Open neuroscience talk at the University of Sussex next Wednesday - by me :)

I've been asked to give a talk to the Artificial Life Reading Group (Alergic) at the University of Sussex next Wednesday (Oct 6). I'm gonna take the opportunity to articulate my understanding of brains as spiking attractor networks that seek operant control through dopamine-mediated reward learning. This is actually a simple mix of old ideas, but my model system - the buccal ganglia - and experimental technique - the multielectrode array - allows me to quantify and visualize the various components of the theory. I'll also have the opportunity to test some new thoughts on brain-computer interfacing.

Network dynamics, dopamine and operant control: lessons from the molluskan buccal ganglia

Time: Oct 6, 16:30-18:00

Abstract

Multielectrode array (MEA) analysis of molluskan nervous systems is an experimental technique recently developed at the University of Sussex (Harris et al., 2010). Here I discuss our current understanding of the molluskan buccal ganglia, with examples from the MEA work, and relate it to more general theories of network dynamics, pattern generation, dopamine-mediated reward and operant control. Variance in the neural pattern for feeding behaviour, which the buccal ganglia continue to generate in vitro, allows the brain to search for optimal feeding strategies in changing environments (Brezina et al., 2006) and can be considered a rudimentary form of free will (Brembs, in press). I argue that this ability to generate variable, reward-sensitive motor output is a central function of brains, and discuss experimental and computational approaches toward a better understanding of it.

References: Brembs (in press) Proc of the Royal Soc; Harris et al (2010) J Neurosci Methods 186(2):171-8; Brezina et al (2006) Neurocomputing 69(10-12):1120-1124.

The talk will be recorded and will hopefully be available on the iPlant channel in a few weeks but please get in touch if you are in the UK and would like to attend the talk (and the enjoyable post-talk discussion in the bar). RSVP on Facebook here.

Spontaneous and dopamine-driven brain activity

Last weekend I started working on a video aimed at giving viewers a feel for the beauty and complexity of the network activity we record in the lab (see this entry for details). The trigger for this was meeting a friend who was able to create for me the Java code necessary to run the spike rate data through JFugue. My own attempts at this had not been completely successful. Simultaneously, Björn Brembs started posting a series of excellent and quite closely related videos over on his YouTube channel, which kept me motivated (also, being able to point to videos from Bill Kristan's lab certainly helped convince my supervisors of the wisdom of the idea).

There will be more of these, at least I intend for there to be, because it was a lot of fun to make and there are a range of variations and improvements on the audiovisual presentation to explore, as well as a whole bunch of specific topics I'd like to address. But for now, this is it.

A brain and a robot walk into a bar.. (MEA2010, day 3)

Not surprisingly there are a few posters here on two-way connections between robots and neurons cultured on multielectrode arrays (MEAs). One of them offer an open source software package to do it, called Cult2Robot. The authors use the software to let a culture of neurons on an MEA move a robot in four directions and avoid obstacles, all through a Bluetooth connection. Spike rates in the culture are monitored and whenever they cross a threshold at one of the four edges of the square MEA the robot goes in that direction. If the robot's sensors detect an object in any of the four directions, electrical stimulation is applied to neurons on the opposite site of the array, making them more prone to fire and bring the robot away from the obstacle.

It's a simple principle but it illustrates a point I've been trying to formulate for months:

1. Activity in brains and neural networks can be understood as a fixed number of neurons with a spike rate 0-200 Hz (120.000 neurons in this particular culture) 
2. Some patterns of activity result in defined actions (here supra-threshold activity along an MEA edge results in ipsilateral movement) 
3. Some actions are adaptive, others maladaptive, depending on the circumstances 
4. Given adaptive sensory- and/or reward-feedback, neurons change their activity to produce more adaptive output (here objects are avoided by stimulation of neurons with contralateral output)
5. The number of adaptive activity patterns a network can reliably assume in the context of changing sensory- and/or reward-feedback is a measure of its operant control (in brains we call this creativity, intelligence, self-discipline etc; as an infant learns new words, its operant control increases)
6. By using sensory- and/or reward-feedback protocols and good electrophysiological or brain imaging techniques we can map the dynamic range of activity states a brain or network can assume and explore/model the network properties that determine its degree of operant control

