Modern technologies allow eye movements to be used as a tool for studying language processing during tasks such as natural reading. Saccadic eye movements during reading turn out to be highly sensitive to a number of linguistic variables. A number of computational models of eye movement control have been developed to explain how these variables affect eye movements. Although these models have focused on relatively low-level cognitive, perceptual and motor variables, there has been a concerted effort in the past few years (spurred by psycholinguists) to extend these computational models to syntactic processing.

During a modeling symposium at ECEM2007 (the 14th European Conference on Eye Movements), Dr. Ronan Reilly presented a first attempt to take syntax into account in his eye-movement control model (GLENMORE; Reilly & Radach, Cognitive Systems Research, 2006). (more…)

Topic 3: Spatial Knowledge

Animal studies have shown that the hippocampus contains special cells called "place cells".  These place cells are interesting because their activity seems to indicate not what the animal sees, but rather where the animal is in space as it runs around in a box or in a maze. (See the four cells in the image to the right.)

Further, when the animal goes to sleep, those cells tend to reactivate in the same order they did during wakefulness.  This apparent retracing of the paths during sleep has been termed "hippocampal replay".

More recently, studies in humans — who have deep microelectrodes implanted to help detect the origin of epileptic seizures — have shown place-responsive cells.  Place cells in these studies were found not only in the human hippocampus but also in nearby brain regions.

The computation which converts sequences of visual and other cues into a sense of "place" is a very interesting one that has not yet been fully explained.  However, there do exist neural network models of the hippocampus that, when presented with sequences, exhibit place-cell like activity in some neurons.

The notion of place cell might also extend beyond physical space.  It has been speculated that computations occur to convert sequences events and situations into a distinct sense of "now".  And indeed, damage to the hippocampus has been found not only to impair spatial memory but also "episodic" memory, the psychological term for memory for distinct events.

Experiments? 

How can we understand the ways in which we understand space? Understanding spatial knowledge seems more tangible than understanding the previous two topics in this series. It seems that researchers are already using some of the most effective methods to tackle the problem.

First, the use of microelectrodes throughout the brain while human participants play virtual taxi games and perform problem solving tasks promises insight into this question.  Second, computational modeling of regions (e.g., the hippocampus) containing place cells should help us understand their properties and how they emerge.  Finally, continued animal research and possibly manipulation of place cells in animals to influence decision making (e.g., in a T-maze task) may provide an understanding of how spatial knowledge is used on-line. 

-PL 

Topic 2: Conflict and Cooperation

Generally, cognitive neuroscience aims to explain how mental processes such as believing, knowing, and inferring arise in the brain and affect behavior.  Two behaviors that have important effects on the survival of humans are cooperation and conflict. 

According to the NSF committee convened last year, conflict and cooperation is an important focus area for future cognitive neuroscience work.  Although research in this area has typically been the domain of psychologists, it seems that the time is ripe to apply findings from neuroscience to ground psychological theories in the underlying biology.

Neuroscience has produced a large amount of information about the brain regions that are relevant to social interactions.  For example, the amygdala has been shown to be involved in strong emotional responses.  The "mirror" neuron system in the frontal lobe allows us to put ourselves in someone else's shoes by allowing us to understand their actions as though they were our own.  Finally, the superior temporal gyrus and orbitofrontal cortex, normally involved in language and reward respectively, have also been shown to be involved in social behaviors.

Experiments?

The committee has left it up to us to come up with a way to study these phenomena! How can we study conflict and cooperation from cognitive neuroscience perspective?

At least two general approaches come to mind. The first is fMRI studies in which social interactions are simulated (or carried out remotely) over a computer link to the experiment participant.  A range of studies of this sort have recently begun to appear investigating trust and decision-making in social contexts.

The second general approach that comes to mind is that of  using neurocomputational simulations of simple acting organisms with common or differing goals.  Over the past few years, researchers have been carrying out studies with multiple interacting "agents" that "learn" through the method of Reinforcement Learning. 

Reinforcement Learning is an artificial intelligence algorithm which allows "agents" to develop behaviors through trial-and-error in an attempt to meet some goal which provides reward in the form of positive numbers.  Each agent is defined as a small program with state (e.g., location, sensory input) and a memory or "value function" which can keep track  of how much numerical reward it expects to obtain by choosing a possible action.

Although normally thought to be of interest only to computer scientists, Reinforcement Learning has recently attracted the attention of cognitive neuroscientists because of emerging evidence that something like it might be used in the brain.

