summary:
Alain Destexhe. Dendrites Do It in Sequences (24 September 2010)
Science 329 (5999), 1611.

article:

Tiago Branco, Beverley A. Clark, and Michael Häusser. Dendritic Discrimination of Temporal Input Sequences in Cortical Neurons (24 September 2010)
Science 329 (5999), 1671.

Consider a a cortical neuron in V1, layer 2/3, whose output shows sharp orientation tuning. What are the orientation tunings of the most important inputs to that neuron? What is the spatial distribution of these inputs in the neuron’s dendritic tree?

Here’s three possibilities. (1) You might expect the neuron to collect inputs which are broadly tuned for that same orientation (the “weak-bias model”). (2) Or, you might expect that the neuron as a whole collects inputs with various tunings, but that each dendritic branches would tend to collect inputs with a certain orientation. (3) Or, neither of these could be the case; maybe the inputs just take all sorts of orientations, randomly distributed among the dendritic tree. Here a picture of these possibilities from the News and Views:

Jia, Rochefort, Chen, and Konnerth analyzed the orientation tuning of such neurons as well as the orientation tuning of the calcium dynamics within the neuron’s dendritic tree. Their results support the third option (inputs with heterogenous tuning, spatially mixed).

While hyperpolarizing the cell, they found “calcium hotspots” in the dendritic tree, that is, places where there was a noticeable, localized calcium signal in response to stimulation. They then analyzed the orientation tuning of these hotspots. Figure 3b shows three hotspots and their calcium response to various drifting gratings (oriented visual stimuli):

Figure 3c shows what the orientation tuning was for all of the hotspots in one neuron:

The main results are that the orientation tuning of the hotspots is heterogeneous (all sorts of different tunings are found), and that there is no discernible spatial pattern to where the differently tuned hotspots are located within the dendritic tree.

Furthermore, they compared the histogram of the orientation tuning of hotspots between sharply tuned neurons and broadly tuned neurons, and found that they were similar, supporting the hypothesis that whatever it is that makes some neurons have sharper orientation than others tuning in their output, the cause is something other than having sharper orientation tuning in their inputs. Fig. 4d (OSI stands for “orientation selectivity index”):

Here’s an excerpt from the Nature editor’s summary: “Whether…. tuning is already encoded in a neuron’s dendritic inputs or whether the neuron itself computes its selective response has been unclear….They discover that, while all neurons receive distributed input signals coding for multiple stimulus orientations, each neuron makes its own ‘decision’ as to the orientation preference of its firing output.”

Some cautionary notes: (A} the News and Views makes it sound as if this study established linear dendritic summation. As far as I can tell, the study didn’t test that directly. (B) above, I said that possiblity 3 is that the inputs are “randomly distributed”; in the study, however, although the distribution SEEMED random, it’s possible that it is just organized in some complicated way that made it look random. (C) I could be wrong about this, but as far as I can tell, there’s no guarantee that the calcium hotspots are the “most important” synaptic inputs; they might be ones which just happen to have a high density of calcium channels (D) they are only looking in about four planes of focus and getting about 13 hotspots per neuron, so this is only a small proportion of all of the synapses (E) even if the set of strong synapses showed heterogeneous tuning, there could be many weak synapses that all have tuning that matches the output tuning. (F) I defined the hotspots as “noticeable, localized calcium signal in response to stimulation”, but this is pretty subjective. The article does not exactly specify an algorithm which was used to pick out the hotspots from within their imaging data. All the methods has to say about it is, “Transient changes in Ca2+ fluorescence (?f/f) were systematically examined by an adaptive algorithm, which involved small regions of interest (ROIs) of 3?×?4?µm, noise filtering and pattern matching.”

