Ordinary neural networks represent objects in a dimensionally compressed latent (Z) space. In this latent space, a grammatical structure emerges if the input is a language, or a finite state automaton that makes predictions if the input is a time series. But the living system perceives and experiences the object without contracting it. The philosopher Meillassoux said, “A discontinuous ring consisting of multiple interruptions.” While discussing the interpretation of this idea, I will analyze and report on an experiment using the android Alter3, which mimics human poses. What is important to make a robot anti-contractive is one's interaction with a human being, and the other is the autonomous rewriting of memories that is created because of its inability to learn.
Inspired by Q. Meillassoux's philosophy, Ikegami (2021) claims that a living system perceives external reality without ‘contracting' the overwhelming flow of information it contains. This idea implies that cognition is not merely the formation of representations inside a living system. Based on this idea, we can reconsider the relationship between the inside and outside of a living system: I claim that this relationship contains both disconnection and connection at the same time. Ikegami assumes that the contracting movements that occur in living systems are continuously disrupted by encounters with external reality. Furthermore, he claims that living systems ‘vividly experience' this disruption involving an overwhelming flow of information. This idea does not seem to be present in Meillassoux's philosophy.
Ikegami (2021) proposes to reinstate excessiveness as a critical subject in cognitive science. In this commentary, I will discuss what this excessiveness is, taking a hint from the philosophies of Merleau-Ponty and Henry. When we create representations from bodily information, it is impossible to make representations of all the information we receive, and what leaks out here should be called “excessiveness”. In this context, the possibility of being able to handle phenomena that are difficult to verbalize, such as qualia and affectivity, emerges. Furthermore, in relation to deep learning, I would like to discuss contraction and expansion of the representation. Autoencoder is a technology for compressing data while maintaining the original information, and can be divided into an encoder part that compresses the input into latent variables, and a decoder (generator) part that restores the original information by appropriately expanding the latent variables. Generative deep learning is an extension of the generator, which can reconstruct information from appropriate latent variables. I would like to consider how this generative deep learning can regenerate excessiveness from contracted representation and discuss the relationship between the generator and cognitive projection.
Ikegami considers that cognitive systems are constantly exposed to excess observational data, from which it acquires the subtractive representations rather than the contractions that deep learning aims for. Furthermore, he argues that a glimpse of life as the basic principle of cognition can be seen in the moment when the contraction collapses. The current commentary discusses these points by making possible correspondences to the cognitive robotics studies conducted by the author and his colleagues.
Research on Alife, as described by Ikegami, is compared with research on AI, as described by Nakashima. Similarity:They are both based on constructive methodology. They both treats interaction with the environment. Ikegami forms his claim around Meillassoux' s subtracton and contraction. However, I claim the direction of the description is wrong. I favor Uexküll' s concept of Umwelt, which describes the same phenomena from the opposite direction. Difference: ALife seeks for life; AI seeks for intelligence. As the result, ALife creates entities with System 1 intelligence (as described by Kahneman), while AI creates programs with System 2 intelligence. When we talk about life, its survival is essential, which is provided by System 1.
Recent advances in cognitive measurements and machine learning are influencing our way of doing science. Without forming hypothesis, just setting up comprehensive measurements and sending data to machine learning is becoming a new standard. Some even suggest interpretation of data should also be asked to an artificial intelligence. However, I am wondering whether this really is work for scientists. Our work should be to establish a system of explanation accessible for humans. In this light, I am discussing the relationship between compression, encompassing, and subtraction.
Cognitive linguistics views meaning as a central part of language and takes a maximalist, non-reductionist, and bottom-up approach to language. Not by cutting off or contracting the excessive flow of information, but by embracing it, cognitive linguists propose a linguistic theory grounded in the general cognitive capacities of humans. The important point is that this theory can capture the nature of language that keeps changing itself to some extent. We can view this as a mechanism of active generation in language. Based on our studies of neologism in natural and artificial languages, I will introduce aspects of language that can be related to active generation in language. I argue that, through interaction with humans, language can stay fresh and generative.