By providing these agents with a goal that can only be achieved through some measure of coorperation or under some pressure, issues of conflict and coorperation can by studied in a perfectly controlled computer simulation environment.

-PL 

25) The dopamine system implements a reward prediction error algorithm (Schultz - 1996, Sutton - 1988)

It used to be that the main thing anyone "knew" about the dopamine system was that it is important for motor control.  Parkinson's disease, which visibly manifests itself as motor tremors, is caused by disruption of the dopamine system (specifically, the substantia nigra), so this was an understandable conclusion.

When Wolfram Schultz began recording from dopamine neurons in mice and monkeys he was having trouble finding correlations with his motor task. Was he doing something wrong? Was he recording from the right cells?

Instead of towing the line of dopamine = motor control he set out to find out what this system really does. It turns out that it is related to reward.

Schultz observed dopamine cell bursting at the onset of unexpected reward. He also observed that this bursting shifts to a cue (e.g., a bell sound) indicating a reward is forthcoming. When the reward cue occurs but no reward follows he saw that the dopamine cells go silent (below resting firing rate).

This pattern is quite interesting computationally. The dopamine signal mimics the error signal in a form of reinforcement learning called temporal difference learning.

This form of learning was originally developed by Sutton. It is a powerful algorithm for learning to predict reward and learn from errors in attaining reward.

Temporal difference learning basically propagates reward prediction back in time as far as possible, thus facilitating the process of attaining reward in the future.

Figure: (Top) No conditioned stimulus cue is given, so the reward is unexpected and there is a big dopamine burst. (Middle) The animal learns to predict the reward based on the cue and the dopamine burst moves to the cue. (Bottom) The reward is predicted, but since no reward occurs there is a depression in dopamine release.
Source: Figure 2 of Schultz, 1999. (News in Physiological Sciences, Vol. 14, No. 6, 249-255, December 1999)

Implication: The mind, largely governed by reward-seeking behavior on a continuum between controlled and automatic processing, is implemented in an electro-chemical organ with distributed and modular function consisting of excitatory and inhibitory neurons communicating via ion-induced action potentials over convergent and divergent synaptic connections altered by timing-dependent correlated activity often driven by expectation errors. The cortex, a part of that organ organized via local competition and composed of functional column units whose spatial dedication determines representational resolution, is composed of many specialized regions forming specialized networks involved in perception (e.g., touch: parietal, vision: occipital), action (e.g., frontal), and memory (e.g., short-term: prefrontal, long-term: temporal), which depend on inter-regional connectivity for functional integration, population vector summation for representational specificity, dopamine signals for reinforcement learning, and recurrent connectivity for sequential learning.

[This post is part of a series chronicling history's top brain computation insights (see the first of the series for a detailed description). See the history category archive to see all of the entries thus far.]

-MC

22) Recurrent connectivity in neural networks can elicit learning and reproduction of temporal sequences (Jordan - 1986, Elman - 1990, Schneider - 1991)

Powerful learning algorithms such as Hebbian learning, self-organizing maps, and backpropagation of error illustrated how categorization and stimulus-response mapping might be learned in the brain. However, it remained unclear how sequences and timing discrimination might be learned.

In 1986 Michael Jordan (the computer scientist, not the basketball player) developed a network of neuron-like units that fed back upon itself. Jeff Elman expanded on this, showing how these recurrent networks can learn to recognize sequences of ordered stimuli.

Elman applied his recurrent networks to the problem of language perception. He concluded that language relies heavily on recurrent connectivity in cortex; an unproven but well-accepted statement among many scientists today.

The year after Elman's demonstration of sequence learning with language, Walter Schneider (Schneider & Oliver, 1991) used a recurrent network to implement what he termed a 'goal processor'. This network can learn arbitrary task sequences, effectively expanding recurrent networks beyond language learning to learning new tasks of any type.

See this article for a review of a model implementing a goal processor.

The goal processor has been likened to a part of neocortex (dorsolateral prefrontal cortex) shown to be involved in maintaining goal information in working memory. Also, this maintenance is believed to occur via local (and/or via long-range fronto-parietal connections) recurrent connectivity.

