麻豆传媒

DURING the second world war, aircrews who had to calculate mission routes and bomb trajectories found that their instruments 鈥� mechanical computers packed with cogs and gears 鈥� performed better in the air than on the ground. Realising that the plane鈥檚 vibrations were helping to make the instruments鈥� sticky moving parts move more freely, engineers began building small vibrating motors into them to make them more accurate. This was one of the earliest applications of dither, or the deliberate addition of noise.

Noise is usually a nuisance, as anyone who lives under a flight path or has tried to listen to a distant AM radio station can testify. But to engineers it can be a godsend, and now its benefits are cropping up in biology, too. More than a decade of research suggests that under some circumstances, a small injection of noise can sharpen up the way in which an organism senses its environment. For example, crayfish are better at detecting the subtle fin movements of predatory fish when the water is turbulent rather than still. Humans are better able to recognise a faint image on a screen when a dash of noise is added to it.

In these cases the noise source is external to the organism, but they raise an intriguing possibility: could evolution have beaten the wartime engineers to it and incorporated dither into the brain itself? A group of neuroscientists is now claiming to have found just that, in the form of neural circuits that are 鈥渘oisy by design鈥�. If they鈥檙e right, it may be that dither is a common feature in nature.

A working definition of noise is that it is a broadband signal containing a jumble of frequencies 鈥� the hiss of white noise, for example, is made up of the full range of audible frequencies, from very low to very high, in equal amounts. In contrast, meaningful signals concentrate their energy on a comparatively narrow band of the spectrum.

The phenomenon of noise increasing the detectability of a faint signal is called stochastic resonance. Stochastic resonance applies specifically to non-linear systems, where the output is not proportional to the input. Neurons are a good example of a non-linear system, firing only when the electrical potential across their membrane reaches a critical threshold. In such a system, a weak input which fails to reach the threshold can be lifted above it by the injection of noise.

Numerous theoretical models suggest that stochastic resonance could improve how neurons process signals, and there is good experimental evidence that adding external noise can enhance the brain鈥檚 abilities under certain circumstances. Stochastic resonance explains why water turbulence helps a crayfish鈥檚 sensory hair cells detect a distant fin movement, and why noise helps the human eye to pick out a faint image. External noise has since been harnessed to enhance human performance, for example, in cochlear implants to help pick up faint sounds and in vibrating insoles that reduce swaying in people who have suffered strokes (麻豆传媒, 2 November 2002, p 22).

Till now, however, no one has found any evidence that the brain generates its own internal noise to exploit stochastic resonance. That is where the work of Gero Miesenb枚ck, a neuroscientist at the University of Oxford, comes in. Miesenb枚ck thinks he has found a brain circuit, part of the olfactory system of the fruit fly Drosophila, that exists specifically to generate noise and thus enhance brain function. He says his discovery has implications for the human brain because the basic architecture of the Drosophila olfactory system is common not only to all insects but also to all vertebrates 鈥� including humans.

Miesenb枚ck didn鈥檛 set out in search of noise. He was trying to solve a mystery that has troubled olfactory-system researchers for many years.

The fly olfactory system is a huge piece of neural circuitry (see Diagram). It starts in the fly鈥檚 antennae with around 1200 olfactory receptor neurons (ORNs), each of which carries a single type of odour-receptor molecule. There are about 60 different receptor molecules and hence about 60 different types of ORN.

From the antennae, these odour-specific ORNs converge on nodes called glomeruli where they make synaptic connections with cells called projection neurons. Each glomerulus receives inputs from only one type of ORN, so for a long time neuroscientists assumed that each projection neuron would only respond to a single odour.

But a few years ago, neuroscientists discovered that this is not the case (). Electrical recordings from individual projection neurons show that they sometimes respond to odours other than those picked up by their ORNs.

But how do they do this, when each glomerulus receives inputs from only one type of ORN? While at Yale School of Medicine a few years ago, Miesenb枚ck and his colleague Yuhua Shang managed to solve this puzzle.

