientist attempting to
to things that aren’t real. Humans often care explain the evolved purpose of dreams – and
fiercely about events that never happened, if we can identify the biological reason for
whether in TV shows, video games, novels, dreaming, we can ask if it applies to the
movies. Why care so much about fictions? artificial dreams we call fictions.
Perhaps, these aliens might hypothesise, As a neuroscientist, I’ve been working
humans are too stupid to distinguish on a hypothesis that draws on what we’ve
between truth and falsehood. Or perhaps learned about artificial neural networks
they pay attention to fake events for to cast dreaming as a way to improve our
the same reason that they eat too performance in waking life, just not in
much cheesecake: both are non-natural the way we might think. If correct, it may
outcomes of evolved interests. also explain some of this strange human
The aliens’ confusion might deepen attraction to the unreal in our waking lives.
when they learned that humans fall asleep The study of dreams, also known as
and dream. For dreams are also fictions. oneirology, suffered something of a false
Dreaming takes time and energy, so start in the first decades of the 20th century,
presumably has an evolutionary purpose. when it was tainted by association with
The aliens might begin to wonder what Sigmund Freud’s ideas about psychosexual
HARDZIEJ STUDIO

they are missing about the importance of development. Freud argued that dreams are
experiencing things that never happened. an expression of repressed desires resulting
As someone who grew up in my family’s from traumatic experiences in early life.

34 | New Scientist | 7 November 2020

These ideas have been discredited, but dream
research never quite shook the association.
Luckily, over recent decades, neuroimaging
and behavioural research have reinvigorated
the field by giving us insight into the
biological mechanisms underlying dreams.
We now know that dreams are the result of
localised firing of neurons that is probably
induced by the brain’s many feedback
connections and not dependent on
information from external stimuli. Dreaming
represents a unique physiological state in
which activity similar to that we see when
we are awake is promoted while behaviour
is essentially cut off by powerful chemical
systems that induce paralysis.
Yet although we now know a good amount
about the mechanisms of dreaming, we have
little insight into its function. Some argue
that we don’t need to understand what
dreams are for. Perhaps they are just a
by-product of sleep, which may have
evolved for some other reason, such as
to clear the metabolic detritus generated
by neuronal activity.
But this “null hypothesis” of dreams
has been challenged by a slew of ideas
about how dreams have an evolved purpose.
After all, we spend hours every night
dreaming in a distinct stage of sleep.

Making memories?
Generally, these dream hypotheses have
trouble accounting for the distinct
phenomenology of dreams: their unique,
highly specific nature, which is what sets
them apart from waking experience.
Dreams are sparse, in that they mostly don’t
contain the vivid sensory detail of waking
life. Dreams are hallucinatory, in that they
contain warped concepts and perceptions
that are biased or unrealistic. And dreams
are narrative, in that they are fabulist
versions of the kinds of events we might
encounter in real life, just rendered strange.
Consider the leading hypothesis, which
is that dreaming is somehow involved in
the process of memory storage. This idea
draws on the metaphor of the brain as a >

7 November 2020 | New Scientist | 35

computer: explicit memories are created
and then stored, the way one encodes data “Dreams may
on a hard drive. Neuroscience has long drawn
on such metaphors, even from before it was serve as ‘noise
called “neuroscience”, when the metaphors
were pneumatic pressures or mechanical
clocks. But sometimes metaphors can lead
injections’ to
their proposers astray. In the case of sleep
and memory, it is well known that various
counteract the
improvements can occur after a good
night’s sleep, such as performance on risk of learning
some tasks, but it is less clear that acts of
pure memorisation, like lists of numbers, too narrowly”
are actually significantly improved.
What would it even mean to help store
a memory over a night? The clearest

SCOTT MACBRIDE/GETTY IMAGES

hypothesis about memory storage and
sleep is based on studies showing that
memories, in the form of the specific
neural sequences of firing that are seen while
we are awake, are sometimes “replayed”
during sleep in mammals. Perhaps dreams
are just that: replays of memories.
While neurons that learn do seem to
increase in their firing frequency during in one night and then demonstrate it in front
sleep, two facts suggest the idea falls short. of the class. I practised all evening, tossing
The first is that replay has been more strongly tennis balls helplessly, but eventually fell
associated with non-REM sleep than the into bed, certain of embarrassment the next
REM stage, where the most intense narrative day. On waking, I immediately jumped out
dreaming occurs. The second is that it is The purpose of of bed, picked up the tennis balls, and found
unclear whether memories are actually being dreaming remains I could juggle perfectly. It was an incredible
replayed during so-called “replay.” Indeed, one of the great lesson. It seemed that something had
careful studies have demonstrated that mysteries of the brain happened in my sleep that had built on
the brain more commonly produces never- my waking experience.
before-seen patterns during these periods Even so, I find it hard to accept that I had
rather than previously seen waking patterns. stored or replayed memories of my juggling
Behavioural evidence is also a problem during sleep. When I went to bed, I couldn’t
for the idea that dreams are somehow juggle. If I had replayed my failures, what
replays of memories, or even just by-products would be the gain? Most importantly, it
of the integration of memories. If this were is doubtful I dreamed of precise juggling
the case, we would expect to dream actual events. More probably, if I dreamed of
memories, yet dreaming specific previous juggling at all, it was of sparse and
memories is actually so rare that it is hallucinatory fragments.
considered pathological, often a sign This is backed up by studies that have had
TANG MING TUNG/GETTY IMAGES

of post-traumatic stress disorder. participants play games like Tetris, which

Yet there is no doubt that dreams do play they were novices at, and found that they
a role in memory and learning. Consider reported Tetris-like dreams – imagine falling
how I learned to juggle. As an undergraduate, hallucinatory blocks – but no replays of
I took a class on memory and as part of my specific Tetris games. It seems that the
homework I was assigned to learn to juggle best way to get someone to actually dream

36 | New Scientist | 7 November 2020

 How can someone who
went to sleep failing
to learn to juggle wake
up the next morning
as a juggler?

about something is to have them learn the network’s connections are tweaked until the research company OpenAI trained a deep
a difficult and novel task, and then have it can parse the training data set effectively, neural network to learn how to manipulate
them overtrain on it, as with playing Tetris which would be things like classifying a robot hand to solve Rubik’s cubes.
for hours and hours. images, playing a game or driving a car. There is good reason to think the brain
The hope is that the performance faces an identical challenge of overfitting.
generalises beyond the training data set Animals’ days are, after all, statistically pretty
Deep lessons to new, unseen data sets. But it doesn’t self-similar. Their “training set” is limited
A new and growing trend in neuroscience always work so well because training data and highly biased. But still, an animal
might help explain why this is the case, and sets are often inherently biased in all sorts needs to generalise its abilities to new and
offer a clear explanation for why dreams of impossible-to-notice ways. Often a unexpected circumstances, both in terms
possess their distinct phenomenology. network gets so fine-tuned to the specifics of physical movement and reaction, and
This trend seeks to apply the lessons of the data set it is trained on that it fails cognition and understanding. It doesn’t
of deep learning and the study of to generalise to new ones. need to remember everything perfectly;
artificial neural networks to the brain. This is called overfitting, and it is a it needs to generalise from the limited
These techniques are, after all, originally ubiquitous problem in deep learning. things it has seen and done.
inspired by how the brain functions, A number of common techniques have This is the overfitted brain hypothesis
and remain the only set of techniques by been adopted to deal with this issue. Most (OBH): that animals, being so good at
which machines can reach human-level involve exposing the network to some sort learning, are constantly in danger of
cognitive performance on complex tasks. of stochasticity, introducing noise and fitting themselves too well to their daily
From a deep-learning perspective, learning randomness into the system. lives and tasks.
isn’t like storing memories on a computer. One such strategy is “domain I’ve recently been working on developing
Instead, it is about fine-tuning a huge, randomisation”, wherein the inputs are the OBH, exploring how dreams could be a
layered network of connections based on an warped in a highly biased way during way to beat back the tide of daily overfitting.
inherently limited set of example data – the learning, effectively inducing a hallucination Essentially, under the OBH, dreams are
“training” data set. With every example that in the network. This sort of thing has been “noise injections” that serve the purpose
the system sees, the pattern and strength of found to be indispensable, for example, when not of enforcing what is learned when >

7 November 2020 | New Scientist | 37

awake, but rather counteracting the
overfitting associated with that learning.
You can’t do domain randomisation on
an awake brain because most organisms are
negotiating a high-wire act during daily life;
they would certainly hurt themselves in
myriad ways. However, you can use an offline
period to do something similar by creating
sparse and hallucinatory inputs, driven by
top-down activity, that resemble the events
and actions an animal might encounter,
but that are corrupted and biased away
from the drudgery of daily life.
According to the OBH then, dreams are
exactly this: self-generated corrupted inputs.
And the act of dreaming has the effect of
improving generalisation and performance
in waking life. This is how someone can
HARDZIEJ STUDIO

go to sleep failing on their training task

of juggling, and then wake up a juggler.
The advantage of this hypothesis is that
it takes the phenomenology of dreams
seriously, rather than as some sort of
epiphenomenon or unexplained by-product
of some other neural background process.
Indeed, it is the strange phenomenology during sleep – are all still being investigated all help to stop our minds becoming too fixed
of dreams that makes them so effective at more generally. in their ways. They don’t just expand the
combating overfitting. While it may seem But by viewing dreams through this new “training set” that humans have access to, but
weird, experiencing events that are related lens, we can at least move beyond computer do so in ways that assist with generalisation
to a task, but fundamentally different and storage metaphors and begin to think and therefore cognition more broadly.
from it, can actually help performance. of learning as a set of trade-offs, where Perhaps the hypothetical aliens wouldn’t
Dreaming of flying may help you keep your memorisation competes with generalisation, be so puzzled by our obsessions with
balance while running. And deep-learning and learning the specifics of something too fictions once they figured this out. They
practitioners should perhaps take a lesson well can be as bad as not learning at all. wouldn’t be shocked either that as human
from the brain and make their efforts to If dreams have this functional purpose, civilisation developed, daily life became
combat overfitting look as “dream-like” and the OBH is true, then the artificial dreams more complex, and so it became easier
as possible for their networks. we call fictions might satisfy some of that for us to overfit to it – until eventually we
same fundamental drive. I spent 10 years humans began to spend more time with
writing my first novel, The Revelations, artificial dreams than we do with biological
Waking dreams which is about consciousness and murder. ones. Just like how the invention of cooking
Of course, this is still very much a I can give all the standard cultural reasons essentially allowed us to expand digestion
hypothesis – and an untested one at that. for why fictions are important, entertaining, beyond our stomachs, maybe the invention
There is much work that needs to be done revelatory – but the OBH implies there of fictions allowed us to get the benefits of
to assess what the behavioural benefits is something more. Maybe art is also dreams when we are awake. ❚
of dreams are and whether they match pleasurable for humans because we are
the sort of reductions in overfitting that constantly being overfitted to reality.
we might expect in humans and other In this view, the sparse, sometimes Erik Hoel is a neuroscientist at Tufts
animals according to the OBH. Additionally, hallucinatory, corrupted unreality put University in Medford, Massachusetts.
dream physiology – how synapses change forward by authors, film-makers, and those His debut novel, The Revelations,
during dreams and when dreaming occurs first early shamans around some campfire, will be published in April 2021

38 | New Scientist | 7 November 2020