If you want to understand how a commercial on TV or on paper, a brand, a package, or a product are perceived by a consumer, you have to look inside their brain. That’s at least what Nielsen, a giant  providing information and measurement for a comprehensive understanding of consumers and consumer behavior, believes. Nielsen made $5 billion of revenues out of this business last year.

Now Nielsen has decided to strongly invest on neuromarketing, a new methodology based on looking inside the consumers’ brain using brain imaging techniques. That’s why Neurofocus, a neuromarketing leading company, was acquired by Nielsen a few days ago.

The New York Times, while talking about Neurofocus and Nielsen, explains that:

Neuromarketing is the application of neuroscientific research to marketing, advertising, and entertainment content and messages. Neuroscience research has in recent decades revealed important new discoveries about how the human brain is structured and how it functions. These findings enabled NeuroFocus to develop patented technologies and proprietary techniques that provide greater accuracy and insight into consumer research.

The NeuroFocus mind wireless EEG measurement headset

But how is Neurofocus investigating our brain reactions? Let’s take the case of studying a TV commercial. First of all you have to wear an EEG cap on your head, in order to measure the brain activity while being exposed to the media stimuli (watching the commercial…). Afterward, some keywords that have been carefully associated to the commercial under investigation, are shown onto the screen: those words can refer to the brand, to the message, to the emotions that specific experience is evocating.

Let’s say that we want to understand if  people are perceiving “speedness”, “aggressiveness”, “safety”, “design” in a sportscar commercial. While displaying those words, the EEG measures the brain reaction and creates a reference basis for the test. The commercial itself is then projected, followed again by the same keywords.

In the second round of the keyword projection, the brain reacts differently to each word because of the experience just elicited by the commercial. This reaction is measured with a technique called P300, that in neuroscience is a well-known electrical signal change in the neural activity that follows a relevant visual experience.

Thus the words that for that person were relevant while watching the commercial are identified, without having to ask anything or submitting a traditional questionnaire.

My bet is that in the future those systems will be used to support us also for daily tasks, for instance to instantly find images out of large personal database just following our brain inspiration (which in fact Microsoft Research is somehow already investigating), or to navigate through a movies catalog just being guided by our emotions.

Future networks will be able to share a same physical infrastructure: this will guarantee multiple networks simultaneously running without interfering with one another’s operation. Actually, allowing multiple networks to share the same physical infrastructure requires virtualizing and managing network resources (i.e., processing, storage, routing/forwarding nodes and links).

So let’s suppose we wish designing a virtual network capable of self-management operations: one can imagine a set of atomic functions to be performed on links (e.g. Label Switched Path – LSP) and virtual resources (slices of processing, storage and routing/forwarding). For example, some typical operations on a link may include: set-up/tear-down, modification of bandwidth, forwarding, monitoring, etc. In other words, we wish a virtual network able to function and learn over time without (or with a limited) supervision by either a human or some other external centralised system. As in humans, emotions seem to play the role of an evaluation instrument for learning, we wish including “emotions” in our network architectures to serve the self-management purposes.

Let’s start by referring to two well-known cognition theories (the Global Workspace and the Pandemonium Theory).

Global Workspace Theory

There is considerable empirical evidence from fMRI brain imaging experiments to show that individual human consciousness involves a single global workspace. Global Workspace Theory (GWT) (Baars, 1988, 1997, 2003) argues that human cognition is implemented by a multitude of relatively small, special purpose processes. Direct communication between them is rare and over a narrow bandwidth, whilst coalitions of such processes find their way into a global workspace.

Workspace serves to broadcast the message of the coalition to all the unconscious processors, in order to recruit other processors to join. All this takes place in a certain contexts which is itself a coalition of processes. According to GWT, brain can be viewed as a collection of distributed specialized networks (processors). Consciousness, associated with a global workspace in the brain, is a shared memory capacity whose contents are broadcasted to many unconscious specialized networks. A global workspace can also serve to integrate many competing and cooperating input networks.