A few days later, Henry Markram, leader of the Blue Brain Project at EPFL, sent off an e-mail to IBM CTO Bernard Meyerson harshly criticizing the IBM press release, and cc’ed several reporters. This brought a spate of shock media into the usually placid arena of computational neuroscience reporting, with headlines such as “IBM’s cat-brain sim a ‘scam,’ says Swiss boffin: Neuroscientist hairs on end”, and “Meow! IBM cat brain simulation dissed as ‘hoax’ by rival scientist”.  One reporter chose to highlight the rivalry as cat versus rat, using the different animal model choice of the two researchers as a theme.  Since then, additional criticisms from Markram have appeared online.

Find out more after the jump.

In the aftermath, IBM has stood behind the announcement, citing for Network World their team’s involvement with “Stanford University, University of Wisconsin-Madison, Cornell University, Columbia University Medical Center, University of California-Merced and Lawrence Berkeley National Laboratory” as defense.  Who are the researchers they are standing behind?  According to Modha’s blog, they are:

• Stanford University: Brian A. Wandell (Prof of Psychology, Electrical Engineering), H.-S. Philip Wong (Prof of Electrical Engineering)
• Cornell University: Rajit Manohar (Prof of Electrical Engineering)
• Columbia University Medical Center: Stefano Fusi (Prof of Theoretical Neuroscience)
• University of Wisconsin-Madison: Giulio Tononi (Prof of Psychiatry)
• University of California-Merced: Christopher Kello (Prof of Cognitive Science)

For this neurodude, it is interesting how this disagreement may be symbolic of the gap that still remains between neuroscience and AI.  Markram is a neuroscientist turned technologist, while Modha is a computer engineer who wants to derive technological insight from biological  systems.  They are approaching the ideal of reverse engineering the brain from very different perspectives, and its only natural that they value different milestones.  The IBM team, even with the additional professors on their team, still lacks mainstream neuroscientists to help validate their claims.  That being said, the public realization of this could be a positive thing for both fields.  Although some frustration has resulted from this, this could be a great opportunity for the breakdown of walls between these fields.

In the end though, it does seem like Markram has a point.  The IBM press release clearly went too far.  Whether the angry public e-mail was the best strategic way to make the point remains to be seen.  It will be interesting to see what the next move from the IBM team will look like.

Watch it online at the TED website.

I’m a fan of it because it is an open-access journal featuring a “tiered system” and more.  From their website:

The Frontiers Journal Series is not just another journal. It is a new approach to scientific publishing. As service to scientists, it is driven by researchers for researchers but it also serves the interests of the general public. Frontiers disseminates research in a tiered system that begins with original articles submitted to Specialty Journals. It evaluates research truly democratically and objectively based on the reading activity of the scientific communities and the public. And it drives the most outstanding and relevant research up to the next tier journals, the Field Journals.

I’m a big fan of the variety of specialty journals they have:

• Aging Neuroscience
• Behavioral Neuroscience
• Cellular Neuroscience
• Computational Neuroscience
• Enteric Neuroscience
• Evolutionary Neuroscience
• Human Neuroscience
• Integrative Neuroscience
• Molecular Neuroscience
• Neural Circuits
• Neuroanatomy
• Neuroenergetics
• Neuroengineering
• Neurogenesis
• Neurogenomics
• Neuroinformatics
• Neuromethods
• Neuropharamacology
• Neuroprosthetics
• Neurorobotics
• Synaptic Neuroscience
• Systems Neuroscience

Click on for some of my favorite passages from the Abbott piece.

Abbott uses the following problem of input decoding

Spike counts and neuronal firing rates are positive quantities. This simple fact has important implications for neural coding and neural circuits that provide a framework for thinking about a number of research directions taken over the past 20 years.

to highlight new work in synchrony, dendritic compartments, and balanced excitation-inhibition. This is probably the best part of the whole article. With some simple arithmetic, he motivates and explains solutions to the problem of correlating neural activity with real events.