Implication: The mind, largely governed by reward-seeking behavior on a continuum between controlled and automatic processing, is implemented in an electro-chemical organ with distributed and modular function consisting of excitatory and inhibitory neurons communicating via ion-induced action potentials over convergent and divergent synaptic connections altered by timing-dependent correlated activity often driven by expectation errors. The cortex, a part of that organ organized via local competition and composed of functional column units whose spatial dedication determines representational resolution, is composed of many specialized regions involved in perception (e.g., touch: parietal, vision: occipital), action (e.g., frontal), and memory (e.g., short-term: prefrontal, long-term: temporal), which depend on inter-regional connectivity for functional integration and recurrent connectivity for sequential learning.

[This post is part of a series chronicling history's top brain computation insights (see the first of the series for a detailed description). See the history category archive to see all of the entries thus far.]

-MC

21) Parallel and distributed processing across many neuron-like units can lead to complex behaviors (Rumelhart & McClelland - 1986, O'Reilly - 1996)

Pitts & McCulloch provided amazing insight into how brain computations take place. However, their two-layer perceptrons were limited. For instance, they could not implement the logic gate XOR (i.e., 'one but not both'). An extra layer was added to solve this problem, but it became clear that the Pitts & McCulloch perceptrons could not learn anything requiring more than two layers.

Rumelhart solved this problem with two insights.

First, he implemented a non-linear sigmoid function (approximating a neuronal threshold), which turned out to be essential for the next insight.

Second, he developed an algorithm called 'backpropagation of error', which allows the output layer to propagate its error back across all the layers such that the error can be corrected in a distributed fashion. See P.L.'s previous post on the topic for further details.

Rumelhart & McClelland used this new learning algorithm to explore how cognition might be implemented in a parallel and distributed fashion in neuron-like units. Many of their insights are documented in the two-volume PDP series.

Unfortunately, the backpropagation of error algorithm is not very biologically plausible.  Signals have never been shown to flow backward across synapses in the manner necessary for this algorithm to be implemented in actual neural tissue.

However, O'Reilly (whose thesis advisor was McClelland) expanded on Hinton & McClelland (1988) to implement a biologically plausible version of backpropagation of error. This is called the generalized recirculation algorithm, and is based on the contrastive-Hebbian learning algorithm.

O'Reilly and McClelland view the backpropagating error signal as the difference between the expected outcome and the perceived outcome. Under this interpretation these algorithms are quite general, applying to perception as well as action.

The backprop and generalized recirculation algorithms are described in a clear and detailed manner in  Computational Explorations in Cognitive Neuroscience by O'Reilly & Munakata. These algorithms can be explored by downloading the simulations accompanying the book (available for free).

Implication: The mind, largely governed by reward-seeking behavior on a continuum between controlled and automatic processing, is implemented in an electro-chemical organ with distributed and modular function consisting of excitatory and inhibitory neurons communicating via ion-induced action potentials over convergent and divergent synaptic connections altered by timing-dependent correlated activity often driven by expectation errors. The cortex, a part of that organ organized via local competition and composed of functional column units whose spatial dedication determines representational resolution, is composed of many specialized regions involved in perception (e.g., touch: parietal, vision: occipital), action (e.g., frontal), and memory (e.g., short-term: prefrontal, long-term: temporal), which depend on inter-regional communication for functional integration.

[This post is part of a series chronicling history's top brain computation insights (see the first of the series for a detailed description). See the history category archive to see all of the entries thus far.]
-MC

19) Neural networks can self-organize via competition (Grossberg - 1978, Kohonen - 1981)

Hubel and Wiesel's work with the development  of cortical columns (see previous post) hinted at it, but it wasn't until Grossberg and Kohonen built computational architectures explicitly exploring competition that its importance was made clear.

Grossberg was the first to illustrate the possibility of self-organization via competition. Several years later Kohonen created what is now termed a Kohonen network, or self-organizing map (SOM). This kind of network is composed of layers of neuron-like units connected with local excitation and, just outside that excitation, local inhibition. The above figure illustrates this 'Mexican hat' function in three dimensions, while the figure below represents it in two dimensions along with its inputs.

These networks, which implement Hebbian learning, will spontaneously organize into topographic maps.

For instance, line orientations that are similar to each other will tend to be represented by nearby neural units, while less similar line orientations will tend to be represented by more distant neural units. This occurs even when the map starts out with random synaptic weights. Also, this spontaneous organization will occur for even very complex stimuli (e.g., faces) as long as there are spatio-temporal regularities in the inputs.

Another interesting feature of Kohonen networks is that the more frequent input patterns are represented by larger areas in the map. This is consistent with findings in cortex, where more frequently used representations have larger cortical areas dedicated to them.

There are several computational advantages to having local competition between similar stimuli, which SOMs can provide.

