Study Guide with article summaries for Artificial Intelligence at Leiden University

The subject that is most researched of the human brain is vision. However, all available data shows that we still know very little about visual processes. The purpose of this paper is to represent the biological aspects of vision.

Anatomy

The anatomy of the visual system is structured as follows. There are different lobes in the brain and all parts in the brain are connected to each other in a complex way. The visual cortex occupies 40% of the brain. The visual system can be divided into two main streams. First the visual information comes from the eye to the V1 area via the thalamus. From there it can go to the occipital lobe via V1, V2 to MT (the dorsal road; the 'where path') or to the V2 / V3 / V4 regions to the IT (the ventral road; the 'what path') "). The functions of all these areas have only become clear through people with brain damage. V1 has the most important visual role.

The cortex consists of six layers with different input and output layers. The input is mainly in layer 4 and the output is mainly from layer 6. The connections between the cortical areas are probably reciprocal: they can send information to each other (so there is no one-way traffic).

Neurophysiology of visual pattern processes

Contours

The neural responses of the V1 respond primarily to orientation and spatial scales. The cells only respond to their own receptive field. They can be divided into two groups: the dependent and the independent cells. The dependent cells only respond to one kind of stimuli (for example, a line on a dark OR light background), while the independent cells respond generally (for example, to a line regardless of the background). These two groups respond to the orientation of lines. There is another group that not only responds to the orientation but also to the length of the lines, this is the hyper-complex group of cells.

Later research showed that V1 cells can also respond to the environment outside the receptive field. This only happens when the environment matches the stimulus. The V1 cells are arranged retinotopic. This means that the cells that are next to each other also see the world side by side. The receptive fields of neurons therefore partly overlap.

Groups of cells that belong together (such as those that respond to movement or color) also lie together in the cortex and are worked out in columns. The deeper layers of the cortex do not have such a clear distinction. Most neurons respond to one eye.

The work of Hubel and Wiesel showed that neurons from the V1 respond to lines and edges of visual stimuli: contours. This showed that the detection of edges of an object requires integration of information on different spatial scales. This evidence has led to the assumption that the analysis of spatial frequencies is a visual process. In conclusion, the V1 ensures that the contours of an object are determined.

V1 is the most common layer of the visual cortex, then it becomes more detailed. The V2 and V3 layers are more complex and have their own receptive fields. These are somewhat larger than the receptive fields of the V1 neurons. The V2 and V3 neurons are also arranged retinotopically. The V2 layer is sensitive to collinear and continuous contours (these are more general contours). In addition, the V2 layer plays a role in direction and movement. In conclusion, it can be said that contour is determined by neurons from the V1, V2 and V3 layers.

Item attributes

The V4 area responds to color and shape and facilitates the recognition process. Most neurons within the V4 respond to their own color. This area is also arranged retinotopically, but not as sharply arranged as the areas of V1, V2 and V3. Just as with the V2 region, the neurons that have a lot to do with each other are close together.

The last cortical areas of the ventral path are in the IT. The IT responds to both color and shape and has a large receptive field. In V4 and the PIT (posterior inferotemporal), 70% of neurons are sensitive to specific objects. The combinations of neurons indicate for which object, this can be a person. The neurons in the CIT (central inferotemporal) and AIT (anterior inferotemporal) have the most complex sensitivity to objects. All these areas together (V1, V2, V3, V4, IT, PIT, CIT and AIT) provide the ventral path and determine what you see. The further you go into the process, the more sensitive and specific the neurons are.

Grandmother cells

Neurons have been found in the AIT and STPa that respond very sensitively to biologically important visual patterns. They respond strongly to faces, hands and limbs and provide complex and abstract information. It has been proven that they respond specifically by having them look at heads in different ways. Multiple areas are used in the forming process. If an area is defect, it can still be taken care of by other areas (with the exception of the V1 area).

