text stringlengths 128 2.05k |
|---|
These different forms of abstraction are not only useful for improving the tractability of problems, but they also provide a useful form of knowledge transfer, where solving one task provides an encoding of the environment that facilitates better performance on the next task |
Mugan and Kuipers implemented a system that learns qualitative representations of states and predictive models in a bottom-up manner by discretizing the continuous variables of the environment. In another study, Konidaris et al. studied the construction of ``symbols'' that can be directly used as preconditions and effe... |
and probabilistic plans in simulated environments. Note that the usage of the term symbols in this study simply corresponds to internal feature representations, even though it looks like ``symbols'' in terms of symbolic AI. |
These studies all investigated how to form symbols , i.e., internal feature representations, in the continuous sensorimotor space of the robot. However, complex symbols can be formed by combining predefined or already-learned symbols. For example, Pasula et al. |
and Lang et al. studied the learning of symbolic operators using predefined predicates. Ugur et al. re-used the previously discovered symbols in generating plans in novel settings |
In the reinforcement learning literature, most work related to symbol emergence was about the formation of internal feature representation systems, in particular state–action abstractions, for efficient behavior learning. State and action abstraction is regarded as a part of symbol emergence, but not symbol emergence i... |
Recently, several studies extended the framework of reinforcement learning and enabled an agent to learn interpretation of linguistic, i.e., symbolic, input in the context of DRL |
. Most of the studies remain at the preliminary stage from the viewpoint of natural language understanding. However, this is also a promising approach to model symbol emergence in cognitive developmental systems. |
IV-E Dynamical systems viewpoint: from attractors to symbols Any agent (human, animal, or robot), which is physically embodied and embedded in its environment, can be described as a continuous dynamical system. The question, which has been addressed by several researchers, is whether discrete states or proto-symbols ca... |
The notion of attractors in a nonlinear dynamical system provides a natural connection: the attractor (no matter whether a fixed point or a limit cycle) can be seen as a discrete entity and there would typically be only a limited number of them. Pfeifer and Bongard 11 , pp. 153–159] offer the example of a running anima... |
This question is also addressed by Kuniyoshi et al. , who make use of the mathematical concept of structural stability: the claim is that a ``global information structure'' will emerge from the interaction of the body with the environment. Kuniyoshi concludes that because of the discrete and persistent nature, one can ... |
Whereas these studies have addressed artificial agents, we note that their viewpoints are also strongly related to the above discussion about the mirror system in primates as a possible neural instantiation of a proto-symbol structure. |
Integrative viewpoint In this section, we integrate the viewpoints described in Sections III and IV , and provide a unified viewpoint on symbol emergence. |
V-A Wrapping things up (from the perspective of PSS) It has been a generally accepted view in AI (Section IV ) and cognitive science (Section III-C ) to regard symbol systems as internal representation systems. Cognitive neuroscience also follows this way of thinking ( III-B ). However, this idea has been affected by t... |
Furthermore, we discussed that there are two types of usages of the term, symbol system, considered by the different fields: symbol systems in society as compared to the internal representational systems in our brain. Whereas both are clearly linked to each other through processes of externalization, e.g., language gen... |
In this paper, we start with PSS. The focus on PSS chosen here is motivated by its more dynamic nature that better fits the requirements of cognitive developmental systems. Concerning PSS, we can summarize that this theory assumes that internal representation systems, i.e., PSSs, are formed in a bottom-up manner from p... |
The bottom-up development assumed by PSS leads to the fact that such a PSS can be regarded as a self-organization process of multimodal sensorimotor information in our brain. This view is quite similar to the dynamics of the schema model, proposed by Piaget to explain human development during the sensorimotor period |
What we need here for an integration of all this is as follows. (1) If we distinguish symbol systems in society and those in our mental systems, and (2) if we consider the dynamics and uncertainty in symbol systems, and (3) if we also take the context dependency of symbols (i.e., semiosis) into account, and, finally, (... |
V-B Symbol emergence systems Figure shows a schematic illustration of symbol emergence systems, describing a potential path to symbol emergence in a multiagent system |
. In this figure, we distinguish between an internal representational system and a symbol system owned by a community or a society. |
At the start, human infants form their object categories and motion primitives through physical interaction with their environment. Soon, however, they will also generate signs, e.g., speech signals and gestures, to accompany their actions, or to express their wishes. Alongside the development of the infants' ability t... |
The child's symbols, that had originated from bottom-up sensorimotor processes, will in this way drift toward the symbol system shared by their community, at first the infant and parents. Hence, language learning can be regarded as both the bottom-up formation process and the internalization of a symbol system shared i... |
In general, all such systems evolve within their communities and they follow certain rules, which impose constraints on all agents that make use of them. Without following these rules, e.g., semantics, syntax, and pragmatics, communication would be hard or even impossible. Any agent is free to form categories and signs... |
The concept of emergence comes from studies of complex systems. A complex system that shows emergence displays a macroscopic property at some higher level that arises in a bottom-up manner based on interactions among the system's lower-level entities. In turn, the macroscopic property constrains the behaviors of the lo... |
Figure depicts this for our case. The sensorimotor and also the direct social interaction processes of the individual reside at the lower levels. The resulting symbol system represents a higher-level structure, which—as described above—constrains the lower levels. Hence, such symbol systems may well be regarded as an e... |
To become an element of a symbol emergence system, the element, e.g., a person or a cognitive robot, must have the cognitive capability to form an internal representation system, i.e., PSS, and to perform symbol negotiation. Figure shows an abstract diagram of a partial structure of a cognitive system required for an a... |
). Conversely, it has also been hypothesized that the evolution of syntactic planning capabilities became a scaffold of the evolution of language. Some studies have already shown that structures in an internal representation system analogous to action grammar emerge as distributed representations in neuro-dynamic syste... |
One should likely think of this as an upward-winding spiral, where one process drives the other and vice versa. This knowledge should be self-organized as an internal representation system in a distributed manner. We expect that we can develop neural network-based and/or probabilistic models to artificially realize suc... |
V-C Redefinition of the terminology In this section, we try to identify the differences between terms related to symbols , from the viewpoint of a symbol emergence system, to address possible confusion in future interdisciplinary research. |
V-C Concept and category These two terms are—as mentioned above—often used interchangeably across the literature, even in cognitive science. Here, we would like to offer a practical solution to disentangle this. |
Category: A category corresponds to a referent, i.e. an object in semiosis, where exemplars of the object are invoked by a perception or a thought process. Categories are influenced by contextual information. Categories can be formed and inferred by an agent from sensorimotor information, i.e., features, alone without ... |
. A multimodal autoencoder can also perform a similar task. In order to make sense for an agent that acts in some world, categories need to be based on real-world experiences. This can either happen directly through forming a category from sensorimotor inputs, or indirectly through inferring a category from entities in... |
Categories can be memorized in long-term memory, but importantly, they are not static. Instead they update themselves in long-term memory by the interactions between actual objects, instances, relations, or events in the world. We have many relatively static categories in our memory system and society (such as fruit ca... |
In summary, categories are very much an exemplar-based idea, and always have an inference process using category-determining features that make an object a member of the category. This process can be regarded as semiosis. |
Concept A concept, on the contrary, is usually regarded as a context-independent internal representation embedded in long-term memory, as opposed to a category which is an exemplar-based idea. When we follow the idea of PSS, cross/multimodal simulation is a crucial function of a concept, i.e., perceptual symbol. Here, ... |
Difference between semantic essence of exemplars, and categories can also be understood by means of an example. Let us consider the category of all possible Go games, determined by their shared features. As we know, AlphaGo |
can by now beat the best Go players in the world. Within AlphaGo the representation of how to play a game is encoded in the network weights. Neither AlphaGo itself nor anyone else can extract from these weights and explain the goal of Go and the game strategiesr required to achieve it. Goal and strategy constitute in t... |
In summary, the concept refers to an internal representation which goes beyond the feature-based representation of a category. To form a concept of a category an agent needs to be able to extract the semantic essence of the category, and to connect category and semiosis. At the same time, the concept conversely needs t... |
We believe that the distinction offered here between category and concept captures many of the aspects of these two entities discussed above. These definitions can be phrased in computer (or network) terms which should be useful for our purpose of helping to develop improved cognitive developmental systems. |
V-C Feature representations For performing recognition and other tasks, a cognitive system needs to perform a hierarchical feature extraction and obtain various feature representations. Raw data, e.g., sound waves and images, do not represent a category or a concept on their own. Let us consider image classification as... |
that some high-level features correspond to meaningful entities, e.g., a human face or a human body. Thus, high-level features represent more complex information than low-level features or raw signals. However, it goes too far to assume that such high-level features are categories or concepts, because so far all of the... |
V-C Internal representation system The term ``internal representation'' has a very broad meaning. In a cognitive model, any activation pattern of a neural network and any posterior probability distribution in a probabilistic model can be regarded as an internal representation in a broad sense. Note that concepts encode... |
However, from the viewpoint of this review, we would posit that the term internal representation system refers to a structured system that is formed by interacting with symbol systems in the society and somehow internalizing it. This retains categories and concepts. It also retains knowledge of syntax, semantics and pr... |
This is organized by multimodal perceptual information obtained from its sensorimotor input and from the interaction with the society in which the agent is embedded. As discussed above, these bottom-up principles of organization let an internal representation system emerge . Such an internal representation system can b... |
V-C Emergent symbol system When we open a dictionary, we can find many words representing concepts and categories and their relationships. The meaning of each word is described by some sentences, i.e., syntactically composed words. In classical AI, they attempted to give a type of knowledge to robots. That led to amoda... |
Through negotiation between cognitive agents, the standard meanings of words and sentences are gradually organized in a society or a community. Note that the symbol system itself dynamically changes over time in a small community and in the long term. However, it looks fixed in a larger community and over the short ter... |
We argue that the symbol system in our society has an emergent property. The symbol system is formed through interaction between cognitive systems that have capability of bottom-up formation of internal representation systems. When the cognitive systems try to communicate with others, they have to follow the symbol sys... |
V-C Other related terms In this section, we mention several terms that also tend to be confused with symbols or symbol systems, although they are clearly different. |
Discrete state : As discussed above, symbols have strong connections with categories and concepts. These are by construction somehow discrete in the sense that every category/concept has a boundary to other categories or concepts. This definition goes beyond the token-like, more static discreteness assumed by amodal sy... |
Word A word is the smallest element that is written or uttered in isolation with pragmatic and semantic content in linguistics. It also becomes a syntactic element in a sentence, and it consists of a sequence of phonemes or letters. |
In an ASS (or even in formal logic), we tend to misunderstand this term to the degree that we would naively think that a word, e.g., ``apple,'' itself conveys its own meaning. However, as studies following the symbol grounding problem suggested, to interpret the meaning of a word, the system needs to ground the word by... |
Moreover, the arbitrariness of labels and concepts suggested in semiotics should be considered. The word representing an apple, concept and category of the word ``apple'' can be different in different languages and regional communities. Note that a word is just an observed sign. It itself it is not a concept or a categ... |
Language Human beings are the only animals that can use complex language, even though robots might be able to use some type of reduced language in the near future. Generally, symbol systems do not need to have all properties of a language. Language involves syntax, semantics, pragmatics, phonology, and morphology. Conv... |
VI Challenges VI-A Computational models for symbol emergence and cognitive architecture Developing a computational model that realizes symbol emergence in cognitive developmental systems is a crucial challenge. Many types of cognitive architectures, e.g., ACT-R, SOAR, and Clarion, have been proposed to give algorithmic... |
They described possible psychological process in interaction between symbolic and sensorimotor computations. However, their adaptability and developmental nature are limited. We need a cognitive architecture that enables robots to automatically learn behaviors and language from a sensorimotor system to allow them to pl... |
Tani et al. have been studying computational models for internal representation systems based on self-organization in neuro-dynamic systems |
. Their pioneering studies have shown that robots can form internal representations through sensorimotor interaction with their environment and linguistic interaction with human instructors. The central idea is that continuous neural computation and predictive coding of sensorimotor information is crucial for internal ... |
Recent studies on artificial general intelligence based on DRL follow a similar idea. Hermann et al. showed that DRL architectures can integrate visual, linguistic, and locational information and perform appropriate actions by interpreting commands |
Taniguchi et al. have been using probabilistic generative models (i.e., Bayesian models) as computational models of cognitive systems for symbol emergence in robotics |
. It is occasionally argued that their models assume discrete symbol systems in a similar way as ASSs. However, this involves misunderstanding. A cognitive system designed by probabilistic generative models has its internal states as continuous probability distributions over discrete nodes in the same way as a neural n... |
Neural networks can be trained using stochastic gradient descent and related methods, i.e., backpropagation, and this allows a system to have many types of network structure and activation functions. Recently, neural networks with external memory are gaining attention |
Therefore, if we have a large amount of data, a neural network can organize feature representations adaptively through supervised learning. This is an advantage of neural networks. However, in an unsupervised learning setting, it is still hard to introduce a specific structure in latent variables. In contrast, probabil... |
. This has been preventing their models from having feature representation learning capabilities. However, it should be emphasized that the employment of probabilistic generative models or neural networks, e.g., neuro-dynamics systems by Tani, is not a binary choice. For example, Kingma et al. introduced autoencoding v... |
. They proposed variational autoencoder (VAE) as an example of this idea. By integrating VAE and probabilistic generative models, e.g., GMM and HMM, Johnson et al. proposed a probabilistic generative model with VAEs as emission distributions |
. Employing a neural network as a part of the probabilistic generative model and as a part of inference procedure, i.e., an inference network, is now broadening the possibility of applications of probabilistic generative models and neural networks. Edward, a probabilistic programming environment developed by Tran et al... |
, has already been merged into TensorFlow, which is the most popular deep learning framework. Finding an appropriate way to integrate probabilistic generative models and neuro-dynamics models is crucial for developing computational cognitive architecture modeling symbol emergence in cognitive developmental systems. |
Artificial systems involving symbol emergence need to have many cognitive components, e.g., visual and speech recognition, motion planning, word discovery, speech synthesis, and logical inference. All of those must be adaptive and learn through interaction with the system's environment in a lifelong manner. In addition... |
. Computational models and frameworks to develop a large-scale cognitive architecture that can work practically in the real-world environment is our challenge. |
VI-B Robotic planning with grounded symbols Since the early days of AI, symbolic planning techniques have been employed to allow agents to achieve complex tasks in closed and deterministic worlds. As discussed previously, one problem in applying such symbolic reasoning to robotics is symbol grounding |
, i.e., mapping symbols to syntactic structures of concrete objects, paths, and events. In robotics, this problem has been rephrased using the term symbol anchoring, which puts stronger emphasis on the linking of a symbol (in symbolic AI) to real-world representations acquired with robot sensing |
. This is a particularly challenging task when the process takes place over time in a dynamic environment. A number of works in the literature have attempted to solve this problem by relying on the concept of affordances |
: action possibilities that are perceived in the objects. The perception of object affordances relies on an action-centric representation of the object, which is dependent on the sensorimotor abilities of the agent and that is learned from experience: if a robot learns how to perceive affordances based on its own visuo... |
In the AI and robotics literature, comprehensive cognitive architectures for representation, planning, and control in robots have been proposed, involving multiple components addressing different requirements (e.g., sensor fusion, inference, execution monitoring, and manipulation under uncertainty). However, there is c... |
. Examples of such comprehensive systems , while presenting solid theoretical foundations in behavior-based control and robot simulation results, still lack robust and general applicability (e.g., not being restricted to one specific task) on real robot platforms, such as humanoids. |
The above discussion shows that one major challenge arises from having to design systems that can be scaled up to complex problems, while all their components need to allow for grounding. At the same time, such systems must cope with the nasty contingencies and the noise of their real sensorimotor world. Recently, hybr... |
: whereas symbolic planning offers good generalization capabilities and the possibility to scale, its static and deterministic set of rules does not cope well with real uncertain environments where robots operate. Conversely, tight perception–action loops that allow robots to perform simple tasks effectively in the rea... |
From the viewpoint of symbol emergence system, action planning capability and syntax should also be organized in a bottom-up manner (Fig. ). Structural bootstrapping had been introduced as one possible model |
, in which the structural similarity of a known to an unknown action is used as a scaffold to make inferences of how to ``do it with the new one." Additionally, syntactic bootstrapping also becomes a key to accelerate learning action planning and language. If the internal representation of planning and syntax are share... |
This relates to learning hierarchies of action concepts as well. When forming action concepts and symbols from sensorimotor experience, what should be their scale or granularity? If internal representations are very fine-grained ( bend a finger ), they are easy for a robot to learn but hardly represent useful cause–eff... |
Ugur and Piater illustrated this principle in the context of a robot that learns to build towers from experience by child-like play. The central idea behind this work was to allow a robot to first (pre-trial) learn basic concepts asking: ``How will individual tower-building blocks respond to actions?'' In doing so, the... |
VI-C Language acquisition by a robot Language acquisition is one of the biggest mysteries of the human cognitive developmental process. Asada et al. described how cognitive developmental robotics aims to provide new understanding of human high-level cognitive functions, and here specifically, also the language faculty,... |
Along these lines, an integrative model that can simultaneously explain the process of language acquisition and the use of the language obtained is still missing. To understand the process of language acquisition, we need to develop a computational model that can capture the dynamic process of the language-learning cap... |
A series of non-developmental studies using deep and recurrent neural networks made significant progress in speech recognition and natural language processing, e.g., machine translation, image captioning, word embedding, and visual question answering, using a large amount of labeled data to create off-the-shelf pattern... |
Recent advances in modern AI and robotics have brought the possibility of creating embodied computational intelligence that behaves adaptively in a real-world environment. Creating a robot that can learn language from its own sensorimotor experience alone is one of our challenges, which is an essential element for the ... |
and probabilistic models There is quite lengthy list of subchallenges that have to be addressed before we can reach this goal. The following enumeration is incomplete. |
1. A developmental robot should acquire phonemes and words from speech signals directly without transcribed data. Some attempts toward this exist. For example, Taniguchi and coworkers proposed a hierarchical Dirichlet process-based hidden language model, which is a nonparametric Bayesian generative model, and showed th... |
2. Learning the meaning of words and phrases based on multimodal concept formation is another subchallenge. The interplay between different modalities is crucially important for this. For example, the learning process of the meaning of a verb, e.g., ``hit'' or ``throw,'' must depend on the learning process of the motio... |
3. We also need a machine learning method for syntax learning. Reliable unsupervised syntax learning methods are still missing. Many scientists believe that the syntactic structure in action planning can be used for bootstrapping language syntax learning. |
4. Furthermore, we need unsupervised machine learning methods for the estimation of speech acts and meaning of function words, e.g., prepositions, determiners, and pronouns, as well as metaphors without artificially prepared labeled data. |
5. The learning of nonlinguistic symbol systems that are used with linguistic expressions, e.g., gestures, gaze, and pointing, is also important. They might even have to be learned as prerequisites of language learning |
All these subchallenges, in addition, need to deal with the uncertainty and noise in speech signals and observed (sensorimotor) events. |
VI-D Human–robot communication, situated symbols, and mutual beliefs The existence of shared symbol systems is clearly fundamental in human–robot collaboration. Cognitive artificial systems that incorporate the translation of natural-language instructions into robot knowledge and actions have been proposed in many stud... |
and toward statistical symbol grounding In order to support human activities in everyday life, robots should adapt their behavior according to the situations, which differ from user to user. In order to realize such adaptation, these robots should have the ability to share experiences with humans in the physical world.... |
. From this viewpoint, language systems can be considered as shared belief systems that include not only linguistic but also nonlinguistic shared beliefs, where the latter are used to convey meanings based on their relevance for the linguistic beliefs. |
However, the language models existing in robotics so far are characterized by fixed linguistic knowledge and do not make this possible |
. In these methods, information is represented and processed by symbols, the meaning of which have been predefined by the robots' developers. Therefore, experiences shared by a user and a robot do not exist and can, thus, neither be expressed nor interpreted. As a result, users and robots fail to interact in a way that... |
To overcome this problem and to achieve natural dialogue between humans and robots, we should use methods that make it possible for humans and robots to share symbols and beliefs based on shared experiences. To form such shared beliefs, the robot should possess a mechanism, related to mechanisms of the theory of mind ,... |
presented a language acquisition method that begins to allow such inference and makes it possible for the human and the robot to share at least some beliefs for multimodal communication. |
VII Conclusion AI, cognitive science, neuroscience, developmental psychology, and machine learning have been discussing symbols in different contexts, but usually sharing similar ideas to a certain extent. However, these discussions in the different fields have been causing various types of confusion. A central problem... |
Machine learning-based AI, including deep learning, has recently been considered as a central topic in the AI community. In this context, it is often said that the notion of a ``symbol system'' is out of date. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.