Nav

Monday, July 25, 2011

Hallucinating neural networks

Hearing voices is a hallmark of schizophrenia and other psychotic disorders, occurring in 60-80% of cases. These voices are typically identified as belonging to other people and may be voicing the person’s thoughts, commenting on their actions or ideas, arguing with each other or telling the person to do something. Importantly, these auditory hallucinations are as subjectively real as any external voices. They may in many cases be critical or abusive and are often highly distressing to the sufferer.

However, many perfectly healthy people also regularly hear voices – as many as 1 in 25 according to some studies, and in most cases these experiences are perfectly benign. In fact, we all hear voices “belonging to other people” when we dream – we can converse with these voices, waiting for their responses as if they were derived from external agents. Of course, these percepts are actually generated by the activity of our own brain, but how?

There is good evidence from neuroimaging studies that the same areas that respond to external speech are active when people are having these kinds of auditory hallucinations. In fact, inhibiting such areas using transcranial magnetic stimulation may reduce the occurrence or intensity of heard voices. But why would the networks that normally process speech suddenly start generating outputs by themselves? Why would these outputs be organised in a way that fits speech patterns, as opposed to random noise? And, most importantly, why does this tend to occur in people with schizophrenia? What is it about the pathology of this disorder that makes these circuits malfunction in this specific way?

An interesting approach to try and get answers to these questions has been to model these circuits in artificial neural networks. If you can generate a network that can process speech inputs and find certain conditions under which it begins to spontaneously generate outputs, then you may have an informative model of auditory hallucinations. Using this approach, a couple of studies from several years ago from the group of Ralph Hoffman have found some interesting clues as to what may be going on, at least on an abstract level.

Their approach was to generate an artificial neural network that could process speech inputs. Artificial neural networks are basically sets of mathematical functions modelled in a computer programme. They are designed to simulate the information-processing functions carried out by individual neurons and, more importantly, the computational functions carried out by an interconnected network of such neurons. They are necessarily highly abstract, but they can recapitulate many of the computational functions of biological neural networks. Their strength lies in revealing unexpected emergent properties of such networks.

The particular network in this case consisted of three layers of neurons – an input layer, an output layer, and a “hidden” layer in between – along with connections between these elements (from input to hidden and from hidden to output, but crucially also between neurons within the hidden layer). “Phonetic” inputs were fed into the input layer – these consisted of models of speech sounds constituting grammatical sentences. The job of the output layer was to report what was heard – representing different sounds by patterns of activation of its forty-three neurons. Seems simple, but it’s not. Deciphering speech sounds is actually very difficult as individual phonetic elements can be both ambiguous and variable. Generally, we use our learned knowledge of the regularities of speech and our working memory of what we have just heard to anticipate and interpret the next phonemes we hear – forcing them into recognisable categories. Mimicking this function of our working memory is the job of the hidden layer in the artificial neural network, which is able to represent the prior inputs by the pattern of activity within this layer, providing a context in which to interpret the next inputs.

The important thing about neural networks is they can learn. Like biological networks, this learning is achieved by altering the strengths of connections between pairs of neurons. In response to a set of inputs representing grammatical sentences, the network weights change in such a way that when something similar to a particular phoneme in an appropriate context is heard again, the pattern of activation of neurons representing that phoneme is preferentially activated over other possible combinations.

The network created by these researchers was an able student and readily learned to recognise a variety of words in grammatical contexts. The next thing was to manipulate the parameters of the network in ways that are thought to model what may be happening to biological neuronal networks in schizophrenia.

There are two major hypotheses that were modelled: the first is that networks in schizophrenia are “over-pruned”. This fits with a lot of observations, including neuroimaging data showing reduced connectivity in the brains of people suffering with schizophrenia. It also fits with the age of onset of the florid expression of this disorder, which is usually in the late teens to early twenties. This corresponds to a period of brain maturation characterised by an intense burst of pruning of synapses – the connections between neurons.

In schizophrenia, the network may have fewer synapses to begin with, but not so few that it doesn’t work well. This may however make it vulnerable to this process of maturation, which may reduce its functionality below a critical threshold. Alternatively, the process of synaptic pruning may be overactive in schizophrenia, damaging a previously normal network. (The evidence favours earlier disruptions).

The second model involves differences in the level of dopamine signalling in these circuits. Dopamine is a neuromodulator – it alters how neurons respond to other signals – and is a key component of active perception. It plays a particular role in signalling whether inputs match top-down expectations derived from our learned experience of the world. There is a wealth of evidence implicating dopamine signalling abnormalities in schizophrenia, particularly in active psychosis. Whether these abnormalities are (i) the primary cause of the disease, (ii) a secondary mechanism causing specific symptoms (like psychosis), or (iii) the brain attempting to compensate for other changes is not clear.