Learning associations

Research showed that after training some neurons were more sensitive to incoming learned information. The reaction speed became faster and more active. These studies show that the brain can learn and is plastic, but also that one can learn associative at the cell level. From the above information, it can be said that four phases are carried out:

1. collecting contour information;

2. creating a form dictionary (from which the correct form is recognized);

3. retrieving all possible descriptions;

4. recognizing the object.

Rolls of feedback connections

Within psychological models of recognition, top-down influence plays a major role in the visual process. Now we mainly look at expectations. Stimuli outside the receptive field are often completed with existing knowledge and people expect what they will see in advance. Various reactions have also been demonstrated in the V3 area if the expectation was right or wrong.

Movement plays a major role in this. There is almost no reaction to the moment that someone puts a hand in the receptive field, compared to the moment when something else suddenly comes in. Then there are strong reactions. However, often the cells begin to respond before the object is actually in the receptive field. That is why reactions are often very fast. The expectation is not only linked to seeing your own hand, but also to the expectation of where the hand should come into view and what it should look like. So there is no response if you know what is coming and if that is correct.

However, expectations do not only apply to the visual system. The moment you feel something different than you expect, you also respond strongly to it. All this is linked to the memory.

The entire visual system uses a feed forward process. The information is being processed further and further and does not return. So there are many more feedback connections from V1 to the LGN than the other way around. The feedback probably creates the "I told you so" idea. This can mainly be tested with prototype objects. Why prototypes are recognized faster is not yet known. This probably has something to do with the columns in the IT. Activity in a whole column then stands for features of the prototype.

This article provides a graphical perspective on the arithmetic challenges of object recognition. It also looks at which neuronal population is responsible for the representation of objects. Our daily activities are accompanied by quick and accurate recognition of visual stimuli; this way we can recognize thousands of objects within seconds. However, which brain mechanisms are involved in this process is still unknown.

Object recognition is defined as being able to accurately distinguish objects or categories from all kinds of possible stimuli. This is done via the retina through an identification-prescribing transformation. Object recognition is difficult for a variety of reasons. The main reason is that every object can produce an infinite number of different images on the retina, while this image is still recognized from every angle. This is also called the invariance problem: the fact that we never see exactly the same thing twice and can nevertheless recognize things.

Arithmetic processes

When solving a recognition task, someone must use internal neuronal representations. These internal neural representations come from the visual vision and from this a choice is made by the brain. The brain must use a selection function to distinguish between which neurons are fired when object A is presented and which is not. Somewhere in the brain there are the right neurons that react to something and you have to filter it out. The central question in this process remains: which format of neurons represents the choice and which decision functions belong to that representation?

You can see the above problem from two sides. On the one hand, object recognition is a problem for finding complex decision functions and on the other it is a problem for finding operations that progressively transform from the retinal representation in the form of a new representation, followed by decision functions. The latter vision can be well used when investigating the architecture of the visual system (especially the ventral path).

Object recognition is difficult

Our eyes fixate an average of 300 ms on the world and then move on. During every glimpse, a visual image is already made and stored using at least 100 million cells. Such a representation can be seen as high-dimensional. An example of a low dimensional representation is a face.

Then it is about a fixed object that can be seen in many different ways. How you can see an object is called a manifold. Different objects have different manifolds.

The manifolds of all different objects are crumpled together in the brain. This means that the retina does not immediately recognize what we see, but it does pass on the information that we need to make a choice of what we see.

You can see the brain recognition mechanism as a transformation of the incoming visual representation that is easy to build into recognition. However, it is not possible to decode how recognition works.

The ventral visual path

This path translates the manifolds into objects. The ventral path is as follows: V1 to V2 to V4 to IT. Gross's studies have shown that the IT has the most specific complex neurons. The neurons there are likely to cause object recognition, because these neurons respond specifically to certain forms and are reasonably insensitive to changes in object position.

Recognition is not a result of performance, but of how strong the visual representation is in the IT cortex. This also means that the manifolds are less mixed up in the IT cortex. This also means that the V1 cortex is still very mixed up when it comes to the manifolds (just like in the retinal representation). In short: the ventral path ensures that objects are recognized by unraveling the manifolds. It is not yet known how this happens.