A pressing question is how adaptive various networks can be. Could we for example program the MEA-culture-robot above to move not just in four directions but in the 360 directions of a circle? Could the network learn that certain spatial and/or temporal patterns of network output drive power-moves in the robot (like jumping, climbing or crawling) that scale difficult obstacles, the presence of which might be indicated by specific sensory feedback patterns? The link between specific obstacles and appropriate outputs could be strengthened by application of dopamine... you get the idea. Moreover, to make use of the rich and variable activity of neural networks, they should be given the ability to control the robot's actions along continuums like amplitude, duration, and correlation with other actions. And remember, its not just academic curiosity driving these explorations: a constant theme of this conference has been neural prostheses, and the ability of human brains to generate and respond to many arbitrary patterns of activity along continuums is exactly what gives them the ability to control and respond to brain computer interfaces that restore function and improve lives daily, all over the world, but which are still very immature and problematic.

Brains as spiking attractor networks

Thank you to everyone who commented, online and in person, on my previous post. I hope this updated version will address the issues you raised, but I warmly welcome further comments. If the following is good enough I'll use it in a video entitled 'Basic neuroscience' that I'm making for YouTube, and in my analysis of neuronal multielectrode array data (Harris et al., 2010).

What I'm after is agreement on a simple but formally correct quantitative framework for describing brain activity. What determines the spiking (1) or non-spiking (0) of a neuron at a given time t is, as several people pointed out, an enormously complex interplay of cellular and synaptic processes, and it is the job of neuroscientists to discover the conditions (e.g. synaptic weights) necessary for artificial neural networks to replicate the spike patterns of biological brains. However, the spiking/non-spiking binary is clearly the most salient computational property of neurons, and in multielectrode recordings it is often the only value we can reliably detect.

Several people pointed out that the state of a brain (X) at time t, is a vector (an array of values) rather than a sum, i.e.

X(t) = [V₁(t), V₂(t),... V_N(t)]

where N is the number of neurons in the brain and V is the spiking (1) or non-spiking (0) of a particular neuron. Thus, a brain containing 100 neurons has 2¹⁰⁰ possible states, e.g.

X(t₁) = [ 0 1 1 1 1 0 0 0 1 0 1 1 1 1 1 1 0 0 0 0 0 0 1 1 0 1 1 1 1 1 1 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 1 0 1 1 1 1 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 ]^T

X(t₂) = [ 0 1 1 1 1 0 1 1 0 0 1 1 1 1 1 1 0 0 1 1 1 1 0 0 0 1 1 1 1 1 1 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 1 0 1 1 1 1 0 0 0 0 1 1 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 ]^T

X(t₃) = [ 0 0 0 0 1 0 1 1 0 0 1 1 1 1 1 1 0 0 1 1 1 1 0 0 0 1 1 1 1 1 1 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 0 0 1 1 0 0 1 1 0 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 ]^T

where red highlights a change in neuronal activity from the previous time point, and T stands for transpose (change from row to column) since the states of a neuron at successive time steps are usually arranged in a row (e.g. a raster plot, spike density function or voltage trace).

So far so good, but now to the heart of the matter: how do we use this simple framework to describe the complex brain activity manifest in cognition, behaviour and key neural processes?

The 2¹⁰⁰ (1.2676506 × 10³⁰) possible states of a brain with 100 neurons are referred to as its state space. Some of these states are not possible under biologically plausible conditions, and only a fraction of all biologically plausible states will be expressed during the lifetime of a brain. The order in which brain states occur is referred to as the trajectory of the brain's activity in state space. If we plot against one another the spike patterns (e.g. spike density functions) of two or more antagonistic neurons or neural networks, such as those that drive walking, breathing or swimming, they form a cyclic trajectory in state space, called an attractor, e.g.

Cyclic attractors in state space. After displacement (blue) one system (a) returns to its main trajectory while a different system (b) switches to a different trajectory, which drives a different behaviour. Figure from Briggman and Kristan (2008) Multi-Functional Pattern Generating Circuits.

Some commenters questioned the usefulness of the attractor concept for understanding complex cognitive processes and behaviours, such as making coffee, that involve large brain networks. In the absence of an alternative though, I still think the attractor model is worth pursuing, because the activity of sensory and motor neurons is tightly constrained by complex behaviours, and can therefore be described as attractors in the state space of the brain. In making coffee for instance, extending the arm and grasping the kettle must precede pouring hot water into the cup, and all three actions map directly onto tuning curves of specific neurons in the hand and arm regions of the left (typically) motor cortex. Therefore, these neurons will generate very similar spike patterns every time the action is performed. I believe the activity of neurons not directly constrained by the sensory and motor demands of the task, and therefore not obviously part of the behaviour, will nevertheless be selected for its ability, through direct and indirect synaptic connections, to create and maintain the required attractor among the sensory and motor neurons.

In terms of the present framework the question is, for each neuron or group of neurons N, how similar their spike patterns V have to be, morning-to-morning, for coffee to be successfully made. Neurons that are constrained in some way can be considered part of the coffee-making attractor. We can approach the question for instance by hypothesizing that the activity V of neurons in the pre-motor cortex will be more constrained than that of neurons in the temporal cortex, and then use EEG to measure variability in these regions on successive mornings. In animal models the recordings could be at the level of individual neurons and perturbations could be introduced. The relationship of the spike pattern of a neuron to an ongoing attractor may be complex and depend on the spike patterns of many other neurons, and some behaviours clearly require more neuronal resources than others, but these are empirical questions that, to me at least, seem fruitful.

Some further questions:

• In what ways can some neuronal activity V_N(t) be said to participate in an ongoing attractor?
• How do we characterize, mathematically and visually (or perhaps audibly) an attractor in N dimensional space?

Lorenz attractor

How should we think about brains?

This is a simplified but, I hope, formally correct framework for thinking about brain activity. Please help me by pointing out errors and obscurities.

The human brain is a network of some 100 billion neurons. The generation, recognition and modulation of neuronal activity patterns by neural networks in the brain are physical manifestations of thoughts, feelings, actions and sensations. Efforts to understand, reproduce and communicate the activity and capacities of brain networks are hampered by the inaccessibility and extraordinary complexity of these networks. However, a general understanding of brains as dynamical systems has emerged.

Consider a brain consisting of N neurons with an activity state V indicating activity (1) or inactivity (0). At time t the state (X) of this brain is equal to the activity state V_i(t) of neuron i, where i = 1,2,...N labels the N neurons of the brain. Formally, we write

which simply means that at a given time (t) the state of the brain (X) is the sum of the activity or inactivity (V) of its N neurons (i). The neurons are said to be the state variables of the system. Thus, a brain consisting of 100 neurons has total of 100^100 possible states, e.g.

Possible brain states. Red indicates a change in V from the previous time point.

The 100^100 possible states of this brain are referred to as its state space. Given biologically plausible rules for neuronal activation however only a fraction of the 100^100 states are practically possible. For example, it is not practically possible for all the neurons in your brain to become simultaneously active, or for all neurons in one hemisphere to be active and all neurons in the other hemisphere inactive. Furthermore, only a fraction of all practically possible brain states will be expressed during the life of a brain. For example, although it is practically possible for your brain to learn and express brain states associated with the articulation of words in Swahili, in actual fact you probably will never express those states. Conversely, some classes of brain states may recur frequently (see below).

The order in which a brain expresses its various states is referred to as its trajectory through state space. In some brain networks the trajectory is rhythmic and continuous. For example, neurons in the Pre-Bötzinger complex drive breathing from the moment of birth to the moment of death, and can only be temporarily displaced from their oscillatory trajectory in state space. Networks driving episodic rhythmic behaviours such as chewing go into a particular oscillatory trajectory when the behaviour is expressed, but may also be quiescent for long periods of time, or express different trajectories that drive other behaviours involving the same muscle groups (e.g. speaking, licking, coughing).

Rhythmic trajectories (also referred to as neuronal oscillations or "limit cycles") through a 3D state space. After displacement (blue) the system (a) returns to its main trajectory or (b) switches to a different trajectory. Figure from Briggman & Kristan (2008) Multi-Functional Pattern Generating Circuits.

A region or path in state space that attracts nearby network trajectories is called an attractor or attractor basin. For example, consider making coffee in the morning: this behaviour is a precise trajectory through the state space of your brain, involving movement to the kitchen, location of appropriate equipment, pouring hot water into a cup etc. If you're like me, your brain will revolve in the basin of this attractor untill coffee is produced regardless of where you wake up, what the time is, what you dreamt etc. In other words, your brain tends to travel through the coffee-making attractor regardless of its starting point in state space in the morning. This is not to say that every behaviour necessarily corresponds to an attractor - by relying on environmental cues a neural network with just one attractor can express several problem solving states (Buckley et al., 2008) - but it is a very useful simplification for thinking about brain activity.

Attractor basins (green) and hills (red) in a 2D state space.
The brain state spaces discussed here have N dimensions.

Attractors are also useful for understanding sensory classification, memories and habits of thought. For example, sensory categories are thought to result from neural networks associated with aspects of a class of objects (e.g. sensory networks responding to the sight, bark or smell of dogs) being repeatedly activated together and thus linked through Hebbian plasticity. In other words, the networks form an attractor "of" the abstract concept 'dog', such that all nearby brain states (e.g. neurons in the auditory cortex responding to a bark, or neurons in the visual cortex responding to a furry tail) will tend to converge on the same attractor basin and classify as instances of 'dog'.

That's it for now. Topics for future blog posts: How are attractors continually formed and dissolved in the state space of brains and networks? How can we study neural network dynamics and attractors? How can we visualize them, quantify them and use them in computing and in our everyday understanding of ourselves and others?

Studying the formation of patterns in the invertebrate nervous system

Now that my first article has been published (Harris et al., 2010) I should finally able to talk and write about my PhD research. Pre-publication attempts to do so earned me a few slaps on the wrist.

My work builds on the electrophysiological analyses of molluscan nervous systems that my university has specialised in for more than three decades (Benjamin & Rose, 1979; Benjamin & Kemenes, 2008). The use of relatively simple invertebrate systems as models for understanding brains in general goes back to the characterization of the action potential (Hodgkin & Huxley, 1939; 1952), through to the molecular mechanisms of synaptic memory formation (Kandel, 2001) and the study of dynamic neural circuits (Selverston & Ayers, 2006; Briggman & Kristan, 2008; Elliot & Suswein, 2002). Whereas the vertebrate brain, for the most part, remains impossibly complex, invertebrate nervous systems are small (20.000 neurons in the molluscan brain - 100 billion neurons in the human brain) and have large, relatively accessible neurons that survive and remain functional in vitro. Moreover, the location, connectivity and function of invertebrate neurons are virtually identical among members of a species, much unlike the individualistic vertebrate brain.

"In the mammalian brain, the precise relationship between the dynamics of individual neurons and functional networks remains extremely complex. The main reason for this is a lack of knowledge of the detailed cell-to-cell connectivity patterns and the biophysical properties of the individual neurons and their synaptic connections. Attempts to understand brain dynamics by large-scale modeling have been attempted frequently but without knowledge of the detailed parameters such as the number and kind of synaptic connections, the results have been disappointing (e.g., Foldiak and Young 1995). Moreover, each physiological synapse may result from numerous anatomical synapses that may have complex spatial geometries in neuronal branches. The numerically less complex microcircuits of invertebrates have neurons and synapses which are identifiable from animal to animal. Therefore, a much more detailed understanding of neural circuit dynamics is possible." - Selverston & Ayers (2006)

Unlike many other sciences, neuroscience lacks what Gerald Edelman calls a "global brain theory", a conceptual and mathematical foundation for understanding brain and behaviour. In this context, the aim of invertebrate neuroscience is the development of principles and computer models that describe the operations of biological neural networks in general. The tacit assumption here is that vertebrate brains work like their spineless cousins, and though this has proved largely accurate on the molecular and cellular level, it has not been conclusively established on the poorly understood level of neuronal networks, although there are tantalizing similarities (Yuste et al., 2005; Grillner 2006).

Of particular interest, at least in our lab, is the generation, by these networks, of the complex patterns of electrical activation that underlie the adaptive behaviour of living organisms. Invertebrates show a striking degree of adaptive variability in the way they behave, which through processes of sensory feedback, neuromodulation and reinforcement learning allows them to behave intelligently in a constantly changing environment (Horn et al., 2004; Selverston & Ayers, 2006), an ability currently unavailable to computer software and robotics. We are also concerned with reinforcement learning on longer time-scales, and with the competitive and cooperative interactions between neural networks that allow organisms to select different behaviours in different situations.

We study these processes primarily through a detailed analysis of a network of neurons that control the feeding musculature in the freshwater gastropod Lymnaea stagnalis. The network comprises some 500 neurons in two almost identical ganglia (shown below, black dots are extracellular electrodes).



The network occasionally generates a complex pattern of activation (below), which in vivo drives sequential muscle contractions that extend a tongue-like structure, scoop food into the mouth cavity and swallow it (a video of this behaviour can be found here). Remarkably, the neural network continues to generate the appropriate sequential activation of motor neurons for hours even when the brain is isolated and monitored with microelectrodes. Each cycle of activity is then called a "fictive" feeding cycle. Below are five such cycles, recorded on 14 of the extracellular electrodes in the figure above (note the cycle-to-cycle variability).



A number of methods have been developed for experimental initiation of feeding cycles, including intracellular activation of cerebral command-neurons, dopamine perfusion, and perfusion of sensory tissues with gustatory stimuli.

I'll return to this work in future posts, because there is an enormous number of questions and avenues for research here, and I know some of my readers work on similar projects. There are also a number of direct links between this research and reinforcement learning, motivation, dopamine, the iPlant, and the overarching goal of developing a qualitative, easily communicable understanding of the living human brain.

Notes on electrical rhythms in the brain

I'm reading George Buszáki's Rhythms of the Brain at the moment. Here are some notes on what I've read so far. Please correct me if I've got something wrong.

Electrical fluctuations in the cortex are organized into rhytmic oscillations at different spatial and temporal scales. The resting cortex is characterized by oscillations primarily in the alpha band (8-12 Hz, the brain's 'default network'). The active (i.e. behaving, percieving) cortex is characterized by oscillations primarily in the gamma band (25-100 Hz). Buszáki and others argue that cortical neurons that synchronize their membrane oscillations in the gamma band 'bind' their respective functions (e.g. visual feature detection) together into cognitive processes (e.g. object perception). Such formations of neurons are called neuronal groups or assemblies. Particularly striking are neuronal groups in the gamma range emerging in the prefrontal cortex for the duration of time that human patients are asked to hold items in working memory.

(alpha)

(gamma)

Cortical neurons sponaneously synchronize their membrane oscillations in the gamma range and form transient neuronal groups even in the absence of stimuli. This is, at least in part, due to the time constants of GABA curents, synaptic delays and synaptic potentiation. Buszáki writes:

"If neurons are already engaged in internal synchronization, the external stimulus will compete with the central oscillator, and the coutcome depends on the relative timing and strenght of the external input and the propensity of the internal oscillator. The stimulus may be ignored, or it may enhance or quench the internal oscillation." p.255

In other words, the effect of a stimulus on cortical activity depends strongly on the prior state of the brain. This explains the significant variability in brain activity (e.g. on EEG/MEG/fMRI) seen within and between subjects in response to invariant stimuli. Buszáki laments the fact that this variability is usually averaged out and treated as 'noise'. Björn Brembs often makes a similar argument.

Stimuli interact with ongoing cortical activity in various ways. Whereas a weak stimulus may reset the phase of ongoing oscillations, a strong or salient stimulus may completely change the type and distribution of oscillations in the cortex. Several studies have found that strong ongoing oscillations in the gamma, theta or alpha ranges prior to stimulus presentation promote efficient memory encoding. A stimulus that resets a strong rhythm presumably has a larger impact on brain activity than one that resets a weak rhythm. The presence or absence of strong rhythms in the brain is directly related to states of attention and catecholamine concentrations.

A complete cell-by-cell recording from a human brain

Two thought experiments have been bugging me lately. I've not made much progress on either one so I thought I'd write them down here and consider why they're interesting and maybe get some input.

The first thought is, what would be the value of a complete cell-by-cell recording of electrical activity from a human brain? Not sub-threshold resolution, but rather 100 billion channels recording the location and time (in milliseconds) of every action potential fired in the brain, as if an extracellular electrode had been wrapped around each and every axon. 30 billion channels (firing patterns) for the cortex, 10 billion channels for the subcortical forebrain (hippocampal region, basal ganglia and thalamus), 1 billion channels for the brainstem and spinal cord, and the rest for the grainy cerebellum. I asked this question on Mahalo Answers several months ago but nothing came of it.



What I'm really after I guess is the long-term purpose and potential of brain imaging (everything from fMRIs to neural networks cultured on multielectrode arrays). How would neuroscience change if brain imaging data was not limited to diffuse blobs, squiggly lines or a handful of individually recorded neurons? It seems clear that we would learn a tremendous amount, but what exactly?

One impulse is to model the data, but how would we do this without knowing the connectivity of the neurons? Seth (2005) and others would argue that we could work out causal, if not synaptic, connectivity through statistical analysis of the information-content of the firing pattern of each neuron relative to the others. If we had the computational resources to do that, what might we learn? People interested in evolutionary algorithms on the other hand might argue that we could use simple spiking model neurons (Izchikevich, 2003) and let synaptic weights and properties evolve until the model behaved and responded to input like the recorded brain did. (I just made that up, I have no idea if it makes sense. A caveat at this point is that I'm NOT a mathematician or anywhere near as familiar with neural network modelling as I'd like to be. I'm sure there are abstract functional rules and graph-theoretical relationships or whatever that one could retrieve from a complete cell-by-cell recording of the dynamics of a human brain. But I want them named. Control theory? OK, what might happen in control theory? Neural network software? OK, how would artificial neural networks change? Etc.)

A second impulse is to compare the complete cell-by-cell data with existing knowledge and models of the brain. We know a fair bit about gross anatomical and functional structure. We have a lot of rapid EEG and MEG activity which we don't understand but which seems relevant to various cognitive activities. We'd be able to learn an awful lot about what various states of heightened electrical (e.g. gamma buzz, hippocampal theta rhythm) or metabolic (e.g. BOLD-response) activity look like 'from the inside', on a cell-by-cell basis. But then what? What sort of understanding would we gain from turning a BOLD blob into the electrical chatter of millions of neurons?



A third impulse is to look for neural correlates of sensory, cognitive and behavioural events. The English language is composed of 40 or so phonemes (smallest units of sound), each of which presumably corresponds to a specific posture of the mouth and larynx. Each such posture should correspond to a more or less fixed pattern of activity in spinal C1 motorneurons and perhaps even in the motor cortex. It should therefore be possible to ask the brain/person to name an object and PREDICT the activity of some of its/his/her neurons from the word you expect it/him/her to utter. A similarly fixed pattern of activation would be induced by photons impinging on the retina, which might allow one to predict the activity of individual neurons in visual thalamus and the visual cortex. Similar predictions should be possible in the auditory system. We also know of a few deep brain regions where activity of individual neurons is predictable, such as the bursts of midbrain dopamine neurons in response to unexpected rewards. Would these fixed points of neural activity in combination with simple or complex tasks allow us to understand/predict a bit better the activity of the rest of the brain? Not sure where I'm going with this one but I think it's important.

Yet another impulse would be simply to ask the brain/person to describe its/his/her subjective experience/internal processes as cell-by-cell recordings in regions of interest are taking place. Many theories in cognitive and affective neuroscience could be directly applied ('Let's see, you're feeling hungry, a little bit anxious and quite indecisive, right? Not hungry? That's weird, your NPY neurons are quite active..'), and many new ones could be developed. Moreover, I strongly believe the cultural impact of people being able to see and play with visual reconstructions of subjective states and cognitive events would be enormous. Right? What kind of visual reconstructions? Of which states or events? How would people come in contact with them? Statistical analysis of 100s of neurons in the leech nerve cord allowed Briggman et al. (2005) to identify 'decision-making' neurons whose activity predicted which of two possible activity states the nervous system would adopt in response to a stimulus. Benjamin Libet continues to annoy the world with his decision-predicting EEG events which supposedly occur a few hundred milliseconds before awareness, and recent versions of the experiment claim successful decision-making forecasts as early as 10 seconds before consciousness. These experiments use crude scalp electrodes. What awesomely weird conclusions about human beings would a complete cell-by-cell recording generate?

And so on. I'll return later with the other thought experiment, which is a (somewhat) more realistic and practical version of the same question.

Chronic electrical stimulation of cultured hippocampal networks increases spike and burst rate, and changes burst structure

Brewer GJ, Boehler MD, Ide AN, Wheeler BC (2009) Chronic electrical stimulation of cultured hippocampal networks increases spontaneous spike rates. Journal of Neuroscience Methods.



Neural cultures developing in vitro lack the incoming stimulation/information of their natural (in vivo) environment. In this paper Brewer and collegues applied chronic electrical stimulation to cultures of E18 rat hippocampal neurons developing on 60 channel multi electrode arrays.



30 uA paired pulses (50 ms ISI; biphasic, 100 us/phase duration, positive first) with a 5s wait between pulse pairs were delivered to 30 of the 60 electrodes, one-by-one, in a semi-random sequence for 0 (control), 1 or 3 hrs per day, at 7 11 12 14 18 19 and 21 days in vitro. There were five cultures in each condition. Three minutes of activity were recorded and analysed on day 21. My only problem with these methods is that the cultures recieved stimulation just before recordings were made. This confounds the 'chronic' impact of stimulation during network development - the observed effect could be acute. Recordings further from the time of the last stimulation sequence are required to rule out acute effects.



Interestingly, 1 h of stimulation had a greater impact on almost all spike and burst measures compared to 3 h stimulation, which often was not significantly different from control. Spikes per burst, burst duration, burst rate (bursts/minute) and total spike rate were all up to 2-fold higher in the 1 h stimulation group. However, intra-burst spike frequency was significantly higher in the 3 h group, although bursts in this group were shorter than in the 1 h group. In other words, bursts in the 3 h stimulation group were sharper, more distinct. This was perhaps the most interesting finding in the paper as far as I am concerned. In all groups, 90% of spikes occurred in bursts, with no difference between conditions. Other reserachers have reported similarly high levels of bursting in neuron cultures. "can we reasonably conclude that information coding occurs in bursts and less so in individual or smaller groups of action potentials?" ask the authors. If so, chronic electrical stimulation during development certainly has considerable impact on information coding.



The authors also make some interesting observations on the impact of distance-from-stimulating-electrode on spike rate. Weirdly, they find that spike-rate decreases with increasing proximity to stimulating electrodes in the 1 h stimulation group, but increases (albeit modestly) with increasing proximity to stimulating electrodes in the 3 h stimulation group. Another piece of the puzzle.

In conclusion, the presence of incoming stimulation/information clearly changes the behaviour of developing neural networks. Izhikevich and Edelman similarly found that input was required for activity to emerge and maintain itself in their vast thalamocortical network models.

It's hard to know what to make of these cultured or modeled, essentially random neural networks. They need input and can discriminate complex output in ways your average CPU cannot. They also produce complex, often chaotic output that changes over time with some regularity (Wagenaar's report on superbursts is relevant here). But the question of how to link input to output, output to input, in ways that produce self-organizing, useful, meaningful or even intelligent results has not yet been addressed (as far as I'm aware).

Riding a bike

I was 5 when I learned to ride a bike. I remember the place - a thin strip of asphalt surrounded by very green grass, just beside a small patch of forest - where I first managed to ride it for a stretch without dad holding it steady and without falling over. Imagine my brain at the time. Imagine two pulsating groups of active neurons in each prefrontal cortex, slowly circling each other. One, oscillating slowly, projecting to the motor cortex right at the top of my head and down between the lobes, driving the oscillating contractions of leg and foot on the pedal. The other, pulsating a seemingly patternless pattern to the motor cortex below the upper sides of the skull, constantly adjusting the handle bar with a cramped grip, keeping the whole circus upright. Both groups, closely connected to eachother and to their mirror images in the opposite lobe, receiving a constant barrage of input: sight from the back of the brain and the colliculus, balance from the inner ear, kinetics from the spinal cord. Both groups constantly adjusting their output to maximize the flow of reinforcing dopamine from their respective midbrains. And that's why I remember it so well, that moment when the groups finally got the output right, for a time, and were showered with dopamine as all regions of the brain reported success. The dopamine reinforced them, and with them every other process that was active in my frontal lobes at the time - the location, the weather, the color of the grass.

After that of course, I've kept on biking, for years and years and years, milking those two groups for all the dopamine they were worth, until they were neat and trimmed and refined to a point where almost all the oscillation and rotation and complex feedback loops have been moved over to small, dedicated central pattern generators in my motor cortex and spinal cord.

A slice of shark brain with lemon

Politics fatigue. The US election is over and all the politics podcasts and news have lost their bite. Hillary as foreign minister, so what? The economy tanking, who cares? Keith Olberman's attempts to sound as excited about the political news now as he did before the election are off key and Democracy Now won't stop heaping shit on not-even-president-yet Obama. But even Obama is boring now. Another 'stimulus package', wake me up in a month or two, seriously, who gives a shit? (And all the while Israel is starving Gaza.. that still lights a spark somewhere, but is it being discussed? nooo, course not. Because, to paraphrase Obama; 'that shit is sacrosanct'.. yea he actually used that word.. right before bringing Rham Emanuel on board.. but I digress....)

So that's half my podcasts rendered impotent, half my daily entertainment gone. And to top it all Entitled Opinions go on hiatus. Shit. Post-election, post-SfN blues. Even Twit has lost it's umpf.

So, back to basics. Back to work. Think of the absence of distractions as an opportunity to take a closer look at what I'm working on. An intelligent circuit. A biologically inspired circuit. Biologically inspired circuit design.

Sooner or later we all come up against network theory. We all need large data-sets for strong patterns to emerge, but then the variables are too many and the maths are too hard. For this we need intelligent circuits.

NASA knows this. NASA even uses an artificial neural network to handle the plants they're bringing to Mars. But it's unlikely that it's a network modeled on a biological circuit. We don't have programs like that, not yet.

Attention-allocation in open real-time communication and neural networks

One of the many reasons I follow Steve Gillmor's work is that his interest, it seems to me, is in attention-allocation, which is also my interest. How can we discover and attend to those bits of information that, were we aware of their existence, would be of highest quality to us? Attention-allocation is a fundamental problem in all adaptive behaviour of course, but it has to be thoroughly re-imagined in relation to the internet in general - as Google did - and open, real-time communication in particular - which appears to be what Steve is doing.

What is the best way to discover and filter information online? Open, real-time messaging services like Twitter, Identi.ca, FriendFeed and (sooner or later) Google Reader offer two solutions: friends and track. A 'friend' here is anyone whose activities and recommendations you find sufficiently interesting to follow indiscriminantly, be they IRL friends, a Mars rover or unnamed members of a common-interest group on FriendFeed. A 'track' is a word the use of which you find sufficiently interesting to follow indiscriminantly: I use TwitterSpy and IdenticaSpy to have gTalk alert me whenever words like dopamine, serotonin and chrisharris are used in a conversation anywhere on Twitter (some delay) or Identi.ca (instantly). FriendFeed, we are told, will soon enable track on its network, and all three networks are growing fast.

As the aim of my own work in electrophysiology is to understand and model how neural networks achieve intelligent attention-allocation, the question inevitably arises: how do these two forms of attention-allocation relate? Do neurons have friends? Do they track? Nervous systems use dopamine, a scalar signal that increases signal-to-noise and promotes memory formation, to self-organize and allocate attention. Could a similar signal be implimented in open, real-time messaging networks?