One such advantage is that local competition can increase specificity of the representation by ruling out close alternatives via lateral inhibition. Using this computational trick, the retina can discern visual details better at the edges of objects (due to contrast enhancement).

Another computational advantage is enhancement of what's behaviorally important relative to what isn't. This works on a short time-scale with attention (what's not important is inhibited), and on a longer time-scale with increases in representational space in the map with repeated use, which increases representational resolution (e.g., the hand representation in the somatosensory homonculus).

You can explore SOMs using Topographica, a computational modeling environment for simulating topographic maps in cortex. Of special interest here is the SOM tutorial available at topographica.org.

Implication: The mind, largely governed by reward-seeking behavior on a continuum between controlled and automatic processing, is implemented in an electro-chemical organ with distributed and modular function consisting of excitatory and inhibitory neurons communicating via ion-induced action potentials over convergent and divergent synaptic connections strengthened by correlated activity. The cortex, a part of that organ organized via local competition and composed of functional column units whose spatial dedication determines representational resolution, is composed of many specialized regions involved in perception (e.g., touch: parietal, vision: occipital), action (e.g., frontal), and memory (e.g.,short-term: prefrontal, long-term: temporal),which depend on inter-regional communication for functional integration.

[This post is part of a series chronicling history's top brain computation insights (see the first of the series for a detailed description). See the history category archive to see all of the entries thus far.]

-MC

11) Action potentials, the electrical events underlying brain communication, are governed by ion concentrations and voltage differences mediated by ion channels (Hodgkin & Huxley - 1952)

Hodgkin & Huxley developed the voltage clamp, which allows ion concentrations in a neuron to be measured with the voltage constant. Using this device, they demonstrated changes in ion permeability at different voltages. Their mathematical model of neuron function, based on the squid giant axon, postulated the existence of ion channels governing the action potential (the basic electrical signal of neurons). Their model has been verified, and is amazingly consistent across brain areas and species.

You can explore the Hodgkin & Huxley model by downloading Dave Touretsky's HHSim, a computational model implementing the Hodgkin & Huxley equations.

Implication: The mind, largely governed by reward-seeking behavior, is implemented in an electro-chemical organ with distributed and modular function consisting of excitatory and inhibitory neurons communicating via ion-induced action potentials over convergent and divergent synaptic connections strengthened by correlated activity.

[This post is part of a series chronicling history's top brain computation insights (see the first of the series for a detailed description)]

-MC

10) The Hebbian learning rule: 'Neurons that fire together wire together' [plus corollaries] (Hebb - 1949)

D. O. Hebb's most famous idea, that neurons with correlated activity increase their synaptic connection strength, was based on the more general concept of association of correlated ideas by philosopher David Hume (1739) and others. Hebb expanded on this by postulating the 'cell assembly', in which networks of neurons representing features associate to form distributed chains of percepts, actions, and/or concepts.

Hebb, who was a student of Lashley (see previous post), followed in the tradition of distributed processing (discounting localizationist views).

The above figure illustrates Hebb's most original hypothesis (which is yet to be proven): The reverbatory cell assembly formed via correlated activity. Hebb theorized that increasing connection strength due to correlated activity would cause chains of association to form, some of which could maintain subsequent activation for some period of time as a form of short term memory (due to autoassociation).

Implication: The mind, largely governed by reward-seeking behavior, is implemented in an electro-chemical organ with distributed and modular function consisting of excitatory and inhibitory neurons communicating via convergent and divergent synaptic connections strengthened by correlated activity.

[This post is part of a series chronicling history's top brain computation insights (see the first of the series for a detailed description)]

-MC

9) Convergence and divergence between layers of neural units can perform abstract computations (Pitts & McCulloch - 1947)

Pitts & McCulloch created the first artificial neurons and artificial neural network. In 1943 they showed that computational processing could be performed by a series of convergent and divergent connections among neuron-like units. In 1947 they demonstrated that such computations could lead to visual constancy, in which a network could recognize visual inputs despite changes in orientation or size. This computation is relevant for many topics besides vision.

More profound than the visual constancy network was the proof of concept it represented. As illustrated in the above figure, multi-layered perceptrons (as networks of converging and diverging neuron-like units came to be known) can compute logical functions such as AND, OR, and XOR.

This insight provided the first clear glimpse of how actual computation might be carried out in the brain via the many convergent and divergent connections already found in its anatomical projections.