Meanwhile, there are already several ideas about and investigations into the process of the ventral path. Some neurophysiologists have focused on characterizing tolerance in IT neurons towards some objects. This is the same as object tangling. Other research is aimed at understanding the characteristics of shape dimensions. These studies are important for defining complex characteristics of the ventral visual pathway for neuronal tuning, which is related to manifold untangling.

The perspective of the tangling object leads to a different approach. Individual IT neurons are not expected to be responsible for the recognition, but for population representations. In addition, this perspective assumes that the immediate preparation of a goal determines how well the ventral visual path causes detangling of manifolds. This perspective offers a better way to make computer models, because populations can be more meaningful than individual neurons.

In addition, the perspective states that a focus on the cause of the tangling is better than focusing on the characteristics or forms where something reacts. Finally, with this perspective, hypotheses can be tested, which can lead to new biological hypotheses.

Flattened manifolds

By flattening manifolds one can perhaps see what happens. We are looking for transformations that cause a manifold to be flattened without interfering with others. This allows the correct neurons to be identified. At IT-level, the detangling of object manifolds results in the folding of each manifold into one point. This suggests that detangled IT-representations not only directly provide object recognition, but also recognize other tasks such as position, location and size. IT-neurons therefore have large limited receptive fields. Here the limitation works to the advantage. The detangling of manifolds can be viewed with neuronal images. Yet this is very difficult to see.

Further analysis shows that the V1 cortex sees the world through a narrow hole and that the V2 can do the same. After that the recognition gets better and better. There are three consistent mathematical ideas that allow the detangling of the above physiology:

• Idea 1: the visual system projects incoming information to higher dimensional places so that the data spreads more in the space.

• Idea 2: neuronal sources are present at every stage that correspond to the distribution of visual information from the real world.

• Idea 3: time implicitly ensures the supervision of manifold flattering.

The branch of computer science concerned with symbolic, (non)algorithmic methods for problem solving is called artificial intelligence (AI). An algorithm is a procedure that guarantees to find a fitting solution for a problem or that there is no solution. For many years, computers have been doing numerical calculations for data processing. However, knowledge about a certain subject is rarely numerical, but rather symbolical. Therefore, its problem-solving methods are rarely mathematical. It relies on heuristics, that are not guaranteed to work, but often find solutions quicker than trial-and-error. MYCIN is an expert system and designed to:

1. Provide expert level solutions for complex problems.

2. Being understandable.

3. Being flexible enough to easily accommodate new knowledge.

What is the context of the MYCIN system?

The MYCIN system exists from two parts. First, a knowledge base and second an inference mechanism. Sometimes there are subprograms added to facilitate interaction with users. There are two flows of information between the user and the system. The user either inputs information in the form of the description of a new case. The second flow is the output of advice and explanation by the expert system. These interactions all go through the user interface which then communicates to the system. The underlying knowledge base is the store of the program containing facts and associations about a subject. The inference mechanism or control structure can take many forms, for example a chained together set of rules about the facts and statements in the knowledge base. This is called forward chaining or data-directed inference because the data is known, by chaining a solution is provided. MYCIN however uses primarily backward chaining, a goal-directed control strategy. In this strategy, the system starts with an argument (the goal) and works backwards through inference rules to find the data that establishes that goal. Because of the many rule chains and data, the system should inquire, MYCIN is sometimes referred to as an evidence-gathering program. The goal of the work of the MYCIN is to provide diagnostic and therapeutic advice about a patient. The whole system is referred to as performance system as the other subsystems are not directly related to giving advice.

How was the programming language chosen?

For the creation of MYCIN, the LISP programming language was used. The reason is the extreme flexibility based on a small number of simple constructs. Furthermore, this language allows rapid testing and modification. This way, medical rules in the knowledge base were easily separated from inference procedures that use the rules. Also, LISP does not require recompilation of programs to test them. It removes the distinction between program and data, so parts of the program can be examined and edited as if they were data structures.  Currently with the release of Interlisp, additional tools such as EMYCIN have been built on top of Interlisp.

What is the historical perspective on MYCIN?

Production rules for artificial intelligence were probably brought up by Allen Newell. He saw an elegant formalism that was useful for psychological modelling. Production rules can be utilized to encode domain-specific knowledge. For example, Waterman (1970) worked with heuristics and rules of a poker game. This research led to the development of the DENDRAL program, the first AI program to emphasize the generalized problem-solving capabilities over specialized knowledge. The program was able to construct explanations of analytic data about molecular structures. A major concern here was the representation of specialized knowledge of the chemistry domain for the computer to use in problem solving. In a sense, MYCIN is an outgrowth of the DENDRAL program because the design and implementation are based upon it.

Before a system can make appropriate therapeutic decisions, it should contain all knowledge of antimicrobial selection. Also, by modifying the input from patient databases to direct data entry a consultation with physicians was realized. This made the system interactive. The model for MYCIN should be able to diagnose and suggest therapy. One difficulty was seeing how a long dialogue between therapist and patient could be focused on one line of reasoning at a time. Furthermore, translating the ill-structured knowledge of infectious disease into semantic networks remains difficult. In the first grant application of MYCIN, the goals of the project were described:

• A consultation program that provide physicians with advice regarding antimicrobial therapies. This is based on data on microbiology and direct observations of the physician.

• It should have interactive explanation capabilities to explain its knowledge of disease therapy and justify recommendations.

• It should contain computer acquisition of judgemental knowledge. The MYCIN system should be taught therapeutic decision rules, useful in clinical practice.

At the time of building MYCIN, researchers also worked on other subprojects in AI, such as question answering (QA), inference, explanation, evaluation and knowledge acquisition.

What about MYCIN’s task domain; antimicrobial selection?

The nature of the decision problem in MYCIN originates in the task domain: antimicrobial selection. An antimicrobial agent is a drug that is designed to kill bacteria or arrest growth. Selection of antimicrobial therapy is the problem of choosing an agent (or combination) to treat a patient with an infection. A naturally occurring bacteria or fungi is called antibiotic. Some are too toxic in treating infectious diseases. Besides antibiotics, antimicrobial or synthetic antibiotics can be used in treatment of infections. The cause of the infection is used as a clue for the decision what drugs are beneficial. The selection of therapy is done in four parts:

1. The physician decides whether the patient has a significant infection.

2. If this is the case, the organism causing the infection needs to be identified.

3. Select a set of drugs that might be appropriate.

4. Finally, the (combination of) drugs should be chose for treatment.

How can useful drugs be selected for the patient?

First, the isolation of bacteria from a patient is not evidence of a significant infection, they are often important to the homeostasis of the patients’ body. The second challenge is samples can be contaminated with external organisms. It is wise to use several samples. The significance is then based on clinical criteria. These allow the physician to judge the significance of the infection. Second, there are several laboratory tests that can check the causing organism of an infection. The complete test and definition of identity can take up to 24-48 hours. The problem with the process is that the patient cannot wait several days before therapy. Early data becomes important for narrowing down the possible identities. Historical information about the patient is also useful. Lastly, to discover the range of antimicrobial sensitivities of the organism, in vitro tests are run. The bacterium is exposed to several commonly used antimicrobial agents and the sensitivity is analysed. Then the physician knows what drugs can be effective in vivo (in the patient). This data is only available after several days; therefore, a decision often is based on statistical data available from hospital laboratories. Once a list with potentially useful drugs is created, the likelihood of its effect should be considered. The patient’s allergies, sex, age and kidney status should be examined, just like the administration of the drug. As the patient’s status can vary, so should the recommended dosage of the drug.

When came the evidence for needing assistance?

The use of sulfonimines and penicillin cannot be overstated. In the 1950s it became clear that antibiotics were misused. At the time of developing MYCIN this misuse was receiving a lot of attention. Many antibiotics were prescribed without the identification of the offensive organism. Antibiotics is overprescribed because the patient demands a prescription with every visit. Improved public education is one step towards diminishing the problem. Studies have shown that one-third of hospitalized patients receive a kind of antibiotics. The monetary cost is enormous. This issue is summarized by Simmons and Stolley in 1974.

1. Have new resistant bacterial strains have emerged because of the wide use of antibiotics?

2. Has the hospital ecology changed because of this use?

3. Have infections changed because of antibiotics misuse?

4. What are trends in the use of antibiotics?

5. Is there a proper use of antibiotics?

6. Is the more frequent use of antibiotics presenting new hazards?

The answers to the questions are frightening to an extent that the consequences might be worse than the disease. This raises a new question: are physicians basing their prescription habits on rational decisions? In a study of Roberts and Visconti (1972) it turns out only 13% was judged to be rational. The goal of MYCIN is to provide an improved therapy decision. An automated consultation system could support in a partial solution to the therapy selection problem.

The difference between male and female robots is that after they mate, the woman is unproductive for a set period, while the male can mate again. Therefore, males have a greater variance in reproductive success because they are actively looking for a scarce resource. Also, reproductive females are less active compared to males. They wait for males to find them. On the other hand, non-productive females are as active as males when it comes to looking for food. They only look for food and are not interested in anything else. The final difference was in the preference of type of food and offspring care. Only if males do not have parental certainty.

What is the objective of this study?

Mating has different consequences for male and female. The objective of this study is to describe the behavioural consequences of the single difference that males can reproduce after mating while females cannot. The idea behind this is to reproduce this phenomenon in artificial systems. If the system will behave the same as the real phenomenon, the model that was used can help explain the phenomenon. The robots used in the experiment are not created by the researchers, but they are male and female robots living and reproducing in an artificial environment. This will teach the differences in evolved behaviour between male and female robots. Robots can be useful to illustrate the difference between man and woman because when constructing robots, the behavioural assumptions need to be stated specifically.

One of the differences is that reproductive female robots are less active than male robots. However, when females mate successfully this is not the case. The stereotype of active males versus reactive females must be guarded against. It should be objectively measured. The robots that are used in this case are highly simplified, and the results of the study may apply to some species, but not to others. The nature of sexual differences in species depends on the adaptive pattern of certain species. The robots used in this experiment provide a better understanding in observed behaviour in real animals plus their evolutionary origin.

Not many experiments have been done on sexual differences in robots and computational models on the behaviour of male and female organisms. Studies that try to investigate genetic algorithms are done by recombining genotypes of two different animals. The only difference between the male and female robots in this experiment is the consequence of mating. It’s assumed the species are sexually different, without specifying the ways in which they are. The only dissimilarity is appearance. The goal is to find out if male and female robots behave differently and what these differences are.

How is the experiment executed?

A group of hundred Khepera robots was used in this study. They all have an energy level ranging from zero to one, and when the energy level is zero, the robot disappears. Food tokens are used to represent eating. Half of the robots in the experiment are male, the other half is female. Mating happens when two robots of a different sex touch. In the end, the number of mating experiences is counted, and the top ten males and top ten females generate offspring. This represents the evolution and is repeated 10 times for 1,000 generations. The behaviour of the robots is controlled by a neural network encoded into the genes of that robot.

What were the results of the experiment?

The variance in reproductive success is the same for males and females. However, the male robots have a higher number of mating events. There was an increase in number of mating events, so the robots evolve in the ability to successfully mate and eat. Length of life and number of food tokens is correlated, but not perfectly. It seems that females can eat more efficiently. Length of life and mating success was correlated differently for males and females; female mating success is strongly correlated with length of life, which is not true for males. This implies that there are two adaptive strategies used.

Males and females produce different results in terms of mating and eating. Females behave differently depending whether they are reproductive or nonreproductive. Males were always active in looking for food or reproductive females, whereas only nonreproductive females were active. Reproductive females were not active and were waiting for males. Nonreproductive females were active, but were only looking for food, not for males.

