Sunday, 10 August 2008

Straw Dogs and the Bandwidth of Consciousness

Straw Dogs by John Gray is an interesting read. The challenging statements come thick and fast, so that the effect of reading it is somewhat like being in a pillowfight with an angry philosophy professor, with your adversary landing most of the blows. Its full of paragraphs like this:
"Humans think they are free, conscious beings, when in truth they are deluded animals. At the same time they never cease trying to escape from what they imagine themselves to be. Their religions are attempts to be rid of a freedom they have never possesed. In the twentieth century, the utopias of Right and Left served the same function. Today, when politics is unconvincing even as entertainment, science has taken on the role of mankind's deliverer."
I want to write a longer review of the book at some point, but in this post I want to talk about some interesting claims Gray makes about Human consciousness:
"... If we do not act in the way we think we do, the reason is partly to do with the bandwidth of consciousness - its ability to transmit information measured in terms of bits per second. This is much too narrow to be able to register the information we routinely receive and act on. As organisms active in the world, we process perhaps 14 million bits of information per second. The bandwidth of consciousness is around eighteen bits. This means we have conscious access to about a millionth of the information we daily use to survive."
My first reactions when reading this were:
a) did I read the numbers correctly? 16 versus 14,000,000?
b) no-one really has any idea what consciousness is, how can anyone measure the bandwidth of it?
c) is that bits as in 'little chunks', or bits as in binary bits?
d) is he making those numbers up?

After further research, it turns out the answers were
a) yes
b) apparently they can
c) binary bits
d) no

The trail of references

The Consciousness Bandwidth numbers are quoted without references or footnotes, but at the back of the book, in the Further Reading section, it says 'I am grateful to Vincent Deary for information on the bandwidth of consciousness'.

Vincent Deary is a research scientist specialising in cognitive behavioural therapy. I sent him an email and he returned a very helpful reply. It turns out that he emailed John Gray some time ago on the subject, but the numbers are not from his own research:
The figures are based on the body of research, begun in the fifties, that tries to link cybernetics and information theory to human consciousness. For instance Professor Manfred Zimmermann has a chapter in Human Physiology (Springer Verlag, 1986) called Neurophysiology of Sensory Systems in which he argues that "the maximal flow of the process of conscious sensory perception is about 40 bits/sec, many orders of magnitude below that taken in by receptors....our perception then would appear to be limited to a minute part of the abundance of information available as sensory input."
He also directed me towards Tor Norretranders book, The User Illusion, which has a whole chapter on the bandwidth of consciousness.

The User Illusion

Norretranders book (subtitled 'cutting consciousness down to size') turns out to be something of a classic in the field.

Chapter 6 of 'The User Illusion' does a great job of explaining the research into the bandwidth of consciousness. Norretranders also points out lots of simple observations that make the numbers much more believable.
"Precisely because from one instant to the next consciousness can switch from one object to another, it is not perceived as limited in its capacity. One moment you are aware of the lack of space in your shoes, the next moment of the expanding universe. Consciousness possesses peerless agility. But that does not change the fact at any given moment you are not conscious of much at all.

Right now you may be aware of the words on this page, or your posture, or the phone call you are expecting, or the room you are sitting in, or the situation in Central Europe, or the noise in the background. But only one thing at a time. You can switch back and forth between events, processes, and facts that are widely disparate in time and space. The flow of what goes through your consciousness is limited only by the scope of your imagination. But there are limits to the volume if flow at any given moment, even though the next moment something quite different may be passing through."
Soon after Information Thoery was first established as a discipline, researchers started using it to measure conscious actions. For example Norretranders mentions the following experiments:

W R Garner and Harold W Lake "The Amount of Information in Absolute Judgements" - Psychological Review 58 (1951) - they attempted to measure people's ability to distinguish stimuli (such as light and sound) in bits. Result: 2.2 to 3.2 bits per second.

W E Hick "On the Rate of Gain of Information" - Quarterly Journal of Experimental Psychology 4 (1952) - this experiment measured how much information a person could pass on if they acted as a link in a communication channel. That is, faced with a series of flashing lights, subjects had to press the right keys. Result: 5.5 bits per second.

Henry Quastler "Studies of Human Channel Capacity" - Information Theory, Proceedings of the Third London Symposium (1956). Measured how many bits of information are expressed by a pianist while pressing keys on a piano. Result: 25 bits per second.

J R Pierce "Symbols, Signals and Noise" (Harper 1961) - used experiments involving letters and symbols. Result: 44 bits per second.

Norretranders also mentiones a couple of studies by Karl Kupfmuller and Helmuth Frank which reviewed and summarised the earlier experiments. Then, after the early 1960s, no more major experiments seem to have been done.

Although the experiments produced different results, none of them produced any numbers higher than 50 bits per second for the 'bandwidth of consciousness'. When you look at the total amount of information that is transmitted into the brain each second, it is clearly much more than 50 bits. Norretranders writes:
"We can measure how much information enters through the senses. We do so simply by counting how many receptors each sensory organ posseses: how many visual cells the eye has, how many sensitive points the skin has, how many taste buds the tongue has. Then we can calculate how many nerve connections send signals to the brain, and how many signals each connection sends a second.
The numbers are vast. The eye sends at least ten million bits to the brain every second. The skin sends a million bits a second, the ear one hundred thousand, our smell sensors a further one hundred thousand bits a second, our tasts buds perhaps a thousand bits a second."
So that gives a rough total of eleven million bits per second total input into the brain - compared to an apparent information processing rate within consciousness of less than 50 bits per second.

Norretranders joins Karl Kupfmuller in suggesting that the sub-conscious parts of the brain have to do a tremendous amount of work to process and order the raw data coming from the senses into the 'content' on consciousness. The number of neurons and the links between those neurons suggest that the brain has plenty of capacity for this: Kupfmuller suggests a very conservative estimate of ten billion bits per second.

So we have eleven million bits of data flowing in, a massive, un-conscious processing job running at at least ten billion bits per second to sift through the data, and then a tiny, less-than-fifty bits per second consciousness running on top. Connections back out to the motor organs are reckoned to be roughly the same number as those coming in from the senses, so as we head from the brain back to the outside world, the bits of information get back to the millions again.

Is consciousness crap, then?

So, all of this shows that John Gray's 'bandwidth of consciousness' numbers in Straw Dogs are well founded. Although he didn't give the sources, there are plenty of experiments to back up the numbers. There are some questions that could be raised about the results of the experiments - more on that below - but we'll put those aside for the moment.

The 'spin' that Straw Dogs puts on the numbers is very questionable though. After quoting the numbers, the book goes on to say:
"... The bandwidth of consciousness is around eighteen bits. This means we have conscious access to about a millionth of the information we daily use to survive."
The numbers are correct, but evoke an image of a tiny drip of random information into consciousness as an avalanche of unconscious information streams past. To me, this section of Straw Dogs subtly suggests that consciousness is sidelined, and just fed scraps of information.

But Norretranders book paints a different picture - the amount of information that enters consciousness is comparatively tiny, but a huge amount of (unconscious) processing takes place to decide which bits will enter consciousness. Furthermore, we have a large degree of control over which bits enter our consciousness at any moment. So there is a massive summarising and editing process taking place every second, coupled with our conscious ability to choose which information streams we want to be most conscious of.

Straw Dogs suggests that the information entering consciousness is paltry and insignificant, whereas Norretranders book shows that the information entering consciousness is intensively selected and condensed. In short, if a live tiger were to pop up from behind your sofa, your brain would make damn sure that you were conscious of it. Any information that is obviously important does not very often slip past, despite the low bandwidth.

The section of Straw Dogs that deals with consciousness also draws on two other main sources to try and show that consciousness is overrated: Benjamin Libet's famous experiments on conscious volitional acts, and the alleged effectiveness of subliminal advertising. Both these are big subjects in themselves, hopefully I can post something about them someday.

Questioning the Experiments

Norretranders states that no major research has been done on the 'bandwidth of consciousness' since the original experiments in the 50s and 60s. For a field that produced such surprising results, its strange that it has not been visited more. The lack of recent research seems to suggest that the matter is settled, but I'm not sure if thats the case.

Some of the questions and objections people have raised are mentioned by Norretranders:

Richard Gregory, in 'Oxford Companion to the Mind' (1987) pointed out that the people in the bandwidth experiments were still conscious of other unrelated things while they were doing the experiments. For example, the person seeing the flashing lights and pressing the buttons in the W E Hick experiment will also have been hearing things, seeing other things except for the flashing lights, and so on. It would be impossible to design an experiment that blocked off all those other avenues, or measured them all in some way.

More interestingly, the experiments do not really take into account that after a learning period to get used to a processing-intensive task, our consciousness can delegate the process to an 'automatic' subsystem. Driving a car is the classic example; once you have learnt to drive, you can happily navigate your way down a motorway and have a conversation with your passenger at the same time. Norretranders mentions a paper called "Skills of Divided Attention" by Spelke, Hirst and Neisser in 1976, in which people were asked to to read stories while simultaneously taking dictation. Reading and writing unrelated things at the same time is hard at first, but after a few weeks practice, some of the subjects of the experiment could read while writing, and do it just as fast as their normal reading speed.

Norretranders interprets this by saying that the experiments have probably over-estimated consciousness. For example, when measuring the processing capacity of someone playing the piano, the experiments assume that the processing is being done consciously. But if they are a very well practiced piano player, it may be the case that their consciousness is hardly involved in the playing at all, and it is all handed over to an 'automatic' sub-system.

He is right, in a sense. If you want to measure the bandwidth of consciousness, you should restrict the experiment to things that are definitely happening in consciousness. But, on the other hand, someone driving in traffic or playing the piano is not performing the task completely un-consciously either.

Imagine: you are driving across a city and engaged in a conversation with a passenger at the same time. A lot of the driving will be done automatically, without your consciousness being 'aware' of it. However, it is not a totally un-conscious process either; compared to, say, the regulation of your heart-beat, or the work your visual cortex does to process the incoming nerve signals. You are still in control of the process - you know where you are driving to. Also, if something unexpected happens, the process will jump back up into your awareness, and the conversation with your passenger will be put on hold temporarily (imagine reaching a junction where the traffic lights are unexpectedly out, and there is a policeman directing the traffic).

I'd say that we need a new term for this sort of process. It is not fully taking place in consciousness, but it is not un-conscious either. For want of a better term, lets label it a semi-conscious delegated process. The experiments mentioned by Norretranders, and the numbers quoted in 'Straw Dogs', do not take account of the fact that we can learn to multi-task, and substantially increase the processing that is under the control of consciousness, by performing tasks in a semi-conscious delegated process. All we need to do this trick is sufficient practice.

Content vs Output

Another big objection to the experiments is this: Most of the experiments measure the information processing output from consciousness, which might be different to the amount of information held within consciousness. Certainly experiments such as W E Hicks one, where a human become a link in a communication chain, can only measure the physical information output of a human. I don't have access to the original papers, so I don't know how the 'distinguishing' experiments done by Garner and Lake worked, so maybe this objection would not apply to them.

To imagine the difference between the bits per second in consciousness, versus the bits per second output of consciousness, think of it this way: you can watch a movie, and take in quite a lot of what is happening. Imagine a scene with two people talking in a room. You might be conscious of what they are saying, their facial expressions, how far they are from each other, what is in the background, and so on. Now imagine trying to give a running commentary of the movie to someone who could neither hear or see it directly. It would be very hard to keep up. You might manage the main bits of plot, but lots of dialogue and most of the scenery would get skipped or abbreviated. This simple situation suggests that our consciousness can take in must more information than it can put out per second.

Low bandwidth vision

I understand that if repeatable scientific experiments provide us with a counter-intuitive result, we should probably accept it anyway. And I'm as much of a heterophenomenologist as the next guy. But when I think about the 'bandwidth of consciousness' limit of 50 bits per second, and then think about the richness of the visual field in my consciousness, 50 bits per second just doesn't seem enough.

Similarly, as Norretranders points out, if we can only process less than fifty bits per second in consciousness, why do we have televisions that transmit four million bits per second, and telephones that can transmit at least four thousand?

The visual field that we experience in our consciousness seems very detailed. Admittedly, there are some tricks that the brain employs to make it seem more detailed than it really is: the detail sensitive fovea in our eyes is surprisingly small, so we move it around very rapidly with saccades to allow our brain to build up a detailed picture. But still, if you imagine it as pixels, there must be a lot of pixels there.

At the end of this sentence, close your eyes, turn your head away from the screen, open your eyes for about 1 second, and see how much you can take in.

Done it? How many bits of information do you think you became conscious of in that second? (It will depend somewhat on whether your eyes were focused to the correct distance).

Figure 1, below contains roughly 48 bits of information:

(Figure 1: 8 colours = 3 bits per square, and there are 16 squares)

What you saw when you opened your eyes for a second was (hopefully) a lot more detailed than Figure 1. That is, if what you saw were arranged into pixels, it would have a lot more colours and a lot higher resolution than that. On the other hand, if you looked at this simple image for one second, you would probably not be able to recall the exact pattern of coloured squares.

Another thing to take account of here, is that although our 'visual field' in consciousness seems very, well, 'visual', there is also a lot of structural information in it. When you look at a scene for one second, you do not just perceive it as a bunch of pixels - our brains usually divide the scene up into objects, and the objects are related to each other in a three dimensional structure. For example: "the chair is under the table, both table and chair rest on the floor, the mug is on the table, behind the mug there is a book", and so on.

To me, this suggests a whole extra level of information in our conscious 'visual field'. You don't just see a mug, you see where it is in a 3D structural relationship to the other objects around it.

For another comparison, Figures 2 and 3 below show two versions of the same photo: one of them reduced to roughly 48 bits of information, the other one comprised of hundreds of thousands of bits.

(Figure 2: approx 48 bits)

(Figure 3: approx 372,000 bits)

For sure, we are not conscious of every single one of the 65,000 pixels in Figure 3, but this comparison shows just how little visual information can be contained in 48 bits. Of course, the visual field in consciousness is not really made up of pixels like that, but the point is still valid - 48 bits is a very small amount of visual data. Can our consciousness really only handle a bandwidth of about 48 bits a second?

Flavours of Bits

One last thought experiment - imagine watching a 10 second film, in which a couple of deer, like the ones in Figure 3 above, wander through a wood. Then imagine a 10 second film in which a 4 x 4 grid of coloured squares, like Figure 1, is shown, and changes once a second. Which film would be easiest to recall details about?

You would probably be able to recall a fair amount of the deer film - how many deer there were, their sizes, which direction they were walking in, whether they looked scared or relaxed, roughly how dense the trees were, what the light was like. You probably would not be able to recall much about the abstract film with the coloured squares. Apart from that it was an abstract film with a 4 x 4 grid of squares that changed colour. You might remember something like 'at some point there were two red squares on the left' or 'a lot of the squares seemed to be blue'.

In short, you would be conscious of, and be able to recall, must more information from the deer film than from the abstract film. Why is that? Probably because our brains are not general purpose processing machines - they have evolved to deal with certain kinds of information, not information in general.

Information Theory, by definition, treats the 'bits' of information as generalised, abstract units. It simplifies situations by ignoring what the 'bits' actually represent. But our brains do not work like that - they have evolved to deal with certain types of information very well, and other types of information not so well.

My guess is that the experiments done 50 years ago underestimated how much information could be processed consciously by concentrating on abstract information. Light bulbs that flashed on and off, random sequences of letters, that kind of thing. Would experiments that used more 'natural' types of information lead to measurements of higher processing rates?

So taken together, there are some quite significant questions about these experiments:
- how did the experiments deal with the information that subjects were conscious of, but was not explicitly measured by the experiments?
- how did the experiments deal with the phenomenon that I have (clunkingly) called 'semi-conscious delegated processes'?
- did the experiments claim to measure the output or content of consciousness?
- how much measurement of the conscious 'visual field' was done by the experiments?
- was the information in the experiments 'abstract' or 'natural'?

Its probably safe to say that the 'bandwidth of consciousness' is a very small fraction of the total information input to the brain. But I think there's probably enough objections there to warrant some more modern experiments to measure the limits.


Pierre said...

I refer you to my blog:

Kris said...

Excellent post, Memebake.

I followed a reference trail of similar numbers from David Brooks' "The Social Animal", over Timothy D Wilson's "Strangers to Ourselves", and Tor Norretranders "User Illusion" all pointing to Zimmermann's work, before I ended up on your site because I didn't feel satisfied with the sources.

Whereas the estimate of the information present in sensory input seems reliable, the main weakness of the 40 or 50 bit claim lies in one of the points you make: the 50s and 60s experiments measure the information rate of the output of consciousness, not of its internal representation. So the comparison of 11 million (which is in terms of representation) to 50 (which is in terms of output) is essentially meaningless.

Seems the early cyberneticists got carried away and used information theory in an inappropriate way.

Well done for fleshing this out. If all the above authors had done the same, they wouldn't be repeating claims that don't seem scientifically sound. Or perhaps the exact numbers don't matter too much, because the difference between conscious and nonconscious processing is indeed many orders of magnitude.

Richard said...

I still can find no evidence to support the idea that we can absorb more than around 10 bits per second of novel information, however we are able to visually recognise anticipated sensations with an information rate of several Megabits per second. I explain this and more in my new book: Bottleneck -Our Human interface with Reality,
In the User Illusion Tor maintains that we subconsciously absorb information much faster than we do consciously, but I maintain that we can merely recognise at at higher rate.