Highlights and comments on Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts, Stanislas Dehaene, 2014. ISBN 978-0670025435.
Consciousness and the Brain is a book about Dehaene’s global neuronal workspace (GNW) theory, the experimental evidence that underlies it, and what it entails. The GNW is an attempt to explain how consciousness arises from brain organization and behavior.
I’m freely mixing in some comments, so don’t necessarily assume everything in here was in the book. I do try to make it relatively clear when I insert my ideas.
The history of consciousness studies is full of pseudoscience and general nonsense. Dehaene acknowledges this fact at the beginning of the book, and as a reaction consciousness was banished from serious psychological and psychiatrical research for most of the twentieth century, in favor of behavioralist analysis, which took a strictly mechanical view of the brain. Yet advances in probing methods available to science, and a more rigorous approach to reported subjective experiences, reintegrated consciousness at a central place in neuroscience research.
This book understands consciousness as conscious access, the fact that the mind’s object of focus becomes available to the entirety of the brain’s systems. It is contrasted with vigilance, essentially the state of being not-asleep, and attention, the ability of the brain to focus its activity on something in particular. While the conscious access–consciousness equivalence can be debated, I think the book’s insights are valuable even if you don’t agree with that distinction. In fact, in order for humanity to become able to decide whether this equivalence is good, we probably need more data of the kind this book provides.
What makes a state of mind conscious or not? Dehaene exposes several experiments have been proposed to probe at this conscious-unconscious boundary.
One of the earliest such experiments is that of binocular rivalry: each of a subject’s eyes is presented with a different image. While this stimulus is completely static, the conscious perception of the image flickers back and forth between one and the other. FMRI results suggest that, while lower levels of the visual cortex accurately reflect the reality, with each half of it generating a representation of the corresponding image, changes in the perceived image match up with changes in the patterns shown in the prefrontal cortex. This is the first of many pieces of evidence that there exist specialized areas of the brain that can act subconsciously (like the visual cortex), and others that correspond to consciousness in some way (like the prefrontal cortex).
[How valid is fMRI? This is an important question, because fMRI experiments underlie many of the book’s results, and fMRI has been the subject of harsh criticism, in particular as dead salmon have confounded common statistical method. The book does however spend a lot of time discussing alternative interpretations for various experiments, and corrected experiments that have been performed to validate or disconfirm these interpretations, most of which seem to have succeeded in replicating the original effects.]
Another related experiment is that of subliminal images. In this experiment, subjects are shown images, frequently either words or faces, very briefly (~50ms is typical). While the brain is capable of becoming conscious of an image flashed for a brief period of time, this ability disappears if masking is used: images of interest are flashed between other scrambled images. This type of experiment reveals a number of things.
First, we find a relatively sharp conscious–unconscious threshold: below 40 ms of exposure, almost nobody will consciously perceive images being flashed, whereas above 60 ms, almost everybody will. This difference is self-reported by subjects, not determined by fMRI or another objective method. Dehaene argues that not only is this acceptable, it is in fact critical to understanding consciousness: we can’t reason fully from first principles, and, at some point, we have to correlate the experience of consciousness with its manifestations.
Second, we observe the same visual cortex–prefrontal cortex distinction as in the binocular vision experiments. The prefrontal cortex activates only when conscious access is reported, while the visual cortex activates at the sight of an image that is subliminal, but analyzable (such as a real face or English word).
Third, the subconscious mind is able to perform advanced levels of computation. In one example, subjects are presented with a masked subliminal word (say, ‘radio’), then with a non-subliminal word that can either be the same (‘radio’ again) or different (‘house’). Subjects are faster to recognize the second word when it was first flashed subliminally, even when the two appearances used completely different typography, which suggests the subconscious process could recognize the words at a high level of abstraction. In fact, subconscious processing can be more complex: people’s mood states are affected by subliminally flashing words with negative connotations. Subjects are also able to perform various tasks, such as evaluating and add up small numbers, subconsciously.
In fact, it seems that most neural circuits can be activated subconsciously, and perform significant amounts of work in a subconscious state. There is even mounting, but not yet very conclusive, evidence that the unconscious brain can perform complex operations over time, which can manifest as insights after a good night’s sleep. This is much like what Rich Hickey suggests in his Hammock-Driven Development talk, and is something that matches my personal experience very well (it might be the primary reason why cramming for exams last minute doesn’t work as well as I’d like).
Overall, the interactions between the subconscious and conscious mind seem very complex. Subconscious operations can reach a high level of sophistication, but their impact to cognition overall is limited, especially in time, unless they are either kicked off or driven by the conscious mind. Most specialized circuits can (at least partially) function in a subconscious manner.