The successes of circuit models (and principles of circuit models) in primary visual cortex:

We now have plausible mechanisms for how simple and complex cells obtain their basic response characteristics. Although no single consensus abouthow the circuits of primary visual cortex operate has arisen from this body of work, this may simply reflect the fact that multiple mechanisms contribute. In other words, many of these ideas are probably correct in one way or another, and the wealth of ideas in this field should be viewed as a success. Circuit-level modeling is now advancing beyond primary sensory areas (for example, Cadieu etal., 2007) and to the consideration of phenomena such as working memory through sustained activity (Amit and Brunel, 1997,Compte etal., 2000,Seung etal., 2000) and decision making (Wang, 2002,Machens etal., 2005).

And the dangers of an unhealthy obsession with connectomics:

What can we learn from the complete connectome or, indeed, a complete mathematical description of a complex artificial network model?

First, what can’t we learn? It is unlikely, for example, that we could deduce the task that the network was constructed to perform even if we were given the complete equations and connections of the model. If, along with this information, we were told what this task was, it is unlikely that we could figure out how the network performs it. If we somehow managed to make any progress along these lines, the people who constructed the network could probably provide us with another one that performs the same task but has a different connectome. In a similarway, biological systems may operate in a more variable manner than we have suspected, as has been stressed by Eve Marder (Marder etal., 2007). These issues are particularly true of a class of network models known as liquid state or echostate networks (Maass etal., 2002,Jaeger, 2003). In these models, the vast majority of interneuronal connections are not directly related to the task being performed (they are typically chosen randomly and left unchanged), the exceptions being synapses onto the output units of the network. Nevertheless, the tuned values of the synapses onto the output units can only be understood through their relationships to the random synapses. Such systems represent enormous challenges for conventional anatomical and physiological approaches.

The fact that the connectome of an artificial neural network does not typically tell us what the network does or how it does it should not be taken as an indication that this information is useless. Far from it. But we must be willing to be more abstract in our thinking. The important issue for an artificial network is not how it works but how it was constructed, which means what training procedures and modification rules were used to get it to perform a task. Although this information is not provided directly by the connectome, much can be inferred. For example, it is important to know whether the network has a feedforward architecture or has strong feedback loops. Other features of the network layout, whether it has hubs or bottlenecks, how many layers it contains, and its degree of heterogeneity, provide important clues as well. Obtaining a high-resolution connectome in neuroscience will be of great value, but artificial neural networks provide a cautionary tale that reminds us that scientific revolutions tend to render uninteresting as many questions as they answer. We will be fortunate if the connectome project does this for neuroscience, butas we launch ourselves into it we should appreciate that, as artificial neural networks appear to suggest, we may be asking the wrong questions.

Finally a major challenge for the future:

This is where I think the future lies in theoretical investigations of cognitive function. We must learn how to build models that construct hypotheses through their internally generated activity while remaining sensitive to the constraints provided by externally generated sensory evidence.

An amazing paper from Neuron demonstrating adaptive (circadian clock-governed) binning in the retina, based on dopamine modulation of gap junction (electrical) synapses between retinal photodetectors. During the day, abundant dopamine release weakens gap junctions coupling rods and cones together so that visual acuity is high. When light is scarce (at night), there is less dopamine and the electrical coupling between rods and cones is increased. This is analogous to on-chip binning in CCD (digital) cameras. Binning increases signal (in light-limited systems, eg. seeing at night) by increasing optical input area and by reducing single element noise (ie. noise at different photoreceptors should be independent) at the cost of resolution. So, the retina activates photoreceptor binning at night to boost low-light signals and deactivates it during the day to increase resolution. The dopamine comes from cells in the interplexiform layer, whose dopamine release is itself governed by melatonin projections.

Also, I never knew that gap junction strengths were directly modifiable. It looks like the D2 receptors are G-protein coupled to PKA, which acts on the gap junctions.

Neurons come in many shapes and sizes. Frequently, the shape of a neuron is characteristic to its type. Several theoretical papers have demonstrated that the shape of a neuron can crucially determine its pattern of activity, independently of other factors (Mainen & Sejnowski, 1996, for example). Several resources on the web such as neuromorpho.org and the Cell Centered Database are dedicated to maintaining repositories of different neuronal shapes (also known as morphologies).