Pandemonium Theory

Pandemonium Theory (Selfridge’s, 1958) is a connectionist architecture originally used for pattern recognition. Multiple independent processes called demons work simultaneously recognizing specific conditions (or a set of them). Demons have links that allows them to “call” other demons. J. Jackson (1987) extended the original Pandemonium theory by creating the stadium metaphor, organizing demons in two different locations, the equivalent of stands and arena of a stadium. A system consists of a crowd of usually dormant demons located at the stands, from where a few demons could go down to the arena and start exciting the crowd. Some demons in the crowd get more excited and start to yell louder. If a demon yells loudly enough, it gets to go down to the playing field and become active. Whenever a demon enters the playing field, the arena creates associations (if they do not already exist) between the incoming demon and any demons already on the field. These connections between demons are created or strengthened according to the time they are together on the arena, following a Hebbian learning scheme. Processes of the global workspace theory where correspond to demons of the Pandemonium theory. Global workspace is the arena.

Arena can perform the actual input and output functions of our network. It can also calculate a “gain”, i.e., a variable to measure how well the network is performing. Gain is up (goals achieved) if things are going well, down when doing poorly. The higher the gain, the more links are a strengthened by time together on the field. Gain determines how to strengthen or weaken associations between components: so network steers towards goals.

Lightweight components can be associated to each network entities (i.e., processing, storage, routing/forwarding virtual nodes and links). This component is in charge to build-up and tear down perceptual structures (of the associated virtual resource) and to direct the self-management operations.

If gain is modelled as a vector of four numbers each of which can be thought of as analogous to the four basic emotions, anger, sadness, happiness, and fear, then network’s emotional state at any one time is, therefore, considered to be the combination of the four emotions. Emotion components are then also introduced. When an emotion component’s preconditions are met it fires, modifying the value of a global variable representing the portion of the emotion vector associated with the component’s preconditions. The value of an individual element (emotion) in the gain vector can be modified when an emotion component fires.

Now last step: let’s imagine a function workspace (like an arena) populated with a myriad of these resource and emotion components behaving according the models of the GWT or/and the Pandemonium Theory. An interesting consequence is that the emergent network representation and self-management will be driven by a myriad of micro-behaviours rather than in traditional systems, where knowledge representation is driving management and behaviour.

Network will look like having associative memory capability based on a sparse distributed memory mechanisms. New perception associates with past experiences including actions and emotions. Remembered emotions activate emotion components that, in turn, influence actions selection.

Neuroscience is seeing increasing knowledge about the anatomy and electrophysiological properties of brain. Brain networks properties — combining high efficiency of parallel information processing transfer with low connection cost — represent the consequence of evolutionary selection.

Understanding brain functioning is one of the grand challenges of Science: interest is clearly motivated by the fact that some general principles of brain functioning seem could govern also other complex networks, including social, biological and communications networks.

Let’s consider emotions: they give us the ability to assess of situations. One of the main example is to determine whether a given state of the environment is beneficial or detrimental without (or with limited) dependence on some external evaluation. There has been a great deal of research showing that, for humans, emotion is one of the key elements for a “rational” behaviour (meant as the behaviour that avoids non-pleasurable states and/or pursues pleasurable states).  Emotions for humans have been adjusted and pre-wired over millions of years of evolution.  But, even so, many of our decisions are based on our culture and on those complex learned emotions that are not pre-wired.  Understanding how humans manage to learn these complex emotions and how these become coupled to actions is challenging and extremely important, with far reaching implications in many areas.

Let’s imagine we wish to design a network able to interact with its environment (including the Users) in a way that includes “emotional content” at a basic level; this network should be able to ability to display a full range of emotions and to learn complex responses.

They say that emotional state at any one time can be the combination of the four basic emotions, anger, sadness, happiness, and fear, but also disgust and surprise can be included.

So, what sorts of emotional reactions can be expected from our emotive network? Well the network may experience fear or anger in case of faults occurrence or security attacks. It may be annoyed at having reminded human operators to take some actions with no reply. It may be happy at having learned new information and/or knowledge or at seeing its Users happy of the Quality of Experience (many efforts have been developed to determine the emotional state of a human User via behaviour, linguistic expressions, face recognition among other things).

Emotional reactions then results in actions – as in human or animals – selected in such a way as to manipulate environment in order to get “pleasure” or avoids “displeasure”. So the network will learn from its experiences and pursue goals reinforced by the emotional valence that the results of those goals produce.

For our design, I wish referring to the Pandemonium and Global Workspace Theories.  In the following days, I’m posting another piece how adopting these theories to design a network with emotions: I hope convincing you this is not that far in the future.

