No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
A neural network model of curiosity-driven infant categorization
Abstract
Infants are curious learners who drive their own cognitive development by imposing structure on their learning environments as they explore. Understanding the mechanisms underlying this curiosity is therefore critical to our understanding of development. However, very few studies have examined the role of curiosity in infants' learning, and in particular, their categorization; what structure infants impose on their own environment and how this affects learning is therefore unclear. The results of studies in which the learning environment is structured a priori are contradictory: while some suggest that complexity optimizes learning, others suggest that minimal complexity is optimal, and still others report a Goldilocks effect by which intermediate difficulty is best. We used an auto-encoder network to capture empirical data in which 10-month-old infants' categorization was supported by maximal complexity [1]. When we allowed the same model to choose stimulus sequences based on a " curiosity " metric which took into account the model's internal states as well as stimulus features, categorization was better than selection based solely on stimulus characteristics. The sequences of stimuli chosen by the model in the curiosity condition showed a Goldilocks effect with intermediate complexity. This study provides the first computational investigation of curiosity-based categorization, and points to the importance characterizing development as emerging from the relationship between the learner and its environment.
... In [10] we proposed ICAC, an actor-critic RL algorithm that learns an optimal action policy by learning an ensemble of local predictive models of environmental dynamics online and generates an intrinsic reward based on the learning progress of each model. Using the learning progress of predictive models for selecting explorative actions is in accordance with the Goldilocks effect in infant cognition that attributes optimal learning to stimuli of an intermediate degree of difficulty [12]. This is evident in how infants seek increasingly complex learning samples by self-organizing their interaction with the world, moving from well-explored regions to others where they expect to learn new effects of motor activity. ...
... The results presented in this paper prove that the exploration strategy of ICAC driven by the intrinsic reward is effective for finding optimal policies in a reward-sparse world. This is in line with theories from developmental psychology that stress the importance of intrinsic reward in complex environments [10] [12]. ...
In this paper, we present a new visually guided exploration approach for autonomous learning of visuomotor skills. Our approach uses hierarchical Slow Feature Analysis for unsupervised learning of efficient state representation and an Intrinsically motivated Continuous Actor-Critic learner for neuro-optimal control. The system learns online an ensemble of local forward models and generates an intrinsic reward based on the learning progress of each learned forward model. Combined with the external reward, the intrinsic reward guides the system’s exploration strategy. We evaluate the approach for the task of learning to reach an object using raw pixel data in a realistic robot simulator. The results show that the control policies learned with our approach are significantly better both in terms of length and average reward than those learned with any of the baseline algorithms.
... This is also supported by recent studies from developmental psychology that have found infants' curious exploration to be based on both their own learning history and perceptual variability. These studies show evidence of a Goldilocks effect where seeking just the right level of difficulty drives optimal learning with too much or too little difficulty being disruptive [17], [18]. ...
Guiding the action selection mechanism of an autonomous agent for learning control behaviors is a crucial issue in reinforcement learning. While classical approaches to reinforcement learning seem to be deeply dependent on external feedback, intrinsically motivated approaches are more natural and follow the principles of infant sensorimotor development. In this work, we investigate the role of incremental learning of predictive models in generating curiosity, an intrinsic motivation , for directing the agent's choice of action and propose a curiosity-driven reinforcement learning algorithm for continuous motor control. Our algorithm builds an internal representation of the state space that handles the computation of curiosity signals using the learned predictive models and extends the Continuous-Actor-Critic-Learning-Automaton to use extrinsic and intrinsic feedback. Evaluation of our algorithm on simple and complex robotic control tasks shows a significant performance gain for the intrinsically motivated goal reaching agent compared to agents that are only motivated by extrinsic rewards.
Curiosity has a long history of research and rich definitions and classifications as a common mental state and personality trait. The division and coordination of multiple brain regions enable individuals to form a cognitive process of generating and evaluating prediction error, triggering and mediating curiosity, and producing surprise and new prediction error, so as to reduce the prediction error and information gap between internal states and external environment, and eliminate uncertainty. Curiosity has a significant role in improving cognitive function and maintaining mental and physical health during development. Future research can be further considered from a cross-species, interdisciplinary, and multi-domain perspective to promote the deepening of research topics, the development of research methods, and the application of research results in curiosity.
