Conference paper and RSMG4 report

Both of these are now up on my research page. The paper is to be presented at IJCNN 2011 in San Jose, California, in August of this year.

First paper, next steps

Last week I submitted my first ever paper (yay!) to IJCNN 2011 in San Jose, having found some interesting results using my implementation of the Ruppin and Reggia (1995) model, most notably that:

• Synaptic compensation using remote memories (which of course are easier to obtain in AD) actually accelerates cognitive decline due to the decreased variability in the data set used to calculate compensatory terms.
• Where small-world connectivity occurs in the brain, this has the effect of increasing redundancy and resilience to damage, at the expense of lower overall capacity.
• Selectively partially muting, rather than deleting, synapses in a spreading area of damage can be used as a simulation of tau pathology, in which vesicles become blocked and axons degrade. This type of lesioning offers a much more graceful decline in performance as the compensatory mechanisms keep up with the changes, but catastrophic damage occurs after a certain level of lesioning and the decline in performance is much more dramatic than with plain synaptic deletion.

Once I get feedback from IJCNN and make any required changes, I'll put a copy of the paper up on my website. In the meantime, I need to think about next steps.

Firstly, my RSMG4 progress report is due in April. This will be a simple 2000-4000 word write-up of my progress over the last 6 months, including what I learned in Göttingen and Zürich on the neuroscience and reservoir computing courses, and of course the IJCNN paper.

Beyond that, I guess I will have the following tasks to choose from:

• [OPTIONAL] Continue work on the Ruppin and Reggia model -- design experiments to explore the above effects further.
• [ESSENTIAL] Collate definitive medical data against which my models should be compared, and lay out the way in which my model can be shown to be a small part of the overall larger brain organisation (i.e. hippocampal vs neocortical organisation). This is hard and will require much thought!
• Incorporate amyloid (including N-APP and anaesthesia) pathology simulations to test cutting-edge medical theories in computational networks.
• Begin working on implementation of a reservoir computing network which incorporates synaptic compensation (very important, as the reservoir network's dynamics change dramatically with only slight changes in the internal reservoir).
• Can reservoir networks be shown to be better models than, or at least as accurate as, Hopfield-style associative networks (a la Ruppin and Reggia)? What are the differences in behaviour?
• What other symptoms of AD can be represented in a reservoir network, other than just failure to accurately recall a stored pattern? As the computational power is much greater, could a basic model of degradation of language / motor skills or some other feature be implemented?

Lots to do!

A plan?

Another meeting with John today, after getting some feedback from John Jefferys in Neuroscience. The current plan for a framework of modelling and experimentation is as follows:

1. Take one of the older models of Alzheimer's disease from the mid-1990s (e.g. Willshaw, Tsodyks and Feigel'man, Hasselmo).
2. Update it with modern learning algorithms (e.g. LEABRA, Contrastive Hebbian Learning).
3. Use evolutionary computation to set parameters instead of calculating / guessing them (e.g. firing thresholds, Gaussian connectivity radius and density). Fitness function should be adequate "normal" functioning of the network with least amount of energy (= connections?) required.
4. Use this updated and evolved model to repeat the 1990s experiments of Ruppin and Reggia or Hasselmo (that's already potentially one small contribution to knowledge in the bag). (As a footnote, Ruppin and Reggia's synaptic deletion and compensation model uses biologically implausible uniform compensation values across the whole network instead of adjusting the remaining synapses using the learning algorithm. This is another change worth considering).
5. Use the model to attempt to dis/prove a more recent medical hypothesis, such as Nikolaev et al. (2009), which will require showing that up to a certain level of deletion the performance actually increases, and then after that level it degrades. (John argues that I should also dis/prove other hypotheses with the same model to be able to show that it's a valid contribution).