Can we imagine the emergence of a global intelligence out of this network ? Will our brain be able to connect to that intelligence and, eventually, will that “distributed network” able to tap into our own brain to further increase the overall Gaia’s intelligence ? Can we use this to mitigate big issues like energy, global warming, water availability, feeding a growing population ?

If on one side we have a huge amount of empirical data in neuroscience, on the other hand there are relatively few global theories about how the brain works. How highly complex neurons networks (and the nervous system, in general) are able to accomplish certain skills? Shall we be able to extract (from ongoing and future researches) some elementary principles for a unified brain theory and to exploit them for the evolution of information infrastructures?

I’ve recently read this interesting paper proposing the free-energy principle for brain action, perception and learning:

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2660582/

Paper reviews some the brain theories, both in the biological (for example, neural Darwinism) and physical (for example, information theory and optimal control theory) field from the free-energy perspective. The free-energy principle says that any self-organizing system that is at equilibrium with its environment must minimize its free energy.

The principle is a sort of mathematical formulation of how adaptive systems resist a natural tendency to disorder. Specifically, the paper discusses how this principle could be applied to neurons networks evolution, explaining many interesting issues behind the Bayesian brain modelling hypothesis. In summary, brain would be like an inference machine that actively predicts and explains its sensations; furthermore it continuously tries to optimize probabilistic representations of what caused its sensory input. Then it is easy to show that value is inversely proportional to surprise, in the sense that the probability of a phenotype being in a particular state increases with the value of that state.

This means that free energy is an upper bound on expected cost, which makes sense as optimal control theory assumes that action minimizes expected cost, whereas the free-energy principle states that it minimizes free energy (and implicitly entropy).

So, if I’ve understood correctly, value or surprise is determined by those innate categories and generative models which can be heritable through genetic and epigenetic mechanisms.

For example young songbirds innately recognize and prefer to learn the songs of their own species; so it seems there might be neural bases in young songbirds for innate recognition of songs. In the same way, Steven Pinker in his book “The Language Instinct” attempts to convince that language is essentially innate in humans. One could argue that in the brain there might be some innate categories, featuring abstraction schemes (such as songs recognition in birds or in language skill in humans), expected to have direct correlates with the neurons circuits. 

In general, this picture leads us to hypothesize that the (autonomic) nervous system, in its environment, as a hierarchy of systems conforming to a range of abstract schemes; these abstractions relate to particular neural circuits or systems and their evolutionary behaviours in a given environment.

I think this is an interesting perspective of study with exciting applications even beyond neuroscience: think about designing free-energy information networks embedding autonomic and evolutionary biology principles.

PNAS (Proceedings of the National Academy of Sciences) published this week the paper “Network architecture of the long-distance pathways in the macaque brain” reporting interesting results (achieved by IBM) about reverse-engineering the Macaque monkey brain. Specifically, the research claims the mapping of the long-distance network of the monkey brain, structure which is essential for understanding brain’s behavior and dynamics.

Long-distance brain connections are like highways crossing the brain’s white matter, while short-distance gray matter connections (based on neurons) are like “local roads” within a brain region.  

The brain network they found shows a “tightly integrated inner core that might be at the heart of higher cognition and even consciousness” capabilities. This core spans parts of premotor cortex, prefrontal cortex, temporal lobe, parietal lobe, thalamus, basal ganglia, cingulate cortex, insula, and visual cortex. Amazingly, by ranking brain regions, they found also that the prefrontal cortex (physically located in the front of the brain) is a functionally central part of the brain that might act as an integrator and distributor of information.

The long-distance network of the Macaque monkey brain showing 6,602 long-distance connections between 383 brain regions (PNAS)

Learning how brain’s network works is indispensable for progressing in several research avenues of neurosciences, including computational neuroscience. At the same time, if we imagine future networks as a global environment of dynamically interconnected nodes (including terminals, sensors, smart things, etc.) these researches might also be valuable to understand and manage the bottom-up emergence of future computing and networking distributed networks.

Understanding the functioning of brain network is one of the grand challenges of Science: many methods have been applied to analyze and study its structure and function. Interest is clearly motivated by the fact that the some general principles of brain functioning seem to govern also other complex networks, including social, biological and communications networks.