The behaviour of the robots in the experimental laboratory was studied by a controlled experiment. The robot is given the choice between two items. The items can be male, reproductive female, nonreproductive female or a food token. Males prefer the reproductive female over food and even stronger over a nonreproductive female. When the male must choose between a nonreproductive female and food, it prefers food. Reproductive females prefer food over a male, also when she needs to choose between a male and a nonreproductive female, the choice of male is just a little more preferable than the nonreproductive female.

Finally, the choice between different types of food was simulated by colouring the available food. Males do not prefer one type of food over another. Females have specific food preferences dependent on their environment. Reproductive females prefer more energetic food. When female robots are present, the male tents to eat less (energetic) food. This can be explained by males being socially aware of their environment and the presence of females. They will behave accordingly.

What is suggested for further research?

In the article, five main suggestions are made to elaborate on the current research.

• The emergence of families could be studied. An important implication is that the life of the robots should then be linked to specific places. They should be able to recognize the location of these places.

• The robots are described as male and female but are not individually different. The influence of sexual attractiveness could be measured to mirror sexual selection. Things that could be considered are recreational sex, which is mating between a male and a nonreproductive female. Also, it could be that sexual attractiveness is different for nonreproductive versus reproductive females.

• Motivations for animals are different and rely on emotional states and their expression. This is because not more than one motivation can be pursued at one time. Emotional states help guide what decisions are made and are based on the interaction of brain with the rest of the body. However, the expression of emotions that are related to sex and parenting might be a whole other subject for research.

• A limitation of the current study is that stereotypes need to be more realistic. One example is that male robots only inherit genotypes of their father whereas females inherit the genotype of their mother. There is no crossover.

• Lastly, female menopause and grandmothers could be considered. In real life, females become permanently nonreproductive. Female robots in menopause could be used to explore the grandmother hypothesis: that females continue to live after menopause to help their daughters take care of their offspring.

The medial temporal lobe (MTL) plays an important role in memory. The study discussed in this article examines how cell activity patterns transform visual information into long-term memory memories such as faces. The question of how this happens has plagued neuroscientists for decades. Evidence comes from electron physiology and lesion studies in monkeys. This shows that there is a hierarchical organization along the ventral visual pathway from the visual cortex V1 to the inferior temporal cortex (IT). Visual stimuli come in through this way and are processed and stored. How this information is exactly presented in there, remains a mystery.

Hypotheses

Two ideas for this representation have been drawn up based on two visions:

• Idea 1: The "distribution population coding" vision states that a stimulus is represented by the activity of a large number of neuron combinations. Each neuron stands for a certain figure.

• Idea 2: The “sparse coding” vision states that a perception is represented by smaller neuron combinations, of neurons that respond specifically to specific figures, objects and concepts. Because of this, neuron scientists thought that one neuron corresponds to an object or person, regardless of how this was observed. These cells were called grandmother cells. The question is whether these cells exist at all?

Neurons in IT and MTL

The IT sends information to the MTL. It is important to know that the MTL is not a homogeneous structure with a single function. Neurons in the MTL respond selectively when composing gender and facial expression on photos and in real life with both acquaintances and strangers. This combination of selectivity leads to an explicit representation in which a single cell can serve as an indicator to see which person is being depicted.

These cells resemble grandmother cells, but it is impossible that only one cell responds. In any case, it is even more unbelievable that you can find this cell. In addition, the cell may respond to several people, because in one study you cannot visit all the people in the world. It has been proven that some units fire at several people. Finally, theoretical findings estimate that each unit must have 50-150 cells to distinguish between people. 

It is important to investigate whether these abstract cells can also be found in animal species. In rats, these abstract cells would be expected to be located in the neocortical areas for object recognition, such as the IT cortex. However, it is very difficult to detect such a group of firing neurons.

Research with learning tasks showed that after an unsupervised learning task, the previously random-firing neuron units became units that simultaneously fire at one object from the learning task. In this way this object or individual would be recognized. Most of these units respond uniquely to a single individual.

The findings of neurological patients showed that the MTL is not necessary for visual recognition. The hippocampal rhinal system is involved in long-term declarative memory. The MTL cells, described in the previous sections, connect visual perception to memory. So they do not really provide recognition. Recognition happens after about 130 ms, while the units (from the previous section) only started firing at a certain stimulus between 250-350 ms.

If the MTL is involved in memory, it is likely that these neurons do not know all the differences within visual details. The existence of category cells such as the units (which respond to individuals) corresponds to the vision that they encode aspects of the meaning we have of a stimulus that we want to remember. These cells are probably also involved in learning associations. These units are therefore not grandmother cells.

Detecting and recognizing objects in the context

Detecting and recognizing meaningful objects in complex environments is a crucial skill that ensures that we can survive. The recognition process is fast, automatic and is seen as being present standardly. However, it is not constructed as easily as people think. After much research, it has become apparent that the recognition process takes place in the ventral visual system. Broadly speaking, this is done through a number of phases: V1 to V2 to V4 to PIT to AIT (the PIT and AIT together form the IT that you see in other articles). The highest phase of this system is the AIT. It is thought that there are neurons in the AIT that specifically recognize objects.

On a theoretical level there is a growing appreciation for the role of learning in robust, crucial representations of object recognition. The role of learning is approached by studying the visual system. Important points in this regard are the development of the visual system and the way in which the system is used by young people and adults. This paper focuses on experience-related plasticity in the adult visual system. Many adult studies have shown that learning-dependent changes are taking place in distinguishing and recognizing stimuli from tasks. Recent studies have searched for locations in the brain that confirm this.

Objects

The ability to form neuronal representations that are sensitive to pixel combinations of shapes and colors comes from object recognition. This is the quality mark for vision and goes through arithmetic difficulties so that it remains highly sensitive to the identity of image changes. Object recognition requires the visual system to discriminate between different patterns of input. Discrimination is probably done by merging a number of neurons in the early stages of the process. As a result, neurons from higher phases know when to fire at certain patterns, which leads to discrimination and recognition. Arithmetic models have proven that such explicit object recognition can be built using neuronal connections from groups that fire together with similar characteristics of images.

Learning sensitivity

The plasticity of neuronal connections combined with suitable learning rules is a potential mechanism. An example of this is that learning to respond selectively goes hand in hand with learning which things often occur (together) in reality. On a mechanistic level, this occurs when certain neurons respond to the input. An interaction of this strategy per phase of the visual system results in a complex stimulus system, which ensures recognition. Recently evidence has been found for the above theory. The timeframe for learning such recognitions suggests a strong link between neuroplasticity and behavioral improvements.

Learning tolerance

Most studies on visual learning have focused on changes in neuronal or behavioral selectivity. Learning selectivity is not enough for object recognition. Selective object representations must be tolerant of image changes (size, color, angle). Learning from the natural world is the solution. A central idea is that features and objects from the world do not suddenly exist or do not exist, but have (temporary) continuity. If you see an object and you walk on, you still see that object, but always from a different point. This is learning tolerance, because you know that it is still about the same object. That way you recognize it from multiple points of view. Evidence has been found that tolerance is not automatic, but it is not known how it works neuronal.

The visual system must also learn to recognize objects among other objects; called clutter. Therefore, it is suggested that learning has to do with improving correlations between neurons that respond to a characteristic or target against a background of noise. However, you should not always see the noise as noise. This background noise can also ensure that you place something in the right context.

Neuron plasticity

It is often said that the plasticity of perceptual learning is in early visual phases, because this learning is linked to the position of the retina. This means that changes in the receptive field can lead to certain tunings of V1 neurons. Recent image studies have found influence of the V1 when learning object characteristics. However, the evidence remains controversial. One possibility is that V1 learning effects can be found in the average response of a large group of neurons measured with fMRI.