They took a mutant fly in which all the ORNs connected to a particular glomerulus were missing, and looked for other inputs to the projection neuron. What they found was a previously unknown network of 鈥渋nterneurons鈥� connecting the glomeruli to each other and transmitting activity between them (). These 鈥渆xcitatory local neurons鈥� seem to provide a sort of diffuse, stimulatory input to the projection neurons whenever an odour is present.

That solved the immediate problem, but raised another: why add something to the system that means losing the exquisite one-to-one mapping of odour receptors to projection neurons? 鈥淚t seems counter-intuitive,鈥� says Miesenb枚ck. 鈥淲hy would you take the crisp, sharply separated input and blur it out, make it noisier?鈥� The hypothesis he came up with was that the noise was there for a reason. Perhaps the excitatory local neurons deliberately inject noise into the system, taking advantage of stochastic resonance to make faint odours easier to detect.

鈥淚t seems counter-intuitive 鈥� why would you take a crisp input and blur it out?鈥�

Fine-tuning

This makes sense in the light of what subsequently happens to the sensory input signal. Projection neurons send signals to other neurons called Kenyon cells in a structure called the mushroom body, a part of the fly鈥檚 brain involved in learning and memory. Each Kenyon cell receives inputs from many projection neurons, but they have extremely high firing thresholds and are only activated when a large number of their incoming neurons fire simultaneously. Given that projection cells fire more readily in response to their own odour than others, each Kenyon cell only fires in response to a single odour and the system recaptures specificity.

Miesenb枚ck鈥檚 group also came across a 1983 paper by Alexander Borst of the Max Planck Institute of Neurobiology in Martinsried, Germany, describing a network of inhibitory local neurons linking the glomeruli. Miesenb枚ck thinks these may have the opposite effect to his excitatory ones, damping down strong signals from ORNs.

So why bother to boost weak signals and tone down strong ones? Miesenb枚ck suggests this happens to iron out extremes in odour concentrations. 鈥淵ou need to be able to smell a rose, and identify it as a rose, at very faint concentrations and in full bloom, if it is held directly under your nose,鈥� he says. 鈥淭here has to be some mechanism that eliminates the variation based on odour concentration. We think that the middle layer of processing does exactly that.鈥�

Miesenb枚ck鈥檚 group still has some way to go to prove the 鈥渘oisy by design鈥� hypothesis, but they鈥檙e working on it. By tinkering with local neurons, they hope to learn how to change the volume of the noise. Miesenb枚ck predicts that turning it down or silencing it entirely will make faint odours less likely to trigger Kenyon cells. Another prediction is that the flies will become behaviourally less responsive to faint smells, which the researchers can test by looking at their avoidance of bad ones.

Manipulations of this kind are tricky, however, partly because the researchers have no idea how many local neurons there are in a Drosophila brain. They need to modify the majority of them if they are to see the effects they are looking for.

If they succeed they will then attempt to show that something similar is happening in the mammalian brain. But finding a noise-generating cell resembling a fly鈥檚 local neuron in a mammalian brain will be a huge challenge, according to Thomas Klausberger, also at Oxford. Klausberger is busy discovering new kinds of interneuron in the rat hippocampus, a structure that has been compared to the insect mushroom body because of its role in learning and memory. He points out that one region alone contains at least 21 different types of interneuron.

Biophysicist Frank Moss of the University of Missouri in St Louis, who did the crayfish study in 1993, is impressed by Miesenb枚ck鈥檚 findings. He has long suspected that animals take advantage of stochastic resonance to boost their reproductive success and says that Miesenb枚ck may be about to clinch it.

Moss was one of the authors of a (Nature, vol 402, p 291) which showed for the first time that externally applied noise worked via stochastic resonance. He was working on paddlefish, which find food by using electrosensors in their snouts to detect faint electrical signals given off by plankton, their natural prey. Moss put a paddlefish in a tank of water containing plankton, along with two electrodes which generated noise in the form of a randomly varying electric field. When he measured the effect of the noise, he found that there was an intermediate amplitude at which the fish鈥檚 success rate significantly increased.