In principle, a brain network can be modelled as graph representing activities in the brain, where the vertices represent anatomical regions and the edges their functional connectivity. Functional magnetic resonance imaging (fMRI) has been used to extract data about how networks of neurons are performing tasks. In these experiments the activity of the brain is measured, in time steps of a few seconds, in data associated to a number of “voxels” of dimension of about 27 mm³. Data are organised in correlation matrices (also with Multidimensional Scaling techniques) to see where there is a correlation among the variables. A correlation matrix can describe correlation among M variables: it is a square symmetrical MxM matrix with the (ij)th element equal to the correlation coefficient r_ij between the (i)th and the (j)th variable.

Over the last two decades, said data analysis techniques have transformed neuroscience: some scientists are even claiming they can identify the brain regions responsible for musical ability, food preference, fairness and even skills by analysing and correlating voxels data.

Also the recent explosion in gene expression experiments has generated a large amount of data to be correlated: think about The Cancer Genome Anatomy Project (CGAP), a program of the National Cancer Institute (NCI) that is researching the molecular patterns changes occurring when a normal cell is transformed into a cancer cell or the relationships between the gene expression profile of a tissue and the pattern of its neighbours.

These are just two example where the capabilities of correlating data have an incredibly high impact value. Analysing interconnectivity is so fundamental to understanding the behaviour of any complex networks.

Also Telecom Operators have huge amount of data, still to be leveraged: in this sense, methods and techniques adopted by neuroscience and genomics can provide valuable lessons learnt.

Recently, New Scientist has published the following article describing how simple memristors (devices that, like resistors, oppose the passage of current, and whose resistance, at any moment, depends on the last voltage) are being used in a US military-funded project (SyNAPSE) trying to make brain-like computers.

http://www.newscientist.com/article/mg20527515.900-electronics-missing-link-brings-neural-computing-closer.html

Neuroscience often compares brain to a computer but, apart from the trivial fact that both process information, it is not clear whether this remains just a metaphor or it is something more.

Yesterday, New Scientist has publishing another interesting paper by a research biologist of the University of Cambridge. Paper argues, apparently in opposite direction than SyNAPSE, that the brain is not a like a computer where neurons are components or transistors; rather each individual neuron is itself a small processing unit, and the brain is like vast network of said microscopic computers.

http://www.newscientist.com/article/mg20627571.100-the-secrets-of-intelligence-lie-within-a-single-cell.html?page=2

Paper brings several interesting examples how in nature even simple organisms or cells are not “subservient nano-bots”, but they respond to environment changes by performing a certain level of processing and by taking decisions. Biology reductionism seems missing the systemic view on how whole cells behave.

I’m wondering whether developing structural (brain-like) networks of nano-micro processing devices will be a way to test the potential emergence of that sort of intelligence elaborated in yesterday article of the New Scientist.

As a matter of fact, nanotechnologies are making amazing advances in developing nano-devices, nano-wires, nano-sources of energy…etc. Moreover, while microelectronics has experienced miniaturization pushing down the scale towards submicron circuitry, there are significant progresses in (self-)composition of nano-devices, also in order to scale up to bigger dimensions and bridging with microelectronics.

Maybe principles of simple nervous systems and brain evolution can be investigated by comparative analysis of such structural (brain-like) networks of nano-micro processing devices.

A team of neuroscientists and computer experts from the UK, US and Germany have discovered “striking similarities” between human brains, the nervous system of a nematode worm (Caenorhabditis elegans) and electronic chips.

http://www.admin.cam.ac.uk/news/dp/2010042201

They have found that all three share two basic properties, at least. Firstly, all have a Russian doll-like architecture, with the same patterns repeating over and over again, at different scales. Secondly all three show the Rentian scaling – a rule used to describe the relationship between the number of elements in a given area and the number of links between them.

This challenges the amazing idea that studying the nervous system of such simple organisms could offer realistic possibilities of better understanding human brain, and what it has in common with them (maybe also from an evolutionary point of view). Lesson learnt will be valuable also for the evolution of technologies (as above analogies seems showing).

I’m wondering if also the Internet, with its ever growing complexity, will share such basic properties ?

The interesting paper is published today in PLoS Computational Biology

http://www.ploscompbiol.org/article/info%3Adoi%2F10.1371%2Fjournal.pcbi.1000748

Neuroscientists argue that brain networks show small world and exponentially truncated power law characteristics. These are considered, by the way, the best structural basis for coexistence of both informational processing segregation (in specialized regions) and integration (by coherent oscillations in wider regions).