I’m amazed at how efficient these subconscious computations seem to unroll. Firing rates are very vaguely defined, but it seems like a given neuron can’t send a signal with a space of much less than a few milliseconds between two firings. Subconsciously recognizing an image seems to take at most about 200 ms, or a maximum of ~100 firings periods! This is an impressive result compared to the performance of current neural networks. However, I mostly find references to this in CS papers1, not in neurosciencey ones, so it might not be quite right.
The Global Workspace Theory
The Role of Consciousness
With this groundwork laid down, Dehaene delves into the question of the role of consciousness (from an evolutionary standpoint).
First, neurons in the visual cortex can only see the output from a small region of the visual field, but global information is critical to understanding what we’re seeing. If their part of the visual field is ambiguous, neurons are initially confused, but settle on a (generally correct) interpretation after communicating with other visual field neurons. This resolution does not occur under anesthesia. This, and other experiments, suggest that the unconscious parts of the brain simultaneously entertain many possible interpretations of the world, which the conscious mind samples them.
Second, long term memory, as well as performing long sequences of operations or following complex rules, seem to only be possible with the help of conscious access. In particular, recognizing complex patterns that are not tightly concentrated in time only seems possible in a conscious state.
Third, the conscious mind lets us represent abstract concepts and helps us share information efficiently about the outside world and our state of mind to other humans. (In fact, some people have theorized that this was the primary role of consciousness, or that we were essentially unconscious machines with a veneer on consciousness added on to make our experiences shareable. Dehaene disagrees, and so do I.)
So, what patterns should we be looking for in the brain’s activity to detect the signs of consciousness? Here are a number of criteria we would like to fulfill:
- the brain activity in question should be correlated with a subjective perception of consciousness,
- it should encode the conscious thought; in particular, two similar thoughts should have similar manifestations, and different thoughts have different manifestations,
- we should expect it to be more stable than the sensory inputs—our senses are very noisy, but conscious experience is stable over time horizons of several hundred milliseconds,
- inducing a similar brain activity pattern should cause a conscious experience, and inhibiting it must prevent conscious experience from arising.
I had heard of brainwaves before but had never paid much attention: the name sounds like it’s right out of a sci-fi movie, and I assumed it was just a bad name for any electrical activity in the brain. I was wrong: the brain does in fact show some specific wave-like patterns of activity. These work a bit like the clock in a CPU: they are the sign of neurons synchronizing their activities to fire in coherent patterns.
Brainwaves can be captured by an assortment of technologies. In addition to fMRI, which gives fine images of the brain, Dehaene mentions we can also use electroencephalography (EEG) and magnetoencephalography (MEG) for coarser-grained data.
Armed with these tools, Dehaene says we can detect four main brain patterns that correlate with consciousness.
- We’ve seen before that unconscious and conscious stimuli both activate the brain, and can do so relatively deeply. But while this excitation is short-lived for unconscious stimuli, when the stimuli is above the threshold of consciousness, it will persist longer and start amplifying itself.
- This amplification results in a positive P3 “wave” (more of one or two voltage peaks than a prolonged wave) propagates throughout the brain.
- After this peak, gamma waves (of frequency ~40 Hz, which is on the upper range for the brain) becomes much stronger throughout the brain. Experiments suggest that this is a sign of different networks synchronizing.
- After the initial synchronization period, data is exchanged bidirectionally on many channels within the brain. Causality analysis shows that data flows back and forth from neocortex to other areas of the brain.
Overall, we observe a sort of ignition in the brain, where most long-distance connections become saturated with the conscious thought, which helps explain attention blindness.
Finally, the fact that modifying these electrical signal with things like trans-cranial stimulation disrupts consciousness suggests that these signatures of consciousness play an active role in it, instead of being merely signs of it occurring.
A Global Workspace
This is the key point of the book, and it’s pretty simple! Putting all this evidence together, Dehaene argues that consciousness of a piece of information is simply that piece of information reaching a special area of the brain, from which it can be shared with other areas. This is the global neuronal workspace.
This workspace processes thoughts in a relatively slow, sequential and deliberate way, characteristic of conscious access, and which contrasts with the rapid, parallel and automatic computation of the brain’s periphery.
But what's the difference between "neural" and "neuronal"? Some knowledge is not meant for man, I guess.
