Die Traumdeutung*

Are dreams the brain’s training grounds?

As the field of Artificial intelligence (AI) rapidly grows, learning machines’ appetite for more data becomes a problem. In many cases, the amount of available data is simply not nearly enough to keep machines well fed. In such cases, researchers turn to synthetically generated data nuggets to keep the machines happy.

It is a bit like printing food using biological 3D printers, just that the results are not artificial hamburgers made of carefully mixed proteins, but rather chunks of data.

This data is then fed back to the machines, based on which they are further trained to understand the world (or at least the problem domain they are targeted to, e.g. medical, social, etc.).

As our brains are learning machines one wonders whether we, as persons, employ a similar process of generating synthetic data in order to better survive.

A Brief (and fun) Dive into Deep Learning

During the last years, many of us have encountered the term deep learning, however, few are well versed in the field. Let me take you on a short stroll into the minds of computers. Hang on, it’s going to be fun and not at all intimidating.

Deep networks are basically a copy of our biological brains. An artificial brain if you’d like.

A deep network is a web of interconnected neurons, which by themselves are very simple machines that know how to do one thing, and one thing only. Add. Well, sort of.

One may think of a single neuron as a person standing in a crowded room. On his left, there is a group of people providing him with facts and figures, and on his right, another group of people just standing there waiting for him to tell them what to do.

This person-neuron would then carefully listen to all the people on his left, gather each and every fact, opinion, or utterance they convey, weigh each by how well he trusts the opinions and facts coming from each person, and come to some conclusion. Once he has, he’ll broadcast his conclusion to all the people on his right.

Simply putting it, all the people on his left are upstream neurons, while the ones on his right are downstream neurons. Or if you’d like, input neurons and output neurons accordingly. The group to the right and left and the actual person-neuron are in fact a network. When we depict a room where many people are talking to many people in the above-described manner we get an image of what a deep network is.

You may envision it as a set of committees in parliament where each committee is good at something (environmental considerations), feeding its conclusions to yet another committee that is good at something else (building regulations), all working in cascading order in order to formulate new legislation. The law they will eventually pass is the joint effort of all the highly specialized groups of thinkers, while setting it in the rule book for years to come represents a memory, or learned model (of the world), in deep networks lingo.

Generated Reality

The following image was created by a massive deep network.

Yes, that is true.

Created by OpenAI’s DALL-E 2 System

Artificial deep networks have come such a long way, they can now generate images that never before existed. The OpenAI DALL-E 2 system can generate beautiful, some may even say original, paintings and images based on a brief textual description.

The above image was generated from the following description.

An astronaut, Teddy bears, A bowl of soup. Riding a horse lounging in a tropical resort in space playing basketball with cats in space, in a photorealistic style in the style of Andy Warhol as a pencil drawing.

So, should artificial deep networks be able to construct yet unseen dreamlike images and experiences, the human brain, a far more extensive network, should be able to too.

And it does.

The Synthesising Brain

If our brains do generate synthetic life experiences in order to expand our experience (and therefore our model of the world), it would make great sense. Carefully crafted synthetic world experiences can teach us how to behave in situations we have never before been through while experiencing them in the safe and controlled environment of the dreamscape.

It is long known that in order to treat traumatic memories deeply carved into the brain’s neural network, one needs to surface them in a controlled and safe environment (e.g. the shrink’s den) and re-write the narrative of the stressful situation that generated the brain structures exciting the traumatic response.

But have we considered dreamscape to be such an environment?

The short answer is yes. Dreams have been studied in the field of PTSD and are recognized in many ways to be the way of the brain to surface stressful memories and allow the dreamer to cope with them in the relative safety of the dream.

However, I would like to consider an alternative and complementary way of explaining the role of dreams in healing post-traumatic memories as well as fortifying the brain against future ones.

Recent work in cognitive neurosciences has established that the hippocampus, in addition to being involved in the formation of memories, is also part of a brain system that uses memory to construct novel imagined scenarios and simulate possible events

In fact studies on rodents have shown that in phases of wakeful rest during spatial navigation, and during subsequent sleep, the rodent hippocampus spontaneously regenerates sequences of neural activations, which resemble sequences observed during actual animal behavior… Yet, rodent and human studies have shown that the hippocampus does not literally “replay” previous experiences from memory. Internally generated sequences during sleep or wakeful rest can depict paths to future goal locations rather than only past trajectories, possibly supporting planning and imagination beyond purely memory function.

Should our assumption be correct, and the brain uses memories as a means to synthetically create virtual worlds in which the dreamer can safely experience challenging situations, it would follow that such synthetic example of virtual world events may enable the dreaming person to find new ways to respond to those (virtual) situations.

In other words, as deep neural networks in AI are fed artificial data in order to teach them how to respond to yet unseen situations, dreams may be the brain’s way of both mending responses to past trauma (allowing the person to respond differently in the dream as opposed to how she performed in real life, thus re-writing the neural scenery), as well as training the dreamer in situations he has never been through, allowing him to better prepare and learn how to cope with the like as they occur in real life.

For example, the threat simulation hypothesis suggests that dreams may provide a sort of virtual reality simulation in which we can rehearse threatening situations, even if we don’t remember the dreams. Presumably, this rehearsal would lead to better real-life responses, so the rehearsal is adaptive.

Evidence supporting this comes from the large proportion of dreams which include a threatening situation (more than 70 percent in some studies) and the fact that this percentage is much higher than the incidence of threats in the dreamer’s actual daytime life.

Many have dreams where they struggled with difficult experiences beforehand, such as starting a new job, or school, meeting new people, and so on. Dreaming ahead of time prepares us for the situation as it carves the brain’s neural network (i.e. trains it) to be ready to act and respond in those situations.

Motorcycle racers, for example, day-dream the race track in order to learn it ahead of time, generating actual memories which would then trigger responses when they actually take that S-curve at 200mph. In that sense, daydreaming serves the same purpose as REM sleep dreams, as it allows the dreaming person to learn about a (dangerous) situation in relative safety and build his internal neural network to better cope and survive the real ordeal.

In that sense nightmares wake you up as they are synthetic data that went out of whack, and are not optimizing your network to better cope next time you occur a threat, so the brain wakes you up to avoid overloading the system.

Do You Think That’s Air You’re Breathing Now?

One question remains. If our brain is so good at generating virtual reality, how can we tell whether we are in the real world, or the dream world?

The answer is somewhat troubling. We can’t.

You see, the entire facade we call reality is in fact a construction generated by our brains (some would say our memory system).

When you are going about your business in the real world your brain generates what we refer to as experiences. However, those experiences are not actual reality.

The Ding an sich (german for “thing-in-itself”) as Kant described it, is the underlying object or objects in reality. That reality, which we assume our brains store in our memory system. But it doesn’t. That is, it doesn’t store the thing-in-itself, but rather a generated version of it. Such generated versions may suffer from many kinds of biases, causing the resulting facade (of reality) to be conceived quite differently by two people witnessing the same sequence of events.

The good news is that even if we cannot experience actual reality, or for that matter know for sure we actually do, we as a species have thrived and survived using these neural mechanisms.

So … dream on.

* “The Interpretation of Dreams” (German), is an 1899 book by Sigmund Freud