Most of neuroscience investigations, as far as I know, are currently based on the assumption that thoughts, feelings, consciousness emerge from the electrical and chemical communications between brain cells. On one side there are experimental studies, using different anatomical technologies (e.g. magnetic resonance imaging) to observe and elaborate about brain patterns. On the other side, there are studies and simulations on various computational models of complex networks neurons: for example how neurons fire electrically in response to inputs of other neurons, or release neurotransmitters to communicate with each other.

Anyway, how such network of networks of cells can produce the subtlety of mind, or can perform the body autonomic control, is still a big open question.

An interesting approach [1] proposes alternatively to focus on the relevance of abstractions (and their relationships to the natural evolution, or the human design).  Being an electronic engineer (fond of Physics) let me resume the example of the Ohm’s law, V = IR, where V, I and R denote abstract entities respectively known as voltage across a resistor, current through it and resistance. So, the movement of electrons, in a physical resistor, results in the realisation of an abstract scheme satisfying the Ohm’s law. Engineers make use of such (and many others) abstract schemes to design electronic systems performing particular functions (but, basically ignoring the laws behind the movement of electrons). The same ideas probably apply to bio-systems. Bio-systems contain several components of various kinds, conforming to some scheme of abstractions. On the other hand, they differ from machines in that the entities concerned often lack a formal definition, their properties being inferred from investigations (at different levels of observation) of several instances encountered in nature.

 
This reasoning equally applies to autonomic systems functioning.  The human autonomic nervous system, in its environment, is probably characterised by a hierarchy of systems conforming to a range of abstract schemes; typically these schemes relate to particular neural circuits or biological sub-systems and their adaptive behaviour in a given environment.

Natural design and evolution are driving (indirectly) the human autonomic nervous system to conform to such abstract schemes, through development and cognition (learning, at least). I think this is an important point.

Following this reasoning, to engineer autonomic systems requires defining, not only control-loop circuitry (at the proper level), but also hierarchies of abstract schemes and the processes of evolutionary development and cognition in dynamic environments.

In a next post, we’ll elaborate about how cognition may control phase-transitions in autonomic networks.

[1] http://www.tcm.phy.cam.ac.uk/~bdj10/papers/messina2001.html

Indeed Nature is full of patterns: imagine the colorful patterns found on tropical fish or the beautifully complex patterns on seashells. Patterns of the same type may look alike, but they never are exactly the same and noise is the reason.

Alan Turing proposed (in 1952) the reaction-diffusion principle [1] to model biological pattern formation, and morphogenesis (cellular differentiation in early biological development). Diffusion is the relatively slow process employed by nature to equalize concentrations, for example chemicals components; reaction is a faster inhibitory local effect between two components. The overall system is constantly in search of a stable (activator – inhibitor) equilibrium: if the system is unable to find such equilibrium, it will be fully dynamic (e.g. chaos and constant transitions); if the system is able to find a stable state, it can create a certain pattern. What is amazing is that Turing’s model is capable of describing many natural patterns.

Reaction-diffusion systems have some very interesting properties for possible application in future networks engineering, pervasive computing and artificial intelligence. Imagine, for example, a network of nodes each of which performing some type of computation (reaction) and being linked to other nodes by some communications protocol (diffusion). They create a kind of ecosystem whose emerging intelligence may be symbolic and/or structural. As another example, reaction-diffusion could be used to enable highly complex ecosystems of ‘minimally cognitive’ objects [2]: actually, in neuroscience, they consider plausible that reaction-diffusion processes can mimic the formation of synaptic contacts between neurons and computational meshes [3], [4].

Simulations of this reaction-diffusion phenomenon are very fascinating. I’d like offering a short movie showing a simulation that I’ve made this morning using the Breve environment.

[local /files/2009/11/rd-simulation.wmv RD simulation]

Imagine it as a cube of nano/micro-devices/nodes interacting with each other through reaction-diffusion and creating dynamic computational meshes or synaptic patterns. 

References

[1] A.M. Turing, Phil. Trans. R. Soc. Lond., B237, 37-72 (1952).

[2] http://alifexi.alife.org/papers/ALIFExi_pp142-149.pdf

[3] http://www.biomedcentral.com/content/pdf/1471-2202-9-S1-P85.pdf

[4] http://arxiv.org/abs/cond-mat/0211283v2