How are thoughts encoded in the global workspace? A hypothesis is that a small subset of neurons will be active at a time and encode a given idea. The rest are inhibited during conscious access of this idea. The large P3 wave we observe is actually the electric signature of this vast majority of inhibited neurons sending inhibitory signals in a synchronized manner.
The brain’s anatomy provides some support for this view: there are long-distance connections between specialized areas and the prefrontal cortex (the proposed location of the global workspace).
Dehaene also mentions computer simulations of the global neuronal workspace show the signatures of consciousness mentioned above, including an avalanche-like effect when a thought suddenly becomes conscious. The model also exhibits spontaneous activation, which, he says, could provide a basis for wandering thoughts and the important ability of human brains to imagine new thoughts or situations.
Neural Networks and AGI
Dehaene mentions several times what he thinks is missing from current computers to reach general intelligence. The gist of his argument is that contemporary computers have a number of very specialized programs that cannot communicate together in a flexible way, and that a key element of an AGI-capable system would be a sort of “super-clipboard”, to which different systems can contribute and that, presumably, a prefrontal-cortex-like structure would control. This analogy is a little jarring because contemporary computers are not really trying to be AGIs. However, I think the same argument applies to current deep learning models, which do aim towards that goal.
Like Dehaene and many others, I don’t believe the current generation of deep neural networks will gain general intelligence. I think a massive change in architectures is needed, one that is at least of the scale of the previous deep learning revolution, but likely much greater. Current networks seem to require ever larger amounts of data for a much less than proportional decrease in performance and brittleness, and my experience with the field is that some of the best results are suspiciously hard to reproduce, or tied to a particular dataset’s idiosyncrasies2
While it’s very unclear how we would go about implementing one, I wouldn’t be surprised if a GNW-like system turns out to be extremely helpful in achieving better adaptability of neural systems. It wouldn’t seem too unlikely that current-generation neural networks will be able to find some use in a future GNW-based AGI, as the lowest levels of the various specialized cortices.
The book mentions some applications of GNWT to clinical cases, some of which serve as further evidence. While this part is interesting, a lot of it is not that tightly related to the main part of the book. Dehaene and others have developed predictive models and tests that have some accuracy at predicting the outcomes in vegetative, minimally conscious, or locked-in patients based on behavioral tests interpreted in the light of the GNWT.
The Future of Consciousness
The last part of the book is a bit of a grab-bag of related ideas. Dehaene discusses how various apes and baby humans show similar neural behavior patterns. (Presumably, this is the case for many mammals and maybe even many vertebrates; there are studies on the global workspace for cephalopods3).
Perhaps more importantly, he argues that the hard problem of consciousness, i.e. the problem of explaining how brain behavior gives rise to qualia, “just seems hard because it engages ill-defined intuitions,” and will disappear once we understand the brain better.
How much do I agree? There are several prior examples of hard philosophical problems more or less dissolving in scientific advances, but, even as a materialist, I don’t feel confident that the hard problem of consciousness can be solved through non-metaphysical means.
Overall, this book presents tons of satisfactory evidence, and an interesting and cohesive take on the question. It obviously leaves much unanswered (I would be particularly interested in more details on the role of memory in consciousness), I feel fairly confident that it is reducing the wrongness.
What’s happened since Consciousness and the Brain?
“But wait”, you say, “we do live in 2014’s future!” So, what’s happened since the book came out?
A lot of research about the global (neuronal) workspace was written between 2000 and 2010; the subject, at least its foundational part, seems to have partly died off since.
A concurrent of the global neuronal workspace theory is the recurrent processing theory. Unlike the GNW, recurrent processing theory downplays the importance of the prefrontal cortex’s centralization, and asserts that the key to consciousness is in fact the backwards connections from higher to lower levels of the many specialized circuits. In this theory, conscious access is not even necessary for consciousness! The debate about that is still alive.
Jeff Hawkins’ more recent (2018–2019) and high-profile thousand brains theory is an interesting variation on the recurrent processing theory.
- The brain contains many specialized circuits which can act fully subconsciously.
- These circuits involve many layers and feedback connections from the top to the bottom layers.
- The prefrontal cortex contains a centralization and dispatching unit, connected to other parts of the brain via long-distance connections.
- This unit can hold one thought at a time and has a relatively low bandwidth to the rest of the brain.
A good, free read by the author if you don’t want to get the book is Dehaene, Stanislas, et al. “Conscious, preconscious, and subliminal processing: a testable taxonomy.” Trends in cognitive sciences 10.5 (2006): 204-211.
Mather, Jennifer A. “Cephalopod consciousness: behavioural evidence.” Consciousness and cognition 17.1 (2008): 37-48. ↩