Shape representation can shift from high to low visual areas. This supports fast and automatic searching and discovery with attention control (in cluttered scenes). These findings are consistent with the suggestion that object representation is not only helped by bottom-up processing but also by top-down processing. Learning starts in higher visual areas for easy tasks and continues to lower visual areas if higher resolution is required for more difficult tasks.

One of the biggest advantages of fMRI is that it shows global images of the brain. fMRI is therefore a handy method for studying the brain during visual tasks. It has been found that learning is supported by selective processes with crucial characteristics of objects. Recent neuronal imaging studies have shown that learning is supported by functional interactions between occipital temporal and parietal frontal areas. The findings of these image studies are consistent with the top-down approaches to visual processes. These areas together create a perceptual representation of the world.

In summary, it can be said that current studies indicate that there is no fixed place for brain plasticity in visual learning. At the neuronal level, learning can arise from changes in the feed forward network, especially at higher phases, or through changes in interactions between frontal cortical areas and local connections in the primary visual cortex. Such changes can be adaptive and efficient.

The four conclusions that now can be drawn are:

1. The adult visual system is plastic.

2. There is no place of plasticity that supports object recognition learning.

3. Learning is not always the result of simple, static change at the core of the feedforward network.

4. The relationship between neuronal mechanisms and plasticity remains unknown.

Object recognition takes place in the ventral visual system in the cortex. This system runs from the visual V1 area to IT. From there, there are connections with the PFC that ensure that perception and memory are connected. The further down the path, the more specific the neurons are and the greater their receptive fields. Plasticity and learning are probably present in all phases of object recognition.

It is not known what feedback will be given from the phases. The hypothesis is that the basic processes of information go feedforward and are supported by short time limits that are required for specific responses. However, this hypothesis does not exclude feedback loops. The feedforward architecture is a reasonable starting point for the theory of the visual cortex aimed at explaining direct object recognition. The recognition phase can also play a role.

Model based on the feedforward theory

The model used here is a direct extension of the Wiesel and Hubel model. The model is a feedforward theory of visual processes. Every feedforward theory distinguishes between two types of cells; simple and complex cells. This creates a distinction between selectivity and invariance. This model uses an input of 7 × 7 boxes. The boxes are first analyzed by a multidimensional row of simple S1 units that respond best to lines and angles. The S1 units can be on and off if the figure they represent is present. The next level is C1. C1 looks like complex cells from the striatum. Every complex C1 unit receives information from a unit of S1 cells with the same orientation. However, there may be slight differences and C1 can receive multiple S1 units, making the C1 unit less sensitive to position and size.

At the C level, complex cells are thus converted into a two-dimensional image by combining the afferent S cells. The results are calculated with firing neurons. This creates a complex model with how a representation is created in the brain of a visual image. Reversing a C level is also called a MAX operation.

Not all features of a complete feedforward network can be found in this model. That would be too complicated. Every model tries to describe and explain a piece. Because we still have little knowledge of how everything works exactly, this remains very complicated. So you also have to put the different theories (and articles) next to each other and try to understand and apply them to each other. Just as in this model the MAX returns.

Results

The above data was presented and measured with a task where test subjects had to indicate whether or not an animal was presented. The showing of the pictures was very short. Yet people can then quickly and accurately decide whether or not there is an animal. This means that our object recognition works very quickly. If the image was displayed a little shorter this also had little effect. The model was used to view the similarities with people.

A comparison between human performance and the feedforward model in animal tasks was measured with (d'); a monotonic function of the performance of the observers. The results showed that people and the model responded about the same and that there were similar reactions. Recent studies have even shown that animal recognition also works if, for example, the image is shown upside down. People were a bit slower when using a concealment component.

Discussion

This model improves the old model on two important points.

1. A learning phase has been added. The hierarchy of the visual system is viewed better. Learning takes place in two independent phases. First during development with unsupervised learning, then at task-specific circuits.

2. The new model is closer to the anatomy of the visual cortex. This model also suggests (along with other theories) that an initial feedforward step is driven in bottom-up processes for a basic representation, built from a basic dictionary with general figures.