Both over-pruning and alterations to dopamine signalling could be modelled in the artificial neural network, with intriguing results. First, a modest amount of pruning, starting with the weakest connections in the network, was found to actually improve the performance of the network in recognising speech sounds. This can be understood as an improvement in the recognition and specificity of the network for sounds which it had previously learned and probably reflects the improvements seen in human language learners, along with the concomitant loss in ability to process or distinguish unfamiliar sounds (like “l” and “r” for Japanese speakers).

However, when the network was pruned beyond a certain level, two interesting things happened. First, its performance got noticeably worse, especially when the phonetic inputs were degraded (i.e., the information was incomplete or ambiguous). This corresponds quite well with another symptom of schizophrenia, especially those who experience auditory hallucinations - sufferers show phonetic processing deficits under challenging conditions, such as a crowded room.

The second effect was even more striking – the network started to hallucinate! It began to produce outputs even in the absence of any inputs (i.e., during “silence”). When not being driven by reliable external sources of information, the network nevertheless settled into a state of activity that represented a word. The reason the output is a word and not just a meaningless pattern of neurons is that the previous learning that the network undergoes means that patterns representing words represent “attractors” – if some random neurons start to fire, the weighted connections representing real words will rapidly come to dominate the overall pattern of activity in the network, resulting in the pattern corresponding to a word.

Modeling alterations in dopamine signalling also produced both a defect in parsing degraded speech inputs and hallucinations. Too much dopamine signalling produced these effects but so did a combination of moderate over-pruning and compensatory reductions in dopamine signalling, highlighting the complex interactions possible.

The conclusion from these simulations is not necessarily that this is exactly how hallucinations emerge. After all, the artificial neural networks are pretty extreme abstractions of real biological networks, which have hundreds of different types of neurons and synaptic connections and which are many orders of magnitude more complex numerically. But these papers do provide aat least a conceptual demonstration of how a circuit designed to process speech sounds can fail in such a specific and apparently bizarre way. They show that auditory hallucinations can be viewed as the outputs of malfunctioning speech-processing circuits.

They also suggest that different types of insult to the system can lead to the same type of malfunction. This is important when considering new genetic data indicating that schizophrenia can be caused by mutations in any of a large number of genes affecting how neural circuits develop. One way that so many different genetic changes could lead to the same effect is if the effect is a natural emergent property of the neural networks involved.


Hoffman, R., & Mcglashan, T. (2001). Book Review: Neural Network Models of Schizophrenia The Neuroscientist, 7 (5), 441-454 DOI: 10.1177/107385840100700513


Hoffman, R., & McGlashan, T. (2006). Using a Speech Perception Neural Network Computer Simulation to Contrast Neuroanatomic versus Neuromodulatory Models of Auditory Hallucinations Pharmacopsychiatry, 39, 54-64 DOI: 10.1055/s-2006-931496

Friday, July 8, 2011

Environmental influences on autism - splashy headlines from dodgy data

A couple of recent papers have been making headlines in relation to autism, one claiming that it is caused less by genetics than previously believed and more by the environment and the other specifically claiming that antidepressant use by expectant mothers increases the risk of autism in the child. But are these conclusions really supported by the data? Are they strongly enough supported to warrant being splashed across newspapers worldwide, where most readers will remember only the headline as the take-away message? The legacy of the MMR vaccination hoax shows how difficult it can be to counter overblown claims and the negative consequences that can arise as a result.

So, do these papers really make a strong case for their major conclusions? The first gives results from a study of twins in California. Twin studies are a classic method to determine whether something is caused by genetic or environmental factors. The method asks, if one twin in a pair is affected by some disorder (autism in this case), with what frequency is the other twin also affected? The logic is very simple: if something is caused by environmental factors, particularly those within a family, then it should not matter whether the twins in question are identical or fraternal – their risk should be the same because their exposure is the same. On the other hand, if something is caused by genetic mutations, and if one twin has the disorder, then the rate of occurrence of the disorder in the other twin should be much higher if they are genetically identical than if they only share half their genes, as fraternal twins do.

Working backwards, if the rate of twin concordance for affected status are about the same for identical and fraternal twins, this is strong evidence for environmental factors. If the rate is much higher in monozygotic twins, this is strong evidence for genetic factors. Now to the new study. What they found was that the rate of concordance for monozygotic (identical) twins was indeed much higher than for dizyogotic (fraternal) twins – about twice as high on average.