Implication:  The mind, largely governed by reward-seeking behavior, is implemented in an electro-chemical organ with distributed and modular function consisting of excitatory and inhibitory neurons communicating via convergent and divergent synaptic connections.

[This post is part of a series chronicling history's top brain computation insights (see the first of the series for a detailed description)]

-MC

Backpropogation of Error (or "backprop") is the most commonly-used neural network training algorithm.  Although fundamentally different from the less common Hebbian-like mechanism mentioned in my last post , it similarly specifies how the weights between the units in a network should be changed in response to various patterns of activity.   Since backprop is so popular in neural network modeling work, we thought it would be important to bring it up and discuss its relevance to cognitive  and computational neuroscience.  In this entry, I provide an overview of backprop and discuss what I think is the central problem with the algorithm from the perspective of  neuroscience.

Backprop is most easily understood as an algorithm which allows a network to learn to map an input pattern to an output pattern.  The two patterns are represented on two different "layers" of units, and there are connection weights that allow activity to spread from the "input layer" to the "output layer."  This activity may spread through intervening units, in which case the network can be said to have a multi-layered architecture.  The intervening layers are typically called "hidden layers".

(more…)

Most neuroscientists don't use human subjects, and many tend to forget this important point: 
All neuroscience with non-human subjects is theoretical.

If the brain of a mouse is understood in exquisite detail, it is only relevant (outside veterinary medicine) in so far as it is relevant to human brains.

Similarly, if a computational model can illustrate an algorithm for storing knowledge in distributed units, it is only as relevant as it is similar to how humans store knowledge.

It follows from this point that there is a certain amount of uncertainty involved in any non-human research. An experiment can be brilliantly executed, but does it apply to humans?

Circumventing this uncertainty problem by looking directly at humans, another issue arises:  Only non-invasive techniques can be used with humans, and those techniques tend to involve the most uncertainty.

For instance, fMRI is a non-invasive technique that can be used to measure brain processes in humans. However, it measures the oxygenation levels, which is only indirectly related to neural activity. Thus, unlike with animal models, measures of neuronal activity are surrounded by an extra layer of uncertainty in humans.

So, if you're a neuroscientist you have to "choose your poison": Either deal with the uncertainty of relevance to humans, or deal with the uncertainty of the processes underlying the measurable signals in humans.
(more…)

Most neurocomputational models are not hard-wired to perform a task. Instead, they are typically equipped with some kind of learning process.  In this post, I'll introduce some notions of how neural networks can learn.  Understanding learning processes is important for cognitive neuroscience because they may underly the development of cognitive ability.

Let's begin with a theoretical question that is of general interest to cognition: how can a neural system learn sequences, such as the actions required to reach a goal? 

Consider a neuromodeler who hypothesizes that a particular kind of neural network can learn sequences. He might start his modeling study by "training" the network on a sequence. To do this, he stimulates (activates) some of its neurons in a particular order, representing objects on the way to the goal. 

After the network has been trained through multiple exposures to the sequence, the modeler can then test his hypothesis by stimulating only the neurons from the beginning of the sequence and observing whether the neurons in the rest sequence activate in order to finish the sequence.

Successful learning in any neural network is dependent on how the connections between the neurons are allowed to change in response to activity. The manner of change is what the majority of researchers call "a learning rule".  However, we will call it a "synaptic modification rule" because although the network learned the sequence, it is not clear that the *connections* between the neurons in the network "learned" anything in particular.

The particular synaptic modification rule selected is an important ingredient in neuromodeling because it may constrain the kinds of information the neural network can learn.

There are many categories of mathematical synaptic modification rule which are used to describe how synaptic strengths should be changed in a neural network.  Some of these categories include: backpropgration of error, correlative Hebbian, and temporally-asymmetric Hebbian.

• Backpropogation of error states that connection strengths should change throughout the entire network in order to minimize the difference between the actual activity and the "desired" activity at the "output" layer of the network.
• Correlative Hebbian states that any two interconnected neurons that are active at the same time should strengthen their connections, so that if one of the neurons is activated again in the future the other is more likely to become activated too.
• Temporally-asymmetric Hebbian is described in more detail in the example below, but essentially emphasizes the importants of causality: if a neuron realiably fires before another, its connection to the other neuron should be strengthened. Otherwise, it should be weakened. 

Why are there so many different rules?  Some synaptic modification rules are selected because they are mathematically convenient.  Others are selected because they are close to currently known biological reality.  Most of the informative neuromodeling is somewhere in between.

