Seminár z kognitívnej vedy a umelej inteligencie je (od zimného semestra 2015-2016) pokračovaním spoločného seminára z umelej inteligencie, ktorý organizovali Ústav aplikovanej informatiky FIIT STU (prof. Kvasnička) a Katedra aplikovanej informatiky FMFI UK (prof. Farkaš, a pred ním doc. Šefránek). V zimnom semestri je seminár určený hlavne pre študentov kognitívnej vedy, aby mali možnosť zorientovať sa v existujúcom výskume v našom okolí, v letnom semestri má rozšírený záber aj na umelú inteligenciu. Po celý rok sú však vítaní všetci záujemcovia, ktorí sú chcú vypočuť ponúkané prednášky.
Organizátorom seminára je prof. Igor Farkaš.
Čas a miesto konania seminára: utorok 16:30-17:45 v miestnosti I-9 (pavilón informatiky) na FMFI UK.
Mental imagery is a phenomenon that has been given various theoretical accounts in cognitive science. We present an enactive approach to visuospatial mental imagery, implemented in a hybrid computational model, using the so-called perceptual actions. The model consists of a forward model, an inverse model, both implemented as neural networks, and a memory/controller module, that grounds simple mental concepts, such as a triangle and a square, in perceptual actions, and is able to reimagine these objects by performing the necessary perceptual actions in a simulated humanoid robot. We tested the model on three tasks – salience-based object recognition, imagination-based object recognition and object imagination – and achieved very good results showing, as a proof of concept, that perceptual actions are a viable candidate for grounding the visuospatial mental concepts as well as the credible substrate of visuospatial mental imagery. (The core of this work was done with students of cognitive science - Jan Jug, Tine Kolenik and André Ofner - during their mobility semester).
Standard error backpropagation is still the most prominent algorithm for supervised training of artificial neural networks, although it has been claimed to be biologically implausible. Algorithms based on local activation differences such as GeneRec (O'Reilly, 1996) were designed as an alternative to error propagation. In continuation with our previous research (Malinovská/Rebrová and Farkaš, 2013), we present a model based on GeneRec that learns heteroassociative mappings. We show how our Universal Bidirectional Activation-based Learning algorithm can learn various other tasks if different parameters are used and that its performance in canonical neural network tasks, such as 4-2-4 encoder or XOR as well as in machine learning benchmarks such as the MNIST dataset, is comparable with a standard error backpropagation model.
Since the mid 1970s connections between music and language have been widely studied in cognitive science. The first part of this talk will present the main directions such studies have taken in recent years (structural, neurobiological, evolutionary, and semantic). In particular, I will discuss major breakthroughs from linguistics that made an impact on music cognition and then reflected back on the study of language. In the second part, I will provide a precis of four recent studies by my group. The first of these looks into the relative weights of perceptual clues for constructing well-formed metrical and melodic patterns in music; the second suggests that seemingly disparate cross-linguistic conceptualizations of music-theoretic constructs, such as scales ''moving upward'' or ''getting thinner'', do not necessarily support the strong case for linguistic relativity, but may rather emerge from more universal, higher-order schematic invariants; the third study revisits the referential power of program music, suggesting that in constructing musical meaning participants are highly sensitive to contextual priming; the fourth segment integrates such results into a larger-scale theoretical program on the nature of musical meaning, vouching for ''multi-level grounding'' in (musical) semantics. Altogether, I hope to suggest that some constructs originally defined in linguistics and then taken over by music cognition can help shed additional light on some major dilemmas of cognitive science in general.
Už dnes jazdí po našich cestách veľa vozidiel s rôznymi elektronickými systémami, ktoré informujú vodiča o hroziacom nebezpečenstve a preberajú riadenie, keď vodič nekoná. Postupne pribúdajú aj vozidlá s umelou inteligenciou, ktorej môže vodič úmyselne odovzdať riadenie a následne sa venovať inej činnosti. Na rozhodovanie vozidlá vyplýva nielen znalosť ovládania vozidla, znalosť dopravných predpisov a dopravných značiek, ale aj nešpecifikované okolnosti, ktoré musí vozidlo zvládnuť aspoň na úrovni ako priemerný vodič človek.
We investigate the role of perceptual similarity in visual metaphor comprehension process. In visual metaphors, perceptual features of the source and the target are objectively present as images. Moreover, to determine perceptual similarity, we use an image-based search system that computes similarity based on low-level perceptual features. We hypothesize that perceptual similarity at the level of color, shape, texture, orientation, and the like, between the source and the target image facilitates metaphorical comprehension and aids creative interpretation. We present three experiments, two of which are eye-movement studies, to demonstrate that in the interpretation and generation of visual metaphors, perceptual similarity between the two images is recognized at a subconscious level, and facilitates the search for creative conceptual associations in terms of emergent features. We argue that the capacity to recognize perceptual similarity—considered to be a hallmark of creativity—plays a major role in the creative understanding of metaphors.
Capturing the movement of a human body is a challenging problem. The surface of the body deforms in a complex way that is hard to be described in a precise way. Therefore, in computer graphics, this complex deformation is usually approximated by the so-called animation skeleton. The skeleton segments the body surface into a set of several rigid areas, and surface deformation of the body is approximated by a rigid transformations of the skeleton. The topic becomes even more challenging if we want to capture not only the movement, but also the point cloud frames of the body, and later reconstruct the whole 3D model of the subject. Since the body can be in different poses each frame, the so-called non-rigid surface fusion has to be performed. This step consists of transforming each point cloud into a normalized pose, where the individual surface sheets need to be stitched. Finally, we would like to reconstruct the 3D model of the subject and transform it into a desired pose. Having a single camera only, a visible surface area is limited, and thus the deformations on the back faced body parts need to be approximated. This can be be either computed using physical simulation that is very time and power consuming, or approximated using machine learning and statistical models.
Artificial neural networks recently became the most successful approach in machine learning with many successful applications in various domains such as image recognition, machine translation, reinforcement learning or generative modeling. The majority of artificial neural network models are currently trained using the gradient descent method with an error backpropagation mechanism. However, the main disadvantage of these models is the required large number of training examples for achieving reasonable performance. This limits the applications to domains offering abundance of labelled training examples or, in the case of reinforcement learning, the large number of interactions with an environment makes it possible to learn the task in that environment. In this talk, we will explore the few-shot learning approach, requiring only a few labeled examples and review various few-shot learning methods based on artificial neural networks. We will also propose a novel approach to the few-shot learning approach called Categorical Siamese neural networks.