text stringlengths 128 2.05k |
|---|
Human brain imaging studies suggest that significant conceptual binding of objects, events, actions, feelings etc. exists in different mirror systems. For example, feeling disgust and seeing someone disgusted activates overlapping regions in the brain |
. This is also the case for pain . It would, therefore, be possible to claim that these overlapping areas represent the concept of pain, and the neural (population) activity pattern there could be considered as the neural symbol representation for pain. Opposing this view, however, is that these are gross symbols that ... |
From an evolutionary viewpoint such a neural system prepares the ground for symbol/concept externalization, and the start of language when combined with the recursive structure of action planning and imitation as elaborated in Arbib's language evolution theory |
Instead of the evolution of a syntax processing enabled brain related to speech production, Arbib argues that the neurons for manual actions have evolved into mirror systems for imitation and simple manual communication, which facilitated the biological evolution of the human language-ready brain that can sustain a pro... |
. The transition from protolanguage to language is then explained by a long history of human cultural evolution. The theory is supported by the findings that overlapping brain activity is observed in Broca’s for perceiving language and tool use |
, both of which involve manipulation of complex hierarchical structures. Furthermore, as Broca’s area is the possible homologue of the monkey ventral premotor cortex where mirror neurons for grasping are located, neural processes for manipulation and action understanding may be the evolutionary bases of language |
The crucial finding, refered to in this theory, was that the mirror system activity in the human frontal cortex was found to be in or near Broca’s area, which had been considered as an area of language processing. They hypothesized that changes in the brain that provide the substrate for tool use and language and that ... |
Therefore, it appears that in highly evolved animals, brains have ``invented'' basic symbols that may be used in rudimentary planning and decision making. In pre-humans, these basic symbols might have triggered an accelerated evolution once they were externalized via utterances, paving the way for speech and language e... |
III-C Cognitive science viewpoint: from concepts to symbols Cognitive scientists have been considering symbol systems for many decades |
. Similar to our discussion above (Section III-A ), most cognitive scientists and psychologists agree that the notion of a concept is very important in understanding symbols as well as cognition, as it enables us to classify, understand, predict, reason, and communicate |
However, researchers do not necessarily agree on the definition of concept . Furthermore, there is a related term, category and its cognitive competence called categorization , which needs to be distinguished from concept |
The following may help. When we saw a dog, we usually think ``we saw a dog,'' but not ``we saw a hairy living organism with two eyes, four legs, pointy teeth, etc.'' That is, instead of simultaneously interpreting the abundant individual features of an entity, we usually categorize the entity into one class (or a few c... |
Let us now try to distinguish concept from category. One of the most accepted definitions of concept is that it corresponds to an internal (memory) representation of classes of things |
. Different from this, one of the most accepted definitions of category is that of a set of entities or examples selected (picked out) by the belonging concept |
Clearly, those two terms are closely related and, thus, often used interchangeably in cognitive science and psychology, leading to controversies. For example, one notable source of disagreement concerns the internal representation asking about possible systems of symbols on which the internal representation of concept ... |
A symbol system that uses symbols independently of their (perceptual) origin as atoms of knowledge is called an amodal symbol system (ASS). However, Barsalou more strongly concentrated on the perceptual aspects of symbol systems and proposed the PSS |
. A PSS puts a central emphasis on multimodal perceptual experiences as the basis of its symbols. Let us describe the differences between ASS and PSS in more detail. |
First, the notion of category is different in ASS and PSS. ASS assumes that categories are internally represented by symbols without any modality, where symbol means a discrete variable, which is often given by a word (a name of a concept). By contrast, in PSS, every category is internally represented by symbols from s... |
PSS furthermore assumes that categories are (implicitly or explicitly) consolidated by simulating several candidate categories. For this, distributed and complexly connected perceptual symbols with background information (i.e., context or situations) are used. This suggests that categorization is a dynamic process and ... |
Second, how concepts are used to realize categories is also different. ASS assumes that a concept is acquired knowledge that resides in long-term memory, and, thus, concepts are more or less static. On the contrary, PSS assumes that concepts are dynamic. Note that in this way, there is a certain similarity between PSS ... |
There is an interesting phenomenon called ad hoc category . An ad hoc category is a spontaneously created category to achieve a given goal within a certain context |
. For example, ``things to take on a camping trip'' is an ad hoc category. Ad hoc categories are shown to have the same typical effects seen in more conventional categories such as dogs and furniture. This somewhat odd but interesting cognitive phenomenon can be well described by dynamic simulation processes that utili... |
To date, the ASS perspective still dominates in cognitive science and psychology, probably because it has been more manageable to conduct research assuming static ASS-like representations. PSS, however, seems to be better matched to the dynamics of brain structures and activities, and to robots that need to deal with r... |
. So far, however, it has been quite difficult to empirically evaluate the PSS theory as a whole. Potentially, building better computational models for symbol emergence in cognitive developmental systems might aid in making progress concerning the ASS–PSS controversy in cognitive science. This review certainly leans mo... |
III-D Developmental psychology viewpoint: from behaviors to symbols The idea that symbols and symbol manipulation emerge from infants' physical and social interactions with the world, prior to their first words and linguistic activity, has been actively discussed since the times of the pioneers of developmental psychol... |
III-D Theories of symbol development The theories of symbol emergence and development in developmental psychology can be separated into two categories: classical and contemporary. Classical accounts were developed by Piaget, Vygotsky, and Werner and Kaplan, and contemporary accounts can be separated into three main app... |
as well as the book of Werner and Kaplan entitled Symbol formation already discussed the gradual emergence of symbols from nonlinguistic forms of symbol-like functioning, and the integration of the formed structures into language skills. In contemporary accounts, empiricist and nativist theories consider a child to be ... |
According to Mandler, the traditional view of infants’ inability of concept formation and use is flawed. Her studies on recall, conceptual categorization, and inductive generalization have shown that infants can form preverbal concepts. An infant can form animals and inanimate distinction from only motion and spatial i... |
The levels of consciousness model developed by Zelazo consists of four levels . The first level is called stimulus bound, ranging from birth to 7 months of age, in which an infant has representations tied to stimuli. In the second level, which is decoupling of symbols, ranging between 8 and 18 months, infants can subst... |
. For Tomasello, symbolic communication emerges as a way of manipulating the attention of others to a feature of an object or an object in general. Up to 9 months of age, behaviors of children are dyadic. Within the age range of 9–12 months, new behaviors emerge that include joint attention and triadic relationship, i.... |
. This view reminds us that the social abilities of an agent would also be critical. Theories of symbol emergence in developmental psychology have been briefly mentioned here. For further information, please see |
III-D Precursors of symbolic functioning Bates et al. discuss that the onset and development of communicative intentions and conventional signals (such as pointing or crying for an unreachable toy), observed between 9 and 13 months, can be viewed as a precursor of symbolic communication. |
Before 9 months, infants can use some signals (such as crying). However, these signals are geared more toward the referent object than toward the adult that can help. By 9 months, signals and gestures become clear, consistent and intentional. For example, previously object-targeted cry signals are now aimed at adults. ... |
The period before the first words also corresponds to Piaget's sensorimotor stage V, the stage where infants can differentiate means from ends and use novel means for familiar ends. After approximately 9 months of age, infants start using learned affordances to achieve certain goals, predicting desired changes in the e... |
. By 12 months, they can make multistep plans using learned affordances and perform sequencing of the learned action–effect mappings for different tasks. For example, they can reach a distant toy resting on a towel by first pulling the towel or retrieve an object on a support after removing an obstructing barrier |
. These data suggest the importance of early sensorimotor and social skill development in infants that are probably integrated into the language and reasoning network, where symbols and symbol manipulations play an important role. Learning a symbol system is also a continuous process, in which infants benefit from diff... |
When we are away from ASS and do not assume pre-existing linguistic knowledge, our view has common grounds with Tomasello's usage-based theory of language acquisition |
. The usage-based approach to linguistic communication has two aphorisms, i.e., meaning is use, and structure emerges from use. The first aphorism argues that the meanings should be rooted in how people use linguistic conventions to achieve social ends. The second argues that the meaning-based grammatical constructions... |
The first words, which can be seen as the clear indication of symbolic knowledge, are observed at around 12 months of age, when infants discover that things have names. For a thorough review of symbol emergence in infants and the role of nonlinguistic developments in symbol and language acquisition, see |
III-D Language, graphical, and play symbolic systems The emergence of symbolic functioning is observed in infants in different domains, such as comprehension and production of verbal or sign language, gestures, graphic symbols, and (pretend) play. Efforts to clarify these range from investigating the capabilities of di... |
to assessing the abilities of children in pretend play, which involves the ability to play with an object as if it were a different one |
. Whereas the former is related to dual representation problem referring to understanding the symbol-referent relationship, the latter is related to the triune representation problem referring to using objects in a symbolic way. It is important to note that related abilities of infants in different domains appear to re... |
. Both language and play acquisition begin with presymbolic structures, where action and meaning are fused (e.g., plays only related to infant's own sensorimotor repertoire), then abstract and agent-independent symbols emerge in both domains, and finally combinatorial use of the abstract symbols is observed |
Acquisition of language, among these domains, is special in humans. Young infants are capable of producing language-like sounds early in their first year: they can produce canonical babbles followed by protowords at month 7 and first words shortly after the first birthdate |
. As we discussed in the previous subsection, the emergence of language comprehension is estimated to be around 9 to 10 months of age and language production in the first half of the second year |
. The symbolic status of some of the early words might be specific to events and contexts, whereas other words might be used across a range of contexts |
. With the development of naming insight early in the second year, when the child discovers that things have names, and names and their references are arbitrarily related, one can argue that symbolic functioning is in operation in the language domain. |
Graphical or pictorial symbolic capability, however, is observed later: 2.5 year old children (but not younger ones) can retrieve hidden objects when the pictures of these objects are presented |
, or can match pictorial symbols with imaginary outcomes of actions given pictures of objects to which actions are applied Whereas there is no consensus on the existence of a single symbolic function/system that develops during early childhood and is realized in different domains, the symbolic systems in different doma... |
. Only after 3 years of age, children gain the capability to retrieve objects only using on their pictorial symbolic description without using language support. However, Callaghan and Rankin also showed that with more training, children could acquire the capability of pure pictorial symbolic functioning earlier than 3 ... |
From this discussion, we can conclude that early acquisition of the language symbol system compared to other domains is possibly due to the heavy parental scaffolding from birth |
, the symbol systems observable in different domains have similar developmental stages, they interact with each other, and finally their development depends on the amount of scaffolding and maturity of the system. |
IV Problem history 2: Symbol emergence in artificial systems The main challenge concerning symbol emergence in cognitive developmental systems is to create artificial systems, e.g., robots, that can form and manipulate rich representations of categories and concepts. Our overview will concentrate on this problem in AI ... |
IV-A Artificial intelligence viewpoint: from tokens to symbols Originally, the idea of a symbol system in AI had its roots in mathematical/symbolic logic because all programming languages are based on this. Alas, this led to quite an inflexible image concerning the term symbol |
Inspired by the work of Newell , AI has tended to consider a symbol as the minimum element of intelligence and this way of thinking is hugely influenced by the physical symbol system hypothesis , p.116] (see Section III-C ): |
``A physical symbol system consists of a set of entities, called symbols , which are physical patterns that can occur as components of another type of entity called an expression (or symbol structure). Thus, a symbol structure is composed of many instances (or tokens) of symbols related in some physical way (such as: t... |
This notion creates a strong link between a token in a Turing machine and a symbol. However, there are two major problems with this definition, |
First, the physical symbol system hypothesis assumes that a symbol exists without any explicit connection to real-world information. Owing to the missing link between such symbols and the real world, such a physical symbol system is ungrounded, and therefore unable to function appropriately in complex environments. Lat... |
, symbol anchoring , and the ``intelligence without representation'' argument , challenged the conventional understanding about the implementation of a symbol system. |
Second, the token–symbol analogy is based on the assumption that a symbol deterministically represents a concept that does not change its meaning depending on the context or interpretation. However, this view is very different from that discussed in semiotics (see Section II ). There, the human (interpretant) takes a c... |
This generates confusion among researchers, especially in interdisciplinary fields such as cognitive systems, in which aspects of robotics and AI are combined with psychology, neuroscience, and social science, attempting to advance our understanding of human cognition and interaction by using a constructive approach. P... |
Indeed, a general and practical definition of symbol that includes both aspects would be desirable. At the same time, the different meanings of this broad concept should be kept clear. In some of the literature on symbol grounding |
these two aspects are referred to as physical symbol grounding and social symbol grounding . Another definition is proposed by Steels |
, who refers to these different types of symbols as c-symbols and m-symbols, respectively, pointing out the ill-posed characteristics of the original symbol grounding problem. |
Studies in developmental robotics have been conducted to solve the symbol grounding problem. For example, Steels et al. have been tackling this problem using language games from the evolutionary viewpoint |
. Cangelosi et al. developed robotic models for symbol grounding . For the further information please see Steels even said ``the symbol grounding problem has been solved |
.'' However, it is still difficult for us to develop a robot that can learn language in a bottom-up manner and start communicating with people. |
To deal with these problems, we need to move on from the symbol grounding problem to symbol emergence by taking the process of bottom-up organization of symbol systems into consideration. This means that we need to develop a learnable computational model representing PSS that is also affected by social interactions, i.... |
IV-B Pattern recognition viewpoint: from labels to symbols Pattern recognition methods, e.g., image and speech recognition, have made great progress with deep learning for about five years |
Most pattern recognition systems are based on supervised learning . A supervised learning system is trained with input data and supervisory signals, i.e., desired output data. The supervisory signals are often called label data because they represent ground truth class labels. |
What are symbols in pattern recognition tasks? Indeed, the class labels mentioned above were regarded as signs representing concepts of objects and the input data as signs observed from objects. Image recognition can, thus, be considered as a mapping between an object and a sign representing a concept. Note that, here,... |
The general pipeline for image or speech recognition is shown in Fig. . Conventional recognition systems use pre-defined low-level features, mostly defined by a human, to structure the input data prior to recognition. Modern deep learning performs end-to-end learning and features exist only implicitly therein (as the a... |
, but it also makes supervised-learning-based pattern recognition a well-defined problem. Alas, this assumption led to large side effects on our way toward understanding symbol systems. From the above, it is clear that this type of pattern recognition essentially assumes static symbol systems. |
However, when we look at the interior of the pattern recognition systems using deep neural networks, we can find that neural networks form internal representations dynamically, even though the learning process is governed by static symbol systems. For example, it was found that convolutional neural networks (CNNs) form... |
Those features themselves are not symbols. However, to realize symbol emergence in cognitive systems the cognitive dynamics that form rich internal representations are crucial. |
In summary, pattern recognition has been providing many useful technologies for grounding labels (signs) by real-world sensory information. However, these methods generally assume a very simple dyadic notion about symbols, i.e., the mapping between objects and their labels. From a developmental viewpoint, a neonatal de... |
IV-C Unsupervised learning viewpoint: from multimodal categorization to symbols It has been shown by a wide variety of studies that it is possible to create an internal representation system that internalizes a symbol system, reproduces categories, and preserves concepts by using unsupervised clustering techniques |
. One example where the complete chain from sensorimotor experience to language symbols has been addressed is the work of Nakamura et al. as discussed in the following. |
The central goal of the studies was that a robot creates a certain conceptual structure by categorizing its own sensorimotor experiences. When this happens using multimodal sensorimotor information, this can be regarded as a PSS model. In addition, the use of multimodal sensorimotor information for forming categories i... |
This has been called multimodal object categorization , realized by applying a hierarchical Bayesian model to robotics . It can be achieved by multimodal latent Dirichlet allocation (MLDA), which is a multimodal extension of latent Dirichlet allocation (LDA)—a widely known probabilistic topic model |
. This way, a robot can form object categories using visual, auditory, and tactile information acquired by itself. An important point here is that the sensorimotor information acquired by the robot generates categories that are, at first, only suitable for this robot. However, it is also possible to add words as an add... |
. This way, the created categories become much closer to human categories, owing to the co-occurrence of words and the similar multimodal sensory information. Thus, this represents a probabilistic association of words, signs, with a multimodal category, which related to a certain concept—a process of symbol formation. ... |
One interesting extension of this model concerns the self-acquisition of language combining automatic speech recognition with the MLDA system. Here, unsupervised morphological analysis |
is performed on phoneme recognition results in order to acquire a vocabulary. The point of this model is that multimodal categories are used for learning lexical information and vice versa. This way, the robot could acquire approximately 70 appropriate words through interaction with a user over 1 month |
. This suggests that unsupervised learning, e.g., multimodal object categorization, can provide a basic internal representation system for primal language learning. |
There have been many related studies. For example, Taniguchi introduced the idea of spatial concept and built a machine learning system to enable a robot to form place categories and learn place names |
. Mangin proposed multimodal concept formation using a non-negative matrix factorization method . Ugur et al. studied the discovery of predictable effect categories by grouping the interactions that produce similar effects |
. They utilized a standard clustering method with varying maximal numbers of clusters, and accepted a number of clusters only if the corresponding effect categories could be reliably predicted with classifiers that take object features and actions as inputs. Predictability also performs a central role in the series of ... |
All of these studies suggest that bottom-up concept formation, using sensorimotor information obtained by a robot itself, is possible, and this appears to be a promising avenue for future research. |
IV-D Reinforcement learning viewpoint: from states and actions to symbols Whereas the aforementioned work focuses on finding structure in sensory information, decision-making and behavior learning are fundamental components of cognitive systems as well. This problem is often formulated as a reinforcement learning probl... |
The goal of a reinforcement learning agent is to find an optimal policy that can maximize the total accumulated reward obtained from its environment. In the context of reinforcement learning, the term symbol has been used related to state and action abstraction with the belief that symbol is an abstract discrete token ... |
Reinforcement learning is most commonly formulated as learning over a Markov decision process, which often assumes that an agent has a discrete state–action space. However, if its action and state spaces are too large, the curse of dimensionality prevents the agent from learning adequate policies. Therefore, how to des... |
State abstraction was commonly treated as a discretization of the state space described by a discrete set of symbols, as in some form of tile coding |
. Another way of abstraction is to use a function approximation that can map continuous states or a large discrete state space directly to action values. For example, these abstract states may be represented by activations in neural networks, i.e., distributed representations |
. This has more recently led to an explosion of work in deep reinforcement learning (DRL), which has involved behavior learning against discovered, higher-level representations of sensory information |
. In the context of DRL, state abstraction is represented in the same way as high-level feature representation in CNNs in pattern recognition systems |
Action abstraction has similarly provided a useful mechanism for learning. Operating and learning exclusively at the level of action primitives is a slow process, particularly when the same sequence of actions is required in multiple locations. Action hierarchies can then be defined through options or macro-actions |
, which are defined as local controllers and constructed from primitive actions (or simpler options). These can then be treated as new actions, and can be used in planning as single atoms or symbols referencing more complex behaviors |
. In the reinforcement learning literature, these are learned as behaviors with some functional invariance, for example as controllers that lead to bottleneck states |
, or as commonly occurring sub-behaviors . Additionally, in more recent work, hierarchies of actions can be defined to allow for concurrent, rather than sequential, execution |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.