For males: MZ: 0.58, DZ: 0.21
For females: MZ: 0.60, DZ: 0.27

Those numbers are for the diagnosis of strict autism. The rate of “autism spectrum disorder”, which encompasses a broader range of disability, showed similar results:

Males: MZ: 0.77, DZ: 0.31
Females: MZ: 0.50, DZ: 0.36.

These numbers fit pretty well with a number of other recent twin studies, all of which have concluded that they provide evidence for strong heritability of the disorder – i.e., that whether or not someone develops autism is largely (though not exclusively) down to genetics.

So, why did these authors reach a different conclusion and should their study carry any more weight than others? On the latter point, the study is significantly larger than many that have preceded it. This study looked at 192 twin pairs, each with at least one affected twin. However, some recent studies have been comparable or even larger: Lichtenstein and colleagues looked at 117 twin pairs and Rosenberg and colleagues looked at 277 twin pairs. These studies found eveidence for very high heritability and negligible shared environmental effects.

Another potentially important difference is in how the sample was ascertained. Hallmayer and colleagues claim that their assessment of affected status was more rigorous than for other studies and this may be true. However, it has previously been found that less rigorous assessments correlate extremely well with the more standardised assessments, so this is unlikely to be a major factor. In addition, there is very strong evidence that disorders like autism, ADHD, epilepsy, intellectual disability, tic disorders and others all share common etiology – having a broader diagnosis is therefore probably more appropriate.

In any case, the numbers they came up with for concordance rates were pretty similar across these studies. So, why did they end up with a different conclusion? That’s not a rhetorical question – I actually don’t know the answer and if anyone else does I would love to hear it. Given the data, I don’t know how they conclude that they provide evidence for shared environmental effects.

The methodology involves some statistical modeling that tries to tease out the sources of variance. However, this modeling is based completely on a multifactorial threshold model for the disorder - the idea that autism arises when the collective burden of individually minor genetic or environmental insults passes some putative threshold. Sounds plausible, but there is in fact no evidence - at all - that this model applies to autism. In fact, it seems most likely that autism really is an umbrella term for a collection of distinct genetic disorders caused by mutations in separate genes, but which happen to cause common phenotypes (or symptoms).

If that is the case, then what the twin concordance rates actually measure is the penetrance of such mutations – if one inherits mutation X, how often does that actually lead to autism? For monozygotic twins, let us assume that the affected proband (the first twin diagnosed) has such a mutation. Because they are genetically identical, the other one must too. The chance that the other twin will develop autism thus depends on the penetrance of the mutation – some mutations are more highly penetrant than others, giving a much higher probability of developing a specific phenotype. If we average across all MZ twin pairs we therefore get an average penetrance across all such putative mutations. Now, if such mutations are dominant, as many of the known ones are, then the chance that a dizygotic twin will inherit it is 50%, while the penetrance should remain the same. So, this model would predict that the rate of co-occurrence in DZ twins should be about half that of MZ twins, exactly as observed. (No stats required).

The conclusions from this study that the heritability is only modest and that a larger fraction of variance (55%!) is caused by shared environment thus seem extremely shaky. This is reinforced by the fact that the confidence intervals for these estimates are extremely wide (for the effect of shared environment the 95% confidence interval ranges from 9% to 81%). Certainly not enough to overturn all the other data from other studies.

What about epidemiological studies that have shown statistical evidence of increased risk of autism associated with a variety of other factors, including maternal diabetes, antidepressant use, season and place of brith? All of these factors have been linked with modest increases in the risk of autism. Don’t these prove there are important environmental factors? Well, first, they don’t prove causation, they provide a statistical evidence for an association between the two factors, which is not at all the same thing. Second, the increase in risk is usually on the order of about two-fold. Twice the risk may sound like a lot, but it's only a 1% increase (from 1 to 2%), compared with some known mutations, which increase risk by 50-fold or more.

The main problem with these kinds of studies (and especially with how they are portrayed in the media) is that they are correlational and so you cannot establish a causal link directly from them. In some cases, two different correlated parameters (like red hair and freckles, for example) may actually be caused by an unmeasured third parameter. For example, in the recently published study, the use of antidepressants of the SSRI (selective serotonin reuptake inhibitor) class in mothers was associated with modestly increased risk of autism in the progeny. This association could be because SSRIs disrupt neural development in the fetus (perfectly plausible) but could alternatively be due to the known genetic link between risk of depression and risk of autism. Rates of depression are known to be higher in relatives of autistic people, so SSRI use could just be a proxy for that condition. The authors claim to have corrected for that by comparing rates of autism in the progeny of depressed mothers who were not prescribed SSRIs versus those who were but one might imagine that the severity of depression would be higher among those prescribed an antidpressant. In addition, the authors are careful to note that their findings were based on a small number of children exposed and that "Further studies are needed to replicate and extend these findings". As with many such findings, this association may or may not hold up with additional study.