An Example

Let's look at a example of a learning rule used in a neural network model that I have worked with: imagine you have a network of interconnected neurons that can either be active or inactive.    If a neuron is active, its value is 1, otherwise its value is 0. (The use of 1 and 0 to represent simulated neuronal activity is only one of the many ways to do so; this approach goes by the name "McCulloch-Pitts").

(more…)

In my most recent post I gave an overview of the "simple recurrent network" (SRN), but I'd like to take a step back and talk about neuromodeling in general.  In particular I'd like to talk about why neuromodeling is going to be instrumental in bringing about the cognitive revolution in neuroscience.

A principal goal of cognitive neuroscience should be to explain how cognitive phenomena arise from the underlying neural systems.  How do the neurons and their connections result in interesting patterns of thought?  Or to take a step up, how might columns, or nuclei, interact to result in problem solving skills, thought or consciousness?

If a cognitive neuroscientist believes they know how a neural system gives rise to a behavior, they should be able to construct a model to demonstrate how this is the case.

That, in brief, is the answer.

But what makes a good model?  I'll partially answer this question below, but in future posts I'll bring up specific examples of models, some good, some poor.

First, any "model" is a simplification of the reality.  If the model is too simple, it won't be interesting.  If it's too realistic, it will be too complex to understand.  Thus, a good model is at that sweet spot where it's as simple as possible but no simpler.
Second, a model whose ingredients spell out the result you're looking for won't be interesting.  Instead, the results should emerge from the combination of the model's resources, constraints and experience.

Third, a model with too many "free" parameters is less likely to be interesting.  So an important requirement is that the "constraints" should be realistic, mimicking the constraints of the real system that is being modeled.

A common question I have gotten is:  "Isn't a model just a way to fit inputs to outputs?  Couldn't it just be replaced with a curve fitter or a regression?"  Well, perhaps the answer should be yes IF you consider a human being to just be a curve fitting device. A human obtains inputs and generates outputs.  So if you wish to say that a model is just a curve fitter, I will say that a human is, too.

What's interesting about neural systems, whether real or simulated, is the emergence of complex function from seemingly "simple" parts.

In future posts, I'll talk more about "constraints" by giving concrete examples.  In the meantime, feel free to bring up any questions you have about the computational modeling of cognition.
-PL 

[Image by Santiago Ramon y Cajal, 1914.] 

…from a knee-jerk reaction to its immediate input? 

Although one of the first things that a Neuroscience student learns about is "reflex reactions" such as the patellar reflex (also known as the knee-jerk reflex), the cognitive neuroscientist is interested in the kind of processing that might occur between inputs and outputs in mappings that are not so direct as the knee-jerk reaction. 

An example of a system which is a step up from the knee-jerk reflex is in the reflexes of the sea slug named "Aplysia".  Unlike the patellar reflex, Aplysia's gill and siphon retraction reflexes seem to "habituate" over time — the original input-output mappings are overridden by being repeatedly stimulated.  This is a simple form of memory, but no real "processing" can be said to go on there.

Specifically, cognitive neuroscientists are interested in mappings where "processing" seems to occur before the output decision is made.  As MC pointed out earlier, the opportunity for memory (past experience) to affect those mappings is probably important for "free will". 

But how can past experience affect future mappings in interesting ways? One answer to this question appeared in the year 1990, which began a new era in experimentation with neural network models capable of indirect input-output mappings.  In that year, Elman (inpired by Jordan's 1986 work) demonstrated the Simple Recurrent Network in his paper "Finding Structure in Time".  The concept behind this network is shown in the picture associated with this entry.

The basic idea of the Simple Recurrent Network is that as information comes in (through the input units), an on-line memory of that information is preserved and recirculated (through the "context" units).  Together, the input and context units both influence the hidden layer which can trigger responses in the output layer.  This means that the immediate output of the network is dependent not only on the current input, but also on the inputs that came before it.

The most interesting aspect of the Simple Recurrent Network, however, is that the connections among all the individual units in the network change depending on what the modeler requires the network to output.   The network learns to preserve information in the context layer loops so that it can correctly produce the desired output. For example, if the task of the network is to remember the second word in a sentence, it will amplify or maintain the second word when it comes in, while ignoring the intervening words, so that at the end of the sentence it outputs the target word.

Although this network cannot be said to have "free" will — especially because of the way its connections are forcefully trained — its operation can hint at the type of phenomena researchers should seek in trying to understand cognition in neural systems.

-PL