Optimal performance when the noise level is intermediate is one of the characteristics of stochastic resonance: too little noise and the signal doesn鈥檛 reach the threshold, too much and the signal will be swamped by noise. The noise-benefit relationship is therefore shaped like an inverted U.

More recently, Moss has turned his attention to tiny aquatic crustaceans called Daphnia or water fleas. He believes they provide another strand of evidence pointing to internally generated stochastic resonance.

Daphnia have a characteristic foraging behaviour that follows the sequence of a hop, a pause, a turn through an angle and another hop. The turn angles vary and appear random to the naked eye.

Moss thought otherwise. He and his colleagues videoed five different species of Daphnia as they foraged for food in a shallow tank, and measured hundreds of turning angles. When they plotted the frequency distribution of these angles, they found that it was not completely random: some turning angles were more frequent than others. Their overall distribution could be described mathematically using a parameter called 鈥渘oise intensity鈥� 鈥� a measure of how random, or noisy, it is.

Next they ran computer simulations of foraging Daphnia using different noise intensities. They found that the most successful food-gathering strategy used the noise-intensity level they had measured in real Daphnia. Lower or higher noise intensities reduced foraging success according to the classic inverted-U shape of stochastic resonance (). Though they don鈥檛 yet know how Daphnia generate their distribution of turning angles, they argue it is an example of stochastic resonance in action and that it must be produced internally. 鈥淚t originates somewhere within the Daphnia, maybe its brain, but we don鈥檛 know,鈥� says Moss. He adds that the optimal noise intensity must be the product of natural selection, because Daphnia using it would find more food and so maximise their fitness.

The idea that biological systems exploit internally generated noise still has questions hanging over it, however. One big one is whether what is being generated by the local neurons in the fruit fly is genuine noise. Bart Kosko, an electrical engineer at the University of Southern California, Los Angeles, and author of a , says he is not convinced it is.

Noise has a strict mathematical definition and what looks like noise in a complex biological system usually turns out to be a signal leaking from elsewhere. 鈥淲hat needs to be done is to take that 鈥榥oise鈥� source and show that it has the statistical footprint of noise,鈥� says Kosko. If it isn鈥檛 genuine noise then, by definition, you haven鈥檛 got stochastic resonance.

Neuroscientist Gy枚rgy Buzs谩ki of Rutgers University in New Jersey goes one step further, arguing that if something is boosting faint signals to threshold in the brain, it is unlikely to be noise. 鈥淕enerating noise is very expensive,鈥� he says. 鈥淎 good system, such as we presume the brain is, can鈥檛 afford it.鈥�

Buzs谩ki agrees with Miesenb枚ck that there is probably a noise-like signal which modulates brain activity in mammals, but says there is no need to invoke specialised noise-making circuitry. Instead, he points to spontaneous neural activity occurring across the brain.

Neurons are capable of two types of activity, spontaneous and evoked. The first happens independently of an external stimulus, whereas the second is a response to it. Spontaneous activity is interesting to neuroscientists because it provides a possible mechanism for generating higher mental activity in the human brain. Spontaneous activity can spread over networks of neurons, and transient periods of synchronised neural firing at a rate of about 40 鈥渟pikes鈥� a second. So-called gamma waves have been proposed as a way that different cognitive processes can be bound together to give rise to perception, for example.

Buzs谩ki says that faint incoming signals could piggyback on these spontaneous waves of activity and thus be lifted above threshold. This would be a more cost-effective way of enhancing a weak signal, he says, since spontaneous activity consumes little energy.

There is, of course, one key similarity between these two possibilities: both involve a signal that pushes another signal over a threshold. 鈥淭he principle is the same,鈥� says Miesenb枚ck. But the details matter both from the perspective of understanding the basic workings of the brain and, potentially, in order for us to exploit stochastic resonance in future sensory aids such as retinal implants.

We will have to wait a little longer to find out whether natural selection created a brain with built-in noise, or simply one which is able to borrow some other neural signal to use as noise. Either way it seems that fly brains can鈥檛 function without a little bit of dither 鈥� and that ours are probably dithering too.

The Human Brain 鈥� With one hundred billion nerve cells, the complexity is mind-boggling. Learn more in our cutting edge special report.