As for season and place of birth, those findings are better replicated and, interestingly, also found for schizophrenia. There is a theory that these effects may relate to maternal vitamin D levels, which can also affect neural development. This also seems plausible enough. However, the problem in really having confidence in these findings and in knowing how to interpret them is that they are population averages with small effect sizes. Overall, it seems quite possible that the environment - especially the prenatal environment - can play a part in the etiology of autism. At the moment, splashy headlines notwithstanding, genetic factors look much more important and genetic studies much more likely to give us the crucial entry points to the underlying biology.



Hallmayer J, Cleveland S, Torres A, Phillips J, Cohen B, Torigoe T, Miller J, Fedele A, Collins J, Smith K, Lotspeich L, Croen LA, Ozonoff S, Lajonchere C, Grether JK, & Risch N (2011). Genetic Heritability and Shared Environmental Factors Among Twin Pairs With Autism. Archives of general psychiatry PMID: 21727249

Lichtenstein P, Carlström E, Råstam M, Gillberg C, & Anckarsäter H (2010). The genetics of autism spectrum disorders and related neuropsychiatric disorders in childhood. The American journal of psychiatry, 167 (11), 1357-63 PMID: 20686188

Rosenberg, R., Law, J., Yenokyan, G., McGready, J., Kaufmann, W., & Law, P. (2009). Characteristics and Concordance of Autism Spectrum Disorders Among 277 Twin Pairs Archives of Pediatrics and Adolescent Medicine, 163 (10), 907-914 DOI: 10.1001/archpediatrics.2009.98

Croen LA, Grether JK, Yoshida CK, Odouli R, & Hendrick V (2011). Antidepressant Use During Pregnancy and Childhood Autism Spectrum Disorders. Archives of general psychiatry PMID: 21727247

Monday, July 4, 2011

On discovering you’re an android

Deckard: She's a replicant, isn't she?
Tyrell: I'm impressed. How many questions does it usually take to spot them?
Deckard: I don't get it, Tyrell.
Tyrell: How many questions?
Deckard: Twenty, thirty, cross-referenced.
Tyrell: It took more than a hundred for Rachael, didn't it?
Deckard: [realizing Rachael believes she's human] She doesn't know.
Tyrell: She's beginning to suspect, I think.
Deckard: Suspect? How can it not know what it is?

A very discomfiting realisation, discovering you are an android. That all those thoughts and ideas and feelings you seem to be having are just electrical impulses zapping through your circuits. That you are merely a collection of physical parts, whirring away. What if some of them break and you begin to malfunction? What if they wear down with use and someday simply fail? The replicants in BladeRunner rail against their planned obsolescence, believing in the existence of their own selves, even with the knowledge that those selves are merely the products of machinery.

The idea that the self, or the conscious mind, emerges from the workings of the physical structures of the brain – with no need to invoke any supernatural spirit, essence or soul – is so fundamental to modern neuroscience that it almost goes unmentioned. It is the tacitly assumed starting point for discussions between neuroscientists, justified by the fact that all the data in neuroscience are consistent with it being true. Yet it is not an idea that the vast majority of the population is at all comfortable with or remotely convinced by. Its implications are profound and deeply unsettling, prompting us to question every aspect of our most deeply held beliefs and intuitions.

This idea has crept along with little fanfare – it did not emerge all at once like the theory of evolution by natural selection. There was no sudden revolution, no body of evidence proffered in a single moment that overturned the prevailing dogma. While the Creator was toppled with a single, momentous push, the Soul has been slowly chipped away at over a hundred years or more, with most people blissfully unaware of the ongoing assault. But its demolition has been no less complete.

If you are among those who is skeptical of this claim or who feels, as many do, that there must be something more than just the workings of the brain to explain the complexities of the human mind and the qualities of subjective experience (especially your own), then first ask yourself: what kind of evidence would it take to convince you that the function of the brain is sufficient to explain the emergence of the mind?

Imagine you came across a robot that performed all the functions a human can perform – that reported a subjective experience apparently as rich as yours. If you were able to observe that the activity of certain circuits was associated with the robot’s report of subjective experience, if you could drive that experience by activating particular circuits, if you could alter it by modifying the structure or function of different circuits, would there be any doubt that the experience arose from the activity of the circuits? Would there be anything left to explain?