Any computer scientist worth their salt, noticing this trend, is tempted to say: if neuronal shape is so important, maybe we ought to have good data standards to describe it. That’s just what a paper last year did. It surveyed the popular data standards for modeling, primarily in the NEURON and Genesis simulation packages. The result is a data standard called MorphML, which is part of a larger effort called NeuroML.

Neuronal shape is a weird data type for the computer science world, but I think an incredibly important and fundamental one for deeply coping with the complexity of real brain tissue. It seems to me that many areas of neuroscience research could benefit from the construction of more explicit models of the circuits they study.

However promising the field has been thus far, even the most accomplished neuroscientists will admit that we still do not understand how the brain really works. I would submit that the current reductionist nature of neuroscience has shed much light on the dynamics of how neurons work, but has to a far lesser degree shed light on how neurons process information. The difference between these two lines of inquiry is important for making progress in understanding how the brain works.

A computer, at its core, processes information through the physics of transistors, which are essentially switches that are either on or off. What makes transistors such a powerful foundation for modern computing is that they are controlled by electrical signals, which means that transistors can be controlled by other transistors and therefore structured into useful systems. Understanding the physics of transistors, how quickly they can switch from on to off, how their material composition affects their ability to switch, is crucial for building a microchip. However, this level of understanding is not sufficient to build a microchip. For that, one needs to understand how to structure transistors in such a way to produce digital computation.

Single transistors turn on and off. As a medium for constructing computer architectures, they are relatively straightforward to combine into complex circuits that perform useful functions of logic. Single neurons, quite a bit more complicated, have a vast repertoire of behavior that, among other things, involves integrating signals from multiple sources and sending signals to multiple recipients. It is not at all straightforward to construct explanations of how neurons combine into complex circuits to perform useful behavioral functions. Yet, this is the kind of explanation that neuroscience ultimately must seek in order to fulfill the promise of its potential to unlock the secrets of the brain.

Keeping in mind that computers are different that brains in many important ways, in some sense, we are still at the level of understanding the dynamics of the transistors in neuroscience. Cellular neuroscience, as found in journals such as Neuron, has concerned itself with the dynamics of neurons, rather than their role in processing information. This may seems like a bold statement to some; after all, decades of research has been conducted on sensory systems such as vision, and many aspects of the visual pathway are understood. Furthermore, lesion studies have been demonstrating that certain groups of neurons have certain functions throughout the last hundred years. However, despite the current push to apply information theory into the study of sensory systems, even the most cutting edge work in the field of neuroscience is still only just beginning to incorporate the understanding of what single neurons do with a rigorous account of how they carry out the functions they perform.

Other flavors of neuroscience, such as systems neuroscience and cognitive neuroscience have made inroads towards this goal. For example, excellent progress has been made in understanding the olfactory system of the locust. Here is a system where we understand the inputs, we understand the physiology of the neurons in between, and we have ways of analyzing the dynamics of the system that allow us to predict future behavior. And yet, the difficulty of generalizing these findings to more complex neuronal systems looms large as an obstacle to progress. Some of the best accounts of the activity of neurons in the pre-frontal cortex of monkeys still only provide a descriptive model of the data that fits the observations but does not provide a complete explanation for how the system actually carries out the function that is being modeled. This is the rule, rather than the exception in neuroscience.

One of the key difficulties is that processing information does not happen in single transistors by themselves, nor does it happen in single neurons by themselves. Both systems require the coordinated spatiotemporal organization of an complex system. Engineers over the past 60 years have constructed patterns that help organize transistors into useful components that process information. The most basic functions are those of basic logic, AND, OR, and NOT. Using these tools, arithmetic can be carried out to add, subtract, divide, and multiply numbers encoded in ones and zeros. From there, computer programs can be constructed in a straightforward manner and provide the foundation upon which more complex computer programs can be constructed. We have no equivalent explanation for the functions that assemblies of neurons carry out. We know that neurons excite or inhibit one another, and that the influence between two neurons can change. But neuroscience does not yet have the ability to recombine biologically faithful model neurons into novel circuits to perform novel functions. This indicates that the field lacks principles, or at the very least a sufficient set of well-understood patterns, which explain how neurons are organized together to enable an animal to behave in an appropriate manner in its environment.