Open-ended exploration and learning in the real world is a major challenge of developmental robotics. Three properties of real-world sensorimotor spaces provide important conceptual and technical challenges: unlearnability, high dimensionality, and unboundedness. In this chapter, we argue that exploration in such spaces needs to be constrained and guided by several combined developmental mechanisms. While intrinsic motivation, that is, curiosity-driven learning, is a key mechanism to address this challenge, it has to be complemented and integrated with other developmental constraints, in particular: sensorimotor primitives and embodiment, task space representations, maturational processes (i.e., adaptive changes of the embodied sensorimotor apparatus), and social guidance. We illustrate and discuss the potential of such an integration of developmental mechanisms in several robot learning experiments. © 2013 Springer-Verlag Berlin Heidelberg. All rights are reserved.
Infants must learn about many cognitive domains (e.g., language, music) from auditory statistics, yet capacity limits on their cognitive resources restrict the quantity that they can encode. Previous research has established that infants can attend to only a subset of available acoustic input. Yet few previous studies have directly examined infant auditory attention, and none have directly tested theorized mechanisms of attentional selection based on stimulus complexity. This work utilizes model-based behavioral methods that were recently developed to examine visual attention in infants (e.g., Kidd, Piantadosi, & Aslin, 2012). The present results demonstrate that 7- to 8-month-old infants selectively attend to nonsocial auditory stimuli that are intermediately predictable/complex with respect to their current implicit beliefs and expectations. These findings provide evidence of a broad principle of infant attention across modalities and suggest that sound-to-sound transitional statistics heavily influence the allocation of auditory attention in human infants.
Often, when animals encounter an unexpected sensory event, they transition from executing a variety of movements to repeating the movement(s) that may have caused the event. According to a recent theory of action discovery (Redgrave and Gurney, 2006), repetition allows the animal to represent those movements, and the outcome, as an action for later recruitment. The transition from variation to repetition often follows a non-random, structured, pattern. While the structure of the pattern can be explained by sophisticated cognitive mechanisms, simpler mechanisms based on dopaminergic modulation of basal ganglia (BG) activity are thought to underlie action discovery (Redgrave and Gurney, 2006). In this paper we ask the question: can simple BG-mediated mechanisms account for a structured transition from variation to repetition, or are more sophisticated cognitive mechanisms always necessary? To address this question, we present a computational model of BG-mediated biasing of behavior. In our model, unlike most other models of BG function, the BG biases behavior through modulation of cortical response to excitation; many possible movements are represented by the cortical area; and excitation to the cortical area is topographically-organized. We subject the model to simple reaching tasks, inspired by behavioral studies, in which a location to which to reach must be selected. Locations within a target area elicit a reinforcement signal. A structured transition from variation to repetition emerges from simple BG-mediated biasing of cortical response to excitation. We show how the structured pattern influences behavior in simple and complicated tasks. We also present analyses that describe the structured transition from variation to repetition due to BG-mediated biasing and from biasing that would be expected from a type of cognitive biasing, allowing us to compare behavior resulting from these types of biasing and make connections with future behavioral experiments.
Most previous work on artificial curiosity (AC) and intrinsic motivation focuses on basic concepts and theory. Experimental results are generally limited to toy scenarios, such as navigation in a simulated maze, or control of a simple mechanical system with one or two degrees of freedom. To study AC in a more realistic setting, we embody a curious agent in the complex iCub humanoid robot. Our novel reinforcement learning (RL) framework consists of a state-of-the-art, low-level, reactive control layer, which controls the iCub while respecting constraints, and a high-level curious agent, which explores the iCub's state-action space through information gain maximization, learning a world model from experience, controlling the actual iCub hardware in real-time. To the best of our knowledge, this is the first ever embodied, curious agent for real-time motion planning on a humanoid. We demonstrate that it can learn compact Markov models to represent large regions of the iCub's configuration space, and that the iCub explores intelligently, showing interest in its physical constraints as well as in objects it finds in its environment.