The counter-argument to this thought experiment is that it would never be possible to create a robot that has human-like subjective experience (because robots don’t have souls). Well, all those kinds of experiments have, of course, been done on human beings, tens of thousands of times. Functional magnetic resonance imaging methods let us correlate the activity of particular brain circuits with particular behaviours, perceptions or reports of inward states. Direct activation of different brain areas with electrodes is sufficient to drive diverse subjective states. Lesion studies and pharmacological manipulations have allowed us to map which brain areas and circuits, neurotransmitters and neuromodulators are required for which functions, dissociating different aspects of the mind. Finally, differences in the structure or function of brain circuits account for differences in the spectrum of traits that make each of us who we are as individuals: personality, intelligence, cognitive style, perception, sexual orientation, handedness, empathy, sanity – effectively everything people view as defining characteristics of a person. (Even firm believers in a soul would be reluctant recipients of a brain transplant, knowing full well that their “self” would not survive the procedure).

The findings from all these kinds of approaches lead to the same broad conclusion: the mind arises from the activity of the brain – and nothing else. What neuroscience has done is correlated the activity of certain circuits with certain mental states, shown that this activity is required for these states to arise, shown that differences in these circuits affect the quality of these states and finally demonstrated that driving these circuits from the outside is sufficient to induce these states. That seems like a fairly complete scientific explanation of the phenomenon of mental states. If we had those data for our thought-experiment robot, we would be pretty satisfied that we understood how it worked (and could make useful predictions about how it would behave and what mental states it would report, given enough information of the activity of its circuits).

However, many philosophers (and probably a majority of people) would argue that there is something left to explain. After all, I don’t feel like an android – one made of biological rather than electronic materials, but a machine made solely of physical parts nonetheless. I feel like a person, with a rich mental life. How can the qualities of my subjective experience be produced by the activity of various brain circuits?

Many would claim, in fact, that subjective experience is essentially “ineffable” – it cannot be described in physical terms and cannot thus be said to be physical. It must therefore be non-physical, immaterial or even supernatural. However, the fact that we cannot conceive of how a mental state could arise from a brain state is a statement about our current knowledge and our powers of imagination and comprehension, not about the nature of the brain-mind relationship. As an argument, what we currently can or cannot conceive of has no bearing on the question. The strong intuition that the mind is more than just the activity of the brain is reinforced by an unfortunate linguistic accident – that the word “mind” is grammatically a noun, when really it should be a verb. At least, it does not describe an object or a substance, but a process or a state. It is not made of stuff but of the dynamic relations between bits of stuff.

When people argue that activity of some brain circuit is not identical to a subjective experience or sufficient to explain it, they are missing a crucial point – it is that activity in the context of the activity of the entire rest of the nervous system that generates the quality of the subjective experience at any moment. And those who dismiss this whole approach as scientific reductionism ad absurdum, claiming that the richness of human experience could not be explained merely by the activity of the brain should consider that there is nothing “mere” about it – with hundreds of billions of neurons making trillions of connections, the complexity of the human brain is almost incomprehensible to the human mind. (“If the brain were so simple that we could understand it, then we would be so simple that we couldn’t”).

To be more properly scientific, we should ask: “what evidence would refute the hypothesis that the mind arises solely from the activity of the brain”? Perhaps there is positive evidence available that is inconsistent with this view (as opposed to arguments based merely on our current inability to explain everything about the mind-brain relationship). It is not that easy to imagine what form such positive evidence would take, however – it would require showing that some form of subjective experience either does not require the brain or requires more than just the brain.

With respect to whether subjective experience requires the brain, the idea that the mind is associated with an immaterial essence, spirit or soul has an extension, namely that this soul may somehow outlive the body and be said to be immortal. If there were strong evidence of some form of life after death then this would certainly argue strongly against the sufficiency of neuroscientific materialism. Rather depressingly, no such evidence exists. It would be lovely to think we could live on after our body dies and be reunited with loved ones who have died before us. Unfortunately, wishful thinking does not constitute evidence.

Of course, there is no scientific evidence that there is not life after death, but should we expect neuroscience to have to refute this alternative hypothesis? Actually, the idea that there is something non-physical at our essence is non-refutable – no matter how much evidence we get from neuroscience, it does not prove this hypothesis is wrong. What neuroscience does say is that it is not necessary and has no explanatory power – there is no need of that hypothesis.
s;o