In summary, while understanding neuronal dynamics is necessary to understanding the brain, it is not sufficient. I would posit that we must understand how the brain processes information in order to understand it as a whole. A prerequisite to understanding how the brain processes information is to describe principles of neural information processing, which a) explain how neurons perform functions collectively, b) help us to explain the functions of those parts of the brain where they are still unknown, and c) are rigorous enough to enable the design of circuits of neurons (model neurons, or eventually real physical neurons) that perform known and novel functions–true neural engineering.

To bring you up to speed, I need to remind you what is going on in the world of experimental neuroscience.

Experimentalists are now able to record the single-cell activities of a whole population of neurons simultaneously. From Briggman, Abarbanel, Kristan (2006):

By using multi-electrode arrays or optical imaging, investigators can now record from many individual neurons in various parts of nervous systems simultaneously while an animal performs sensory, motor or cognitive tasks. Given the large multidimensional datasets that are now routinely generated, it is often not obvious how to find meaningful results within the data.

This paper goes on to provide a nice overview on mathematical methods that researchers are using to grapple with the challenge of understanding the dynamics of the neural systems they are recording from. They make the case that conceptual progress needs to be made on the interpretation of the data these results yield. How can we understand what computations these neurons are collectively performing?

(Incidentally, this topic is being explored in a conference happening this week at the Los Alamos National Laboratory, which, according to one of the conference session chairs, is intended to help shape future directions for the lab. Hopefully there will be webcasts from this conference.)

Theorists have worried about what neurons are doing in local populations for some time. Investigations have given rise to all kinds of models of such activity, notably those of Wilson & Cowan (1972), and Hopfield (1982), which viewed a network of neurons as having “attractor states”, invoking the language, for Hopfield, and the mathematics, for Wilson & Cowan, of dynamical systems theory.

In 2002, Maass, Natschlager and Markram proposed a theory of computation for localized populations of neurons that did not require that the population settle into a stable attractor in order to perform computations. The idea is that local populations of interconnected neurons of the right kinds are capable of discriminating temporal input patterns in general, and that this behavior is governed by the network dynamics of those populations. They showed that neuronal circuits can be constructed to create a generalizable computational architecture for continuous analog systems, known as liquid state machines. In addition, these circuit models are some of the first to incorporate short-term synaptic plasticity in a dynamical population model.

This idea, elaborated by Maass & Markram in a book chapter entitled Theory of the computational function of microcircuit dynamics, circa 2005, proposes that the liquid state machine architecture is capable of universal computation, just as Turing machines are.

On the heels of these postulates, two recent papers make a serious effort to combine this theoretical paradigm with experimental data. Haesler & Maass (2007), in Cerebral Cortex, synthesized data from layers of cerebral cortex to build a sophisticated dynamical model and tested to see if it had the kinds of computational properties described by the previous theoretical work (it did). Karmarkar & Buonomano (2007), in Neuron, explored the notion of “clockless computing” by networks of model neurons to understand how neural systems tell time, and used psychophysical experiments to support their theories.

A few concrete things are suggested by these works:

1. Computational models of neuronal networks that take short-term plasticity and other biological details into account can be constructed. They demonstrate relevant computational properties.
2. Testing the computational properties of such networks requires framing experiments in terms of complex analog signal processing.
3. The global dynamical properties of local populations of neurons set the general tone of the computations they perform, while the single cell dynamics shape and refine those computations.

Together, these papers may represent the beginning of a new understanding of the computations that networks of real physiological neurons are capable of. Expectations are high that results from further multi-cellular recordings and from the Blue Brain project will verify and elaborate these ideas. Stay tuned!