From at least two months onwards, infants can form perceptual categories. During the first year of life, object knowledge develops from the ability to represent individual object features to representing correlations between attributes and to integrate information from different sources. At the end of the first year, these representations are shaped by labels, opening the way to conceptual knowledge. Here, we review the development of object knowledge and object categorization over the first year of life. We then present an artificial neural network model that models the transition from early perceptual categorization to categories mediated by labels. The model informs a current debate on the role of labels in object categorization by suggesting that although labels do not act as object features they nevertheless affect perceived similarity of perceptually distinct objects sharing the same label. The model presents the first step of an integrated account from early perceptual categorization to language-based concept learning.
The efficient coding hypothesis posits that sensory systems of animals strive to encode sensory signals efficiently by taking into account the redundancies in them. This principle has been very successful in explaining response properties of visual sensory neurons as adaptations to the statistics of natural images. Recently, we have begun to extend the efficient coding hypothesis to active perception through a form of intrinsically motivated learning: a sensory model learns an efficient code for the sensory signals while a reinforcement learner generates movements of the sense organs to improve the encoding of the signals. To this end, it receives an intrinsically generated reinforcement signal indicating how well the sensory model encodes the data. This approach has been tested in the context of binocular vison, leading to the autonomous development of disparity tuning and vergence control. Here we systematically investigate the robustness of the new approach in the context of a binocular vision system implemented on a robot. Robustness is an important aspect that reflects the ability of the system to deal with unmodeled disturbances or events, such as insults to the system that displace the stereo cameras. To demonstrate the robustness of our method and its ability to self-calibrate, we introduce various perturbations and test if and how the system recovers from them. We find that (1) the system can fully recover from a perturbation that can be compensated through the system's motor degrees of freedom, (2) performance degrades gracefully if the system cannot use its motor degrees of freedom to compensate for the perturbation, and (3) recovery from a perturbation is improved if both the sensory encoding and the behavior policy can adapt to the perturbation. Overall, this work demonstrates that our intrinsically motivated learning approach for efficient coding in active perception gives rise to a self-calibrating perceptual system of high robustness.
We propose that free viewing of natural images in human infants can be understood and analyzed as the product of intrinsically-motivated visual exploration. We examined this idea by first generating five sets of center-of-gaze (COG) image samples, which were derived by presenting a series of natural images to groups of both real observers (i.e., 9-month-olds and adults) and artificial observers (i.e., an image-saliency model, an image-entropy model, and a random-gaze model). In order to assess the sequential learnability of the COG samples, we paired each group of samples with a simple recurrent network, which was trained to reproduce the corresponding sequence of COG samples. We then asked whether an intrinsically-motivated artificial agent would learn to identify the most successful network. In Simulation 1, the agent was rewarded for selecting the observer group and network with the lowest prediction errors, while in Simulation 2 the agent was rewarded for selecting the observer group and network with the largest rate of improvement. Our prediction was that if visual exploration in infants is intrinsically-motivated—and more specifically, the goal of exploration is to learn to produce sequentially-predictable gaze patterns—then the agent would show a preference for the COG samples produced by the infants over the other four observer groups. The results from both simulations supported our prediction. We conclude by highlighting the implications of our approach for understanding visual development in infants, and discussing how the model can be elaborated and improved.
In this brief essay, I seek to demonstrate the significance of exploratory behavior for understanding cognitive development. Historically, organisms were thought to act solely in the service of achieving biologically significant goals, such as satisfying thirst, hunger, and reproductive drives. However, it became apparent that both animals and humans engage in behavior where the adaptive goal is unclear (see Hunt, 1963, 1965). With no obvious external target, this activity is best described as being intrinsically motivated, and often directed toward the unknown and the unexpected (Kagan, 2002). Hence novelty, the discrepancