"Representation" in neuroscience|
Aug 5, 2013, 12:19p - Science
The word "representation" is used pervasively in neuroscience. Here's an example:
"The neural activity in visual cortex represents the visual input received by the eyes."
I've found this word troubling since the very day I started grad school, and I had hoped that my troubles would abate as I learned more in the last 6 years. Alas, they have not, so I now feel compelled to commit my troubles to my blog.
So why do I find this word, this concept, of "representation" in the brain, so troubling? For those not familiar with neuroscience, let me explain how the word is used a bit more.
The canonical neuroscience experiment goes something like this. Let's say you're interested in understanding how the brain senses touch to the back of the neck. The standard method would be to take a mammal (say a monkey or a mouse), strap it down, and insert an electrode into random locations in the brain. Then, you would touch the back of the monkey's neck, and then check for a response on your electrode. You wouldn't see any correlated activity, so you'd move on to a new brain location. This fishing expedition would eventually lead you to a part of what we call "somatosensory cortex" which would reliably respond when you touch the back of the monkey's neck.
The simple thing to say, now, about this area of somatosensory cortex is something like this:
"Area X responds when the back of the neck is touched."
or after further experiments, something like
"Activity in area X correlates with touching of the back of the neck."
Pretty straightforward, and no one can argue with that. However, that doesn't sound so sexy, and given that neuroscience has not yet rooted out the many sophisticated charlatans that still occupy its realm, at some point some neuroscientists decided to say something different:
"Area X represents touch to the back of the neck."
So what's the difference between using the words "respond" and "correlate" vs. the word "represent"? And why does it matter?
The difference is that "respond" and "correlate" are conservative interpretations of the observations made during the experiment, while "represent" attempts to interpret what was observed with greater force, as follows.
I dislike the word "represent" for 3 reasons:
(1) "Represent" implies causality, because it implies that the brain is using that very signal for another purpose. But all we observed was that area X responds to touch. The signal in area X could be important for a reflexive withdrawal of the neck, or the feeling of touch to the neck, or nothing at all, a proverbial "dead end" in the neural circuit. But to distinguish these possibilities we would need to do further experiments, most important of which is to touch the back of the neck while preventing area X from being active, and then observing whether the reflexive behavior is intact, or asking the monkey or person whether they still feel the touch. Until these causal experiments are done and positive results obtained, we have no business saying that a specific set of neural activity represents anything at all.
(2) When I think of the word "represents", I think, to whom? Representation has a lot to do with symbolism, as in "the rose represents my love for her". But in order for such symbolism to be understood, there needs to be a receiver on the other end who can interpret the symbolism. That means that there needs to be a mind to make sense of the representation. But who is the observer in the brain? If the brain creates the mind, then how can the mind internally observe the activity of the brain which creates it? One could make an argument that parts of the mind can observe other parts of the brain, but isn't this just another form of dualism? And don't we reject dualism as scientists, given that there is no convincing evidence for it? (ignore my posts on the soul for now - this is my other mind speaking) "Representations" in the brain also invoke the concept of the homunculus, a "little man" in the brain who is watching all that the sensory nerves provide. The homunculus is problematic because it leads to infinite regress: what part of the brain constitutes the homunculus, and then who's "watching" that area?
(3) When is a signal just a signal, and when does it become a representation? My worms avoid light, and I've found neurons that respond to light and that control the worm's behavioral response to light. Would I say that the activity of this neuron "represents" light? I don't think so, and I'm not sure most neuroscientists would feel comfortable with this use of the word. I would just say that the neuron responds to light, and this signal controls the worm's behavior. No need to invoke the concept of representation at all. Somehow neuroscientists want to invoke this word when more complicated sensory processing is involved (e.g. not just the presence or absence of light, but the identification of an object using light), but this difference doesn't seem very well-defined to me.
I guess I'm more conservative than your average neuroscientist. I dislike the word "representation" because it's vague, it implies function even when no evidence has been presented, and it invokes traditional dualistic mind-body views that science seems to have banished to a more primitive age.
I think we should no longer use the word "representation" but instead be more literal about the signal processing that the brain seems to do. We don't take representations seriously when programming software or designing electrical circuits, and I see no reason to use the word when dissecting the neural circuits of the brain.
Ultimately I find the word "representation" misleading and responsible for perpetuating greater confusion around the mind-body problem.
Read comments (4) - Comment
« Lab pics on MIT CAST
Consciousness as the only true emergence »
- Aug 5, 2013, 9:29a
I don't know. I think, especially when discussing computational models, 'representation' is particularly useful. Say that someone was coming up with a theory for how the visual system works, and had drawn up a computational model that has images being encoded simply by the greyscale value at each pixel. Here, the pixel representation is the wrong representation, because any slight perturbation of the image will drastically shift the response vector in high dimensional space. You need a model that is invariant to size/rotation/translation, because that's how we know our visual system works on a behavioral level. The next model would seek to account for this, and neurobiological experiments could later be performed to validate the model.
To me, 'representation' has little to do with function and everything to do theory—that is, what types of operations we might expect the cells/systems of interest to perform. How we imagine such stimuli being 'represented' in the brain. Thinking in terms of the representation problem allows for good, hypothesis-driven work towards confirming/rejecting these models.
- Sep 21, 2013, 1:04a
Bravo - err, you do know that Skinnerians have been saying similar things for 60 years, right?
- Sep 14, 2015, 7:47a
I just started my neuroscience career, and I felt confused when I met the word "representation". Why don't these neuroscientists use another word?
- Apr 20, 2017, 10:16a
Maps represent the world, right? Representations are just the internal maps by means of which we steer (this is from Dretske). There is no implied dualism.
It seems innocuous enough, there is no dualism implied, any more than when I use a map to get around town.
It is a representation, not the world itself, upon which your behaviors are based. E.g., the representation of the world in V1 is what you base your decision to take your foot off the brake pedal, for isntance. You don't have magical direct access to the stop sign. It is indirect, mediated by sensory channels.
OTOH, if you don't like the word, you could use another word. Visual information, visual signals, etc.. They are pretty much all used synonymously, and people aren't too worried about words as much as the biological mechanisms and explanatory purchase offered by the discovered mechanisms.
When systems are really simple the language of representation is not used much, because organisms more directly coupled to the environment (e.g., venus fly traps are never discussed in such terms). But when layers of information processors are interleaved between stimulus and behavior (e.g., most vertebrates, and many vertebrates), it is more likely to be used because it is natural and useful and innocuous.
I'm not so sure we should be going back to Skinner, Glen. Just sayin'. :)