I discuss the relevance of Massive Data Flow, Valera' s autopoiesis and Maillassoux' s subtraction and contraction as argued in the target paper, citing the brilliant six commentators. I argue that it is important to think about MDF, ALife and automated scientists to remove human cognitive biases from cognitive science.
How does an artist utilize their ecological constraints when creating art? The present case study describes the making process of Rinsho, or imitation drawing, by a professional Chinese calligrapher through 16 experimental trials. To examine the process through which ecological constraints were exploited and utilized, I recorded the calligrapher' s movement data using a 3D motion capture system. My analysis focused on the calligrapher' s gaze search, specifically head height, the number and frequency of head rotation during writing, and the directional distribution of head movement. The results show that the calligrapher performed a flexible gaze search in response to surrounding constraints, while his drawing procedure remained relatively consistent. The calligrapher combined fine brush-tip control and active gaze search with whole-body movement. In addition, the comparison between the present study and the author's previous study suggests that the calligrapher generated his own constraints in movement variables, and explored white space and character shapes on the paper by utilizing these constraints. These results imply that Rinsho is the art of pursuing resourcefulness under a given set of constraints. The present paper proposes that the method to describe the time evolution of an environment-body system is an effective approach for studying artists' making processes.
Foreign language side effect (FoLSE) refers to a temporary decline of thinking ability while nonproficient foreign language is being used. This decline is produced by stronger interference between thinking and a heavier cognitive load of foreign language processing. Although FoLSE was shown to occur in laboratories, it may not occur in daily verbal communication when thinking is accompanied by inner language. The reason is as follows: In general, the more similar two concurrent cognitive tasks, the stronger their mutual interference. Inner language is usually experienced as effortless native language. When outer language used in verbal communication is native language, it is more similar to inner native language. Therefore, outer native language is expected to produce stronger interference with inner native language involved in thinking, and thus could produce a larger reduction in thinking performance. This larger reduction due to outer native language may cancel out the reduction due to outer foreign language (FoLSE). To examine this possibility, Japanese college students performed verbal and thinking tasks concurrently in two dual-task experiments. The verbal task was presented in either Japanese (native language) or English (foreign language). The thinking task was always presented in Japanese (native language). Past empirical studies strongly suggested that inner language should be evoked in the employed thinking tasks (i.e., validity judgment on categorical syllogism and intelligence test problems that loaded heavily on verbal factors of intelligence). The results revealed that performance in the thinking task was lower when the verbal task was presented in the foreign language. This means that FoLSE was stronger than the interference between the inner and outer native language. It follows that FoLSE is likely to occur in daily verbal communication as well even when it is accompanied by inner language.
Deep Learning achieved significant progresses in building intelligent systems. Some state-of-the-art Deep Learning systems are touted as “human parity” or even “super human”. At the same time, it is often criticized for the lack of understanding of internal workings especially when such systems are considered for real world applications. Cognitive Science, on the other hand, is a scientific discipline aiming to understand internal workings of human-like intelligent behaviours. Thus a question arises: Is Deep Learning a subject of Cognitive Science? This article discusses what can be a subject of Cognitive Science and whether Deep Learning qualifies as such. It is also discussed what kind of connections can be established between these two disciplines even if Deep Learning is not be a direct subject of Cognitive Science.
We consider an intelligence as computation and classify it by the way its specifications are given. Based on this classification, we argue that it is difficult to design a human-like intelligence. Then, we discuss the possibility of non-human-like intelligence. We also touch on the necessity of the study of human intelligence from the social point of view.
This article tries to position deep learning in the intersection of artificial intelligence and cognitive science, as a long quest toward human intelligence. First, the recent development of huge language models obtained by transformer-based methods such as BERT and GPT-3 is introduced. Then, I explain what these models can do and can not do, and why. Two essential problems, which is embodiment and symbol grounding, are shown. In order to solve these problems, deep reinforcement learning with world models are currently studied. Disentanglement is shown to be an important concept to find factors to control. Lastly, I explain my perspective toward the future advancement, and conclude the paper.