This should be enough for a bare minimum PhD thesis, but also opens several avenues for further tinkering or completely different approaches:

• Starting from different network architectures in (1). e.g. more recent models of schizophrenia or general neocortex models.
• Implementing other learning algorithms in (2).
• Broadening the number of parameters which could be evolved in (3), e.g. the distribution of DR6 "death receptors" in (5).
• Trying out various different things along these lines until I get bored or run out of time!

Next steps

After speaking with John Bullinaria today we've worked out the next steps I should be taking on the way to creating the Thesis Proposal:

1.  Continue implementing some of the older models (e.g. Ruppin and Reggia (1995)), and repeating the experiments to compare results.
1. Try other lesions, such as adding noise to the connections.
2. What symptoms appear?
3. What do these symptoms represent? (This could lead to a basic paper).
2. Identify weaknesses in the models, and gaps in what they can tell us.
1. Be analytical: are the assumptions of ~15 years ago still correct? Is a Hopfield network still the best type of model?
2. Identify possible improvements (preferably the ones involving the minimum work and the maximum impact).
3. Identify a selection of plausible hypotheses from the medical literature, and investigate how these could be modelled in existing or new model classes.
3. Investigate more recent classes of neural model which can be adapted for studying Alzheimer's disease; for example:
1. Contrastive Hebbian learning.
2. LEABRA learning.
3. Work by John Jefferys, Rolls and Loh, etc.
4. Recent and more comprehensive models of normal working memory, particularly ones involving the two stages of hippocampal and long-term memory storage (e.g. by McLelland).

This last point (3) is perhaps the most important, as the premise that I need to work from recent and comprehensive models of memory is vital. There is not much point in trying to incrementally improve basic 15-year old models of AD when neural modelling has moved on a great deal in that time. A better approach would be to take these more recent models and lesion them to study AD, using the lessons learned from the older research.

John also recommended looking into PDP++ (now called Emergent), a neural network simulator which could save some implementation time, and that for the thesis proposal I shouldn't worry too much about drawing up a specific single hypothesis, but rather focus on the experimental methodology I'll be employing (i.e. applying past work on AD modelling to recent working memory models).

RSMG2 published

Well I've successfully got my first proper report out of the door, which is the 5000 word progress report on what I've been doing so far throughout the year. It was accepted in my Thesis Group meeting with only minor changes to the prospective timetable being required, and it seems that I'm heading along the right track in terms of writing style in my literature reviews etc, which I'm quite pleased about. I've also managed to convince Prof. John Jefferys of the School of Neuroscience to join my Thesis Group unofficially so I can receive feedback from an experienced neuroscience researcher regarding the medical side of my work.

Other discussion points from the meeting were that I'm starting to move away from the original ideas regarding creating a model of routine memory maintenance and trying to evolve schizophrenia and Alzheimer's disease out of this, for two main reasons:

1. Any model of 'normal working' in the brain would have to be incredibly fine-grained for it to be able to degrade in all possible ways that I'd want to test in order to try to develop the disorders. That's a lot of effort for not very much gain.
2. Attempting to shoe-horn evolution into the system in this way "for the sake of it" wouldn't necessarily achieve much. It would be better to use the evolutionary algorithm to find ideal starting parameters (such as amount of synaptic plasticity) for the various models, as current modelling techniques tend to just pluck values out of the air with little justification. So this is one area of existing modelling which I can improve.

I was also encouraged to decide on one specific disorder to investigate (most likely AD in my case), and to use the relevant research into other disorders (schizophrenia, epilepsy, etc) to draw ideas from.

Finally, it would be beneficial to create a framework to be able to experiment on different hypotheses and models whilst keeping environmental conditions constant. This is essential for me to be able to argue that my model actually does (dis)prove a given hypothesis, rather than it simply being the case that my model is badly implemented and simply doesn't actually work.

Antipsychotics and antidementia agents

In a neuroscience lecture this morning given by James Reed of Birmingham and Solihul Mental Health Foundation Trust, we followed the progress of a number of "typical" mental health, including schizophrenia and Alzheimer's disease cases. One interesting point I noted was that in both cases there was administration of drugs (anti-pyschotics in the schizophrenia case and drugs such as donepezil ("Aricept") in the Alzheimer's case). I feel there is something to be learned for my research from how antipyschotics and antidementia agents act on the brain to reduce the symptoms. It's clear in both cases that the drugs do not address the causes of the disorders, but only manage the symptoms of the suffers. A bit of Wikipedia trawling follows... (proper paper references to come when I properly research this!)

Aricept is a type of acetylcholinesterase inhibitor (i.e. it inhibits the enzyme which breaks down acetylcholine), which ties in with the theories I've read in Stein and Ludik (1998) about the role of acetylcholine in memory management and prevention of runaway synaptic modification. The Alzheimer's Society helpfully explain that "Aricept, Exelon and Reminyl prevent an enzyme known as acetylcholinesterase from breaking down acetylcholine in the brain. Increased concentrations of acetylcholine lead to increased communication between the nerve cells that use acetylcholine as a chemical messenger, which may in turn temporarily improve or stabilise the symptoms of Alzheimer's disease ... the action of Ebixa is quite different to, and more complex than, that of Aricept, Exelon and Reminyl. Ebixa blocks a messenger chemical known as glutamate. Glutamate is released in excessive amounts when brain cells are damaged by Alzheimer's disease, and this causes the brain cells to be damaged further. Ebixa can protect brain cells by blocking this release of excess glutamate."

Most antipyschotic drugs appear to work by simply blocking dopamine receptors in the brain, typically the D2 or D4 receptor, and "atypical" antipsychotics also act on serotonin receptors. James mentioned that clozapine is a particularly effective antipsychotic (although it is rarely used now as it has some severe side-effects), and no-one quite knows why, although apparently it acts on D2 less strongly than other antipsychotics. Developing a new drug which works in the same way, but without the side-effects, is a major part of clinical schizophrenia research which would be aided by knowing how clozapine actually works, and how it is different to other antipsychotics! This is potentially something I could attempt to model during my work.

There is some interesting background to various medical and neurological theories about the causes of schizophrenia on Wikipedia too, with particular mention of the Glutamate (related to NMDA receptors; see Greenstein and Ruppin, 1998) and Dopamine (supported by the fact that antipsychotics work by blocking dopamine receptors) theories. Apparently the glutamate hypothesis "does not negate the dopamine hypothesis, and the two may be ultimately brought together by circuit-based models." (Lisman et al., 2008, as cited at Wikipedia).

Interestingly, glutamate is mentioned in both schizophrenia and Alzheimer's disease, although for different reasons. In the glutamate theory of schizophrenia (non-Wikipedia link), treatment is performed by increasing the amount of glutamate available, whereas Ebixa attempts to reduce the amount of damage caused by glutamate in Alzheimer's disease.

Finally, the BBC has produced a Dementia 2010 news feature focussing on dementia and Alzheimer's disease. Of particular interest, if only for soundbites as part of my introduction, are:

• Dementia 'losing out' to cancer in funding stakes (The usual sob story, albeit justified, calling for more funding).
• What is delaying a cure for Alzheimer's? (The answer again is "funding", but the article at least acknowledges that much of the currently-reported Alzheimer's disease research progress in the media simply focusses on seemingly trivial findings, such as that eating cornflakes three times a day and tying your shoelaces immediately before attempting a crossword statistically lowers your chances of developing AD, without going deeper into why).
• Indian village may hold key to beating dementia (Nearly one of the aforementioned "trivial observations" reports, but at least they did some research into more specific factors such as a low-cholesterol, vegetarian diet, and lack of the APO4E gene which apparently predisposes people to AD. This also links to a paper by Sima and Li (2006) on a proposed connection between diabetes and Alzheimer's disease).

References
Greenstein-Messica, A. and Ruppin, E. (1998), "Synaptic runaway in associative networks and the pathogenesis of schizophrenia", in Neural Computation 10:451--465.

Lisman JE, Coyle JT, Green RW, et al. (May 2008). "Circuit-based framework for understanding neurotransmitter and risk gene interactions in schizophrenia". Trends in Neurosciences 31 (5): 234–42.

Sima, A.A.F and Li, Z. (2006), "Diabetes and Alzheimer's Disease - Is There a Connection?". The Review of Diabetic Studies 3(4):161.


Stein, D.J. and Ludik, J. (1998), Neural Networks and Pyschopathology, Cambridge University Press.

I've created a brain!

And now it's going to take over the world, mwahahahahaa....

Well not quite, but in the last two weeks I've implemented (in Java) the model by Ruppin and Reggia (1995) which creates a non-fully-connected Hopfield network, gets it to learn a number of patterns or 'memories', in an activity-dependent Hebbian manner ("cells that fire together, wire together"), and then tries to recall them.

Up to a certain number of patterns (n / 2 log n, where n is the number of neurons in the network) it works beautifully and then, as expected, once its memory is 'full' recall performance drops dramatically. I'm yet to formally carry out any experiments along the same lines as Ruppin and Reggia, but initial results seem to indicate my implementation gives roughly the same behaviour as their model.

Ruppin and Reggia were only able to create networks up to 1600 neurons in size, due to the limited computing power available at the time. With my implementation I've already successfully run networks of 100,000 neurons in size (=10,000 memories!) in the default Java heap space of 64MB. By increasing the heap space -- and being very patient as it will take a long time to run -- I hope to be able to demonstrate networks of maybe up to half a million neurons in size. Whether that will enhance the experimental results at all remains to be seen.

Next steps are to implement network lesioning -- that is, damaging the network by deleting neurons and/or synapses, in an attempt to get it to develop symptoms similar to Alzheimer's disease, and recreate the experiments of Ruppin and Reggia (1995) on a larger scale.

After that, the model will be extended further to try to imitate symptoms of schizophrenia, as in Ruppin et al. (1996). There is also an alternative model of Alzheimer's disease introduced by Horn et al. (1996) which I'd like to implement, and no doubt I'll find some others in the literature during that time as well (including hopefully some more recent than the mid-1990s). During that time I'll also have to write up my current reading for the RSMG2 report, due in May of this year.

References
Horn, D., Levy, N., Ruppin, E. (1996), "Neuronal-based synaptic compensation: a computational study in Alzheimer's disease", in Neural Computation 8:1227--1243.

Horn, D., Levy, N., Ruppin, E. (1998), "Memory maintenance via neuronal regulation", in Neural Computation 10:1--18.

Ruppin, E. and Reggia J. (1995), "A neural model of memory impairment in diffuse cerebral atrophy", in British Journal of Psychiatry 166:19--28

Ruppin, E., Reggia, J., Horn, D. (1996), "Pathogenesis of schizophrenic delusions and hallucinations: a neural model", Schizophrenia Bulletin (22)1:105-123

Memories that can last a lifetime

How does your brain store childhood memories for your entire life, even after the actual neurons present in your brain when you first experienced that memory have long since died and been replaced, many times over?

It's a fascinating question, and one which Horn et. al. (1998) attempt to answer using a computational model. Firstly, memories are not stored one-per-neuron: "Stored memories are not represented at specific neurons of the network, but their corresponding representations are distributed; many neurons participate in a given [memory], and a particular neuron participates in several different [memories] (Ruppin et. al., 1996).  Horn et. al. (1998) show that by adjusting the strengths of the synapses between neurons whenever a neuron dies in order to keep the activation of all of the remaining neurons the same, the overall network is capable of retaining the stored memories for quite a while.

Interestingly, this mechanism also counter-balances the 'pathological attractors' which tend to arise from the deletion or modification of synapses in this way during activity-dependent learning, and which represent symptoms of schizophrenia (Ruppin et. al., 1996).

This got me thinking: what if I could create a copy of this model to show memories being retained in this way during the course of normal neuronal death, but implement some existing medical hypotheses e.g. delayed NMDA receptor maturation (Greenstein and Ruppin, 1998) or excessive neuronal activity leading to excess glutamate production and neurodegeneration (Good Brain, Bad Brain module semester 2) to try to model a change in the normal neuronal death process which leads to a disorder such as Alzheimer's or schizophrenia (or both, if it can be shown that they are sufficiently closely-linked)? Horn et. al. (1998) also suggest that "recent findings [listed] indicate that signaling molecules involved in neuronal regulation are altered in Alzheimer's disease", adding further weight to this idea.

This would be taking a different angle to the research compared to the suggestion of simply trying to explain Rund (2009)'s findings that there may be a degenerative process in schizophrenia in the context of Alzheimer's disease (which is possibly looking like a very tenuous link to make).

So it looks like I could currently take three paths:

1. Model possibly causes of Alzheimer's and/or schizophrenia, in the context of memory maintenance (Horn et. al., 1998).
2. Model the shared elements of the progress of Alzheimer's and schizophrenia (i.e. synaptic runaway (Ruppin et. al., 1996; Greenstein and Ruppin, 1998; Hasselmo, 1994; Ruppin and Reggia, 1995; Horn et. al., 1996)).
3. Model a degenerative process in schizophrenia (possibly drawing from Alzheimer's disease modelling), to corroborate Rund (2009)'s findings.
   

References
Greenstein-Messica, A. and Ruppin, E. (1998), "Synaptic runaway in associative networks and the pathogenesis of schizophrenia", in Neural Computation 10:451--465.

Hasselmo, M.E.  (1994), "Runaway synaptic modification in models of cortex: Implications for Alzheimer's disease", in Neural Networks 7:13--40.

Horn, D., Levy, N., Ruppin, E. (1996), "Neuronal-based synaptic compensation: a computational study in Alzheimer's disease", in Neural Computation 8:1227--1243.

Horn, D., Levy, N., Ruppin, E. (1998), "Memory maintenance via neuronal regulation", in Neural Computation 10:1--18.

Rund, B.R. (2009), "Is there a degenerativeprocess going on in the brain of people with schizophrenia?" in Frontiers in Neuroscience (3)36:1--6.

Ruppin, E. and Reggia J. (1995), "A neural model of memory impairment in diffuse cerebral atrophy", in British Journal of Psychiatry 166:19--28

Ruppin, E., Reggia, J., Horn, D. (1996), "Pathogenesis of schizophrenic delusions and hallucinations: a neural model", Schizophrenia Bulletin (22)1:105-123

Some models to start off with

In a previous post I started to look at the possibility of modelling a shared overlap with Alzheimer's disease and schizophrenia.

The first thing to mention is that, contrary to what I wrote previously, synaptic runaway appears to have very little to do with Alzheimer's disease, so we've lost that potential link with schizophrenia (which does have a lot to do with synaptic runaway).

EDIT: Actually, maybe synaptic runaway does have something to do with Alzheimer's: Greenstein and Ruppin (1998) mentions "Hasselmo's [1994] hypothesis concerning ... synaptic runaway in the progression of Alzheimer's disease".

There's a whole raft of papers by Ruppin, Reggia, Horn and Levy amongst others from the mid-90s on computational modelling of both Alzheimer's disease and schizophrenia. One such paper includes a very clear description of a model of Alzheimer's disease, in which they deleted synapses at random but also caused the remaining synapses to attempt to compensate for these changes, which caused older memories to be retained long after recent (short-term) memories were destroyed by the deletion process (Ruppin and Reggia, 1995), just like in Alzheimer's disease where patients forget things like turning off the cooker long before they forget older memories such as the names of their loved ones.

They then took this model and re-spun it to model schizophrenia via synaptic runaway, such that the model retained all its memories for a long time, but interestingly also spontaneously retrieved parts of memories at random without any stimulus (Ruppin et. al., 1996), mirroring the hallucinations and delusions in schizophrenia and perhaps explaining the sensation that "someone else" (aliens, the CIA, etc.) is controlling the sufferer's thoughts.

This appears to be the last paper in this particular strand of their research, and it contains some interesting proposals for future work, which could feasibly form part of my preliminary research:

1. Putting several copies of the schizophrenia model together to simulate cortical modules communicating with each other, to see if this restricts hallicinogenic symptoms (the 'biased pathological attractors' in the jargon) to individual parts of the brain, or spreads them throughout.
2. Enabling the models to be reset each time they suffer a 'hallucination' so they can be repeatedly re-tested instead of forcing them to remain in the 'pathological attractor' state.
3. Simply running the models with far larger neural networks, now that improved computational power is available.
   

Interestingly, another paper in this series by Greenstein and Ruppin (1998) potentially refutes Rund (2009)'s hypothesis that schizophrenia could be a neurodegenerative disorder like Alzheimer's disease. This indicates that I could be barking up the wrong tree if I try to "model degenerative processes in both disorders" as I wrote previously.

Greenstein and Ruppin (1998) hypothesised that schizophrenia is due to delayed NMDA receptor maturation which leads to the neural network being "overloaded at an early stage, much before it reaches its explicit capacity limits" and therefore to synaptic runaway and schizophrenia, with no mention of unusual cell death -- so perhaps we'll have to find some other process to explain Rund (2009)'s observations of neurodegeneration in some schizophrenia patients.


References
Greenstein-Messica, A. and Ruppin, E. (1998), "Synaptic runaway in associative networks and the pathogenesis of schizophrenia", in Neural Computation 10:451--465.

Hasselmo, M.E.  (1994), "Runaway synaptic modification in models of cortex: Implications for Alzheimer's disease", in Neural Networks 7:13--40.

Rund, B.R. (2009), "Is there a degenerativeprocess going on in the brain of people with schizophrenia?" in Frontiers in Neuroscience (3)36:1--6.

Ruppin, E. and Reggia J. (1995), "A neural model of memory impairment in diffuse cerebral atrophy", in British Journal of Psychiatry 166:19--28

Ruppin, E., Reggia, J., Horn, D. (1996), "Pathogenesis of schizophrenic delusions and hallucinations: a neural model", Schizophrenia Bulletin (22)1:105-123

Skills and interests

Right, time to focus a little more specifically on what I actually want to do (i.e. which techniques to use) during this PhD.

I want to utilise:

• Evolutionary computation
• Neural networks
• Hence --> evolving neural networks
• Methodical experimentation rather than "one great big long think"

This lends itself to the idea of heuristically creating or refining a model to fit a certain fitness function (e.g. expected behaviour of the network under certain conditions).


The fields in which I am particularly interested in using these techniques are:

• Understanding more about Alzheimer's disease
• Model of robustness of musical memory in AD?
• Nothing obviously defined in the literature, but it could be over-ambitious to hope to model musical memory sufficiently in the time available for the PhD.
  
• Models of progression?
• Already many models defined -- probably end up refining or combining existing models.
• OR understanding more about schizophrenia
• Model of ability of music to calm patients in both AD and schizophrenia?
• Nothing obviously defined in the literature, but it could be over-ambitious to hope to model musical processing sufficiently in the time available for the PhD.
• Models of hallucinations?
• Already many models defined -- probably end up refining or combining existing models.
• OR investigating the overlap between aspects of AD and schizophrenia
• Model of potential degenerative processes in both disorders?
• Degenerative process in schizophrenia postulated in Rund (2009).
• Possible supporting evidence for reduced cortical synaptic connection in schizophrenia (Hoffman and McGlashan (2001)) and reduced hippocampal synaptic connection in Alzheimer's disease (Duch (2007) and Wallenstein and Hasselmo (1997)) both leading to runaway synaptic modification (Hasselmo (1994)).
• Would need to show how the runaway synaptic modification follows the same basic pathology but diverges to form the two separate disorders.
• Needs further reading to investigate the initial feasibility of this hypothesis, or whether it's easy to disprove.
  

Well, having written all that down, it looks like the above question (modelling degenerative processes in AD and schizophrenia) actually could present quite a good topic. (On an unrelated note, why is it that I can spend a day moping around, staring blankly at some papers and not really getting much done, but as soon as I sit down and start to write here, ideas come flowing forth?)

For the time being, at least, if I read further into the possibility of a neuropathalogical overlap in AD and schizophrenia, I can get an idea for how feasible as a hypothesis this is (at the moment it's not much more than a hunch, having read through a few related papers).

I'll also need to figure out how I'll actually go about testing any hypothesis in this area experimentally, as I still want to make use of the refine-and-repeat evolutionary computation technique for generating candidate networks according to the rules of whatever model I end up using / designing.

References
Rund, B.R. (2009), "Is there a degenerativeprocess going on in the brain of people with schizophrenia?" in Frontiers in Neuroscience.

Wallenstein, G.V. and Hasselmo M.E. (1997), "Are there common neural mechanisms for learning, epilepsy, and Alzheimer's disease?", in Stein and Ludik (1997),  pp. 314--316.

Stein, D.J. and Ludik, J. (1998), Neural Networks and Pyschopathology, Cambridge University Press.

Duch, W. (2007), "Computational Models of Dementia and Neurological Problems", in Methods in Molecular Biology, pp. 305--336.

Hoffman, R.E. and McGlashan, T.H. (2001), "Neural network models of schizophrenia", in Neuroscientist 7(5):441--454.

Hasselmo, M.E.  (1994), "Runaway synaptic modification in models of cortex: Implications for Alzheimer's disease", in Neural Networks 7:13--40.

So many questions, so little time!

Argh!

Now that I've got that out of the way, what is it that's so frustrating me?

Whilst I've been (slowly) reading my way around the topic of modelling neurological disorders, I've come across some interesting articles. I've looked into models of schizophrenia and Alzheimer's disease, and analysed at least 6 models in the process. I've toyed with the idea of picking a model and refining it, or picking several models and merging them. Some models postulate key differences between themselves and other models, so I've considered trying to confirm or contradict these differences. I've thought about the nature of hallucinations in AD and schizophrenia, and the possibility of investigating differences in how damaged brains dream.

As I read more, the number of questions one could ask grows ever longer. Whilst this is fantastic from a scientific point of view (no shortage of things to research, then), it really doesn't help when trying to narrow down a topic with the aim of writing a specific proposal 8 months from now!

At the moment it feels like my reading is aimlessly meandering from sub-topic to sub-topic, with no clear goal in sight. I need to figure out a way to pick a question and stick to it, for the time being at least, so I can direct my reading around it.

Any suggestions, anyone?

Some questions to think about

After chatting with John (my supervisor) briefly this evening, we've come up with a couple of potentially interesting questions to try to investigate, or at least to use to try to direct my research more specifically:

Firstly, it is accepted that at least some hallucinogenic symptoms of schizophrenia "arise from pathological activation of neurocircuitry involved with ... perception" due to corticocortical (inter-cortical) connectivity disruption (Hoffman and McGlashan, "Neural Network Models of Schizophreia", 2009). Are these hallucinations related to (and could they arise from) the same low-level mechanisms as those which cause the hallucinations and altered personality in mid-to-late Alzheimer's disease, or in drug-induced hallucinations, e.g. following LSD ingestion?

In other words, is it possible to create one universal low-level model for all types of hallucination and prove that all hallucinations, regardless of high-level disease/drug causes, arise from the same low-level activations of perceptive neurocircuitry?

Secondly what, if any, differences there are there between dreaming in normal brains and dreaming in those with AD or schizophrenia? Have there been any studies comparing AD/schizophrenia dreaming with normal dreaming? What significances do any differences have, if there are any? This could start off quite a large body of research if interesting differences are found.

In fact, there's a third question.. according to my Good Brain, Bad Brain neuroscience module, Down's Syndrome sufferers tend to develop AD at a much earlier age (sometimes in their 30's) -- what structural differences in the brain can cause this? Are these pointers to whatever structure it is that breaks down at the onset of AD?

But for now, I'm still just practising implementing some different types of neural networks, so such thinking can wait til later :)

Ambition or achievement?

I have an abstract! I've just submitted my first report (RSMG1) stating that I will be working as follows:

Computational Modelling of Neurological Disorders

This work will begin with the undertaking of a general literature review to better understand and attempt to synthesize the field of computational modelling of neurological disorders, and recent developments within it. Disorders currently of particular interest are Alzheimer's disease and schizophrenia. The literature review is expected to identify a number of current computational models focussing on different aspects of the various disorders, as well as shortcomings of these models and potential ways forward. The main contribution of the work will be the development of improved and refined models that better represent the underlying biological processes. The work will involve testing of the improved models against empirical evidence (either from medical literature or specifically-commissioned experiments) to ascertain the level of support which can be given to them. The approach adopted will involve various techniques from the fields of neural computation and computational neuroscience, and possibly evolutionary computation.

Next steps are a little unclear though, as officially I have nothing to submit until May 2010 now. By then I should have a "brief summary of work done so far, including literature reviewed, any coursework taken, programming languages learnt, talks given, etc." so that should give me some ideas where to start at least.

You may notice I've added references to schizophrenia in my abstract. This is largely because a lot of the current computational modelling literature focusses in this area. Now, schizophrenia is a fascinating disorder, but the amount of attention being devoted to it seems disproportionately large given that a majority of people (thankfully) will not suffer from or be influenced by it; something which unfortunately cannot be said for dementia.

I'm currently in two minds: I can either go down the currently-fashionable route of schizophrenia modelling, which could be easier given the current work being done in the field, and just try to achieve a PhD out of it as quickly as possible. Or I can stick with the less-fashionable field of dementia and Alzheimer's disease and ambitiously hope that I can come up with something really profound during my three years of research that might make a real difference to a lot of people.

Ambition or achievement?

Of course there is a third way... By studying the work being done in schizophrenia modelling I may be able to apply new models emerging from that field into the field of dementia, and through this contribute to both fields. This may prove difficult (not least because I'll have to learn about both disorders in great detail) but could be the most rewarding path to take.

Well, in the meantime I have a handful of interesting papers to read, but it's probably about time I dusted off my programming gloves and neural networks books and bashed out a few multi-layer perceptron and Hopfield networks to get started.

First month

Well, my PhD is now officially underway. It's been about three weeks since I started, and I now have an official title, Computational Modelling of Neurological Disorders, although work on an abstract for my Research Student Monitoring Group first report is progressing slowly (mainly due to my current seeming inability to sit still and read a paper for more than ten minutes at a time).

The change of title is significant; my original proposal was under the title Modelling the Robustness of Musical Memory in Dementia, however it's looking increasingly likely that -- although a valuable future route to take my research down, and certainly an interesting question in its own right -- it may be a bit too much to attempt for the duration of a PhD. By specifying "Computational Modelling" I create a stronger link with the Computer Science part of the PhD, and by broadening my scope to "Neurological Disorders" I can afford not to have to investigate the music question after all.

The original question was along the lines of "How is it possible that an Alzheimer's sufferer can forget their own name and identity, yet still be able to sing along to songs heard in childhood?" but initial investigation shows that one of the principal reasons for this ability lies in the vastly distributed nature of music processing in the brain, and particularly music's strong links to emotion. Investigating and modelling this in itself is likely to take years of research, so whilst remaining an interesting question, I may have to limit myself to the old tried-and-tested model of scientific investigation for a PhD:

1. Read current knowledge
2. Implement a model from the literature
3. Test and improve model
4. Write thesis
5. Collect Nobel prize

I'll let you know how it goes.