This paper introduces an approach to ‘meaning' from a viewpoint of Hallidayan lin- guistics, that is, systemic functional linguistics (SFL).It explains the basic idea of SFL, the comparison between Hallidayan and Chomskyan linguistics in terms of an approach to meaning, and the relation between SFL and Wittgenstein's philosophy.
Despite recent artificial intelligence technologies achieved prominent results, machines cannot behave like humans yet. I compare the ability of humans’ language acquisition
with natural language processing by machine learning technologies, taking learning of
word embedding vectors as an example. And I introduce some notable researches which
may fill in the gap, one-shot learnig, memory models and language game tasks. Al- though these researches have just begun, it will be important for artificial intelligence to aquire human intelligence.
In this study, we define concepts as categories into which a robot classifies perceptual information obtained through interaction with others and the environment, and the inference of unobserved information through the concepts is defined as understanding. Furthermore, a robot can infer unobserved perceptual information from words by con- necting concepts and words. This inference is the understanding of word meanings. We propose probabilistic models that enable robots to learn concepts and language. In this paper, we present an overview of the proposed models.
Automatic creation of concepts is important for various situations. Previous re- searches in the conceptual blending and the concept invention proposed the cognitive models which represent the process by which people combine concepts and those rela- tionships. However, those researches do not allow one to create new concepts automat- ically in the real world, where there are innumerable notions and the meanings of them are time-varying. Because the previous models can not discover which notions should be combined to create successful concepts, it is necessary for a user to find an appro- priate combination of notions. There are approximately 50 million combinations in the business domain. Therefore, we propose a novel model representing concept creation processes, which makes automatic creation of new and successful concepts possible even in such a real world setting. We formalize the concept creation process as discovering new connections between existing concepts and it can be mathematically represented using the chronological change of the semantic networks. The data of the input and output of this process can be built using a large document set. Hence, machine learning technique can reveal a law underlying the concept creation process. After extracting such a law, the machine learning model can provide new concepts in accordance with its law. In experiments, we evaluated the validity of this approach using real successful concepts and document sets, and created new concepts in food category.
Young children produce multi-word sentences including some systematic errors or overproduction. It has been reported that English-speaking children may add a mor- pheme “ed” to an irregular verb as its past tense while Japanese-speaking children may position a case particle “NO” after an adjective. We hypothesize that an insufficient in- crease in grammatical categories causes such overproduction, which can be expected to disappear with a sufficient increase.We assume that hidden states of a hidden Markov model (HMM) correspond to grammatical categories acquired from language input. Based on the HMM, the simulation results could partially verify the above hypothesis. In the English-trained model, the overproduction could appear and then decline. How- ever, it did not completely disappear because categories of regular and irregular verbs did not differentiate even when the model had many categories. In the Japanese-trained model, the overproduction could appear and then disappear through differentiation of categories of nouns and adjectives. The limitations of the proposed model are pointed out and future issues are discussed.
One of major paradigms in cognitive science is to model cognitive process as an in- formation processing in the digital computer. Marr (1982) has proposed to capture the cognitive process by the three levels of information processing, known as the levels of hardware implementation, algorithm and representation, and computational theory. In particular, the “computational theory” level is supposed most important among the three, as it captures the goal of the information process and explains why the process is organized so. It is, however, often controversial what to count as the computational theory, and there are several variations in its interpretation. In this article, we review these views on the computational theory, and overview the potential problems of the computational theory in a narrow sense which have been pointed out in past literature. By doing so, we discuss the aspects of the current paradigm to be extended toward a new alternative paradigm beyond the formulation of cognition as optimization.
Causal knowledge enables us to explain past events, to control present environment, and to predict future outcomes. Over the last decade, causal Bayes nets have been rec- ognized as a normative framework for causality and used as a psychological model to account for human causal learning and inference. This article provides an introduction to causal Bayes nets. According to causal Bayes nets, causal inference can be divided into three processes: (a) learning the structure of the causal network, (b) learning the strength of the causal relations, and (c) inferring the effect from the cause or the cause from the effect. For each process, I describe the predictions of causal Bayes nets, review experimental results, and suggest future directions. Although there are a few excep- tions (e.g., Markov violation), most of the results are consistent with the predictions of causal Bayes nets. The current problems of the Bayesian approach and its future perspective are discussed.
It is propounded that in order to avoid the “frame problem” or “symbol grounding problem” and to create a way to analyze and realize human-like intelligence with higher functions, it is not enough just to introduce deep learning, but it is significant to get out of deeply penetrated “division into functional modules” and to take the approach of “function emergence through end-to-end reinforcement learning.” The functions that have been shown to emerge according to this approach in past works are summarized, and the reason for the difficulty in the emergence of thinking that is a typical higher function is made clear. It is claimed that the proposed hypothesis that exploration grows towards think- ing through learning, becomes a key to break through the difficulty. To realize that, “reinforcement learning using a chaotic neural network” in which adding external ex- ploration noises is not necessary is introduced. It is shown that a robot with two wheels and a simple visual sensor can learn an obstacle avoidance task by using this new reinforcement learning method.
The present study aimed to investigate the association between changes in the mod ular structures of brain functional connectivity networks (BFCNs) and individual vari- ability in foreign language learning ability. Six healthy Japanese students (all male, age range: 16–21 years) completed both pre- and post-training EEG sessions and ex- aminations in English words. BFCNs were constructed for pre- and post-training data. Training sessions required participants to attempt to memorize 200 pairs of the same English words as those in the examinations and their Japanese meanings. The con- nectivity between any two different electrodes (nodes) was calculated by determining the synchronization likelihood (SL) of the EEGs. An edge connecting the two nodes was drawn when statistically significant differences in SL values were observed between successful and unsuccessful trials. BFCNs for the pre-training data were higher in edge density than those for the post-training data, though this difference was not statisti- cally significant. Moreover, changes in the composition of modules of the BFCNs were associated with the individual difference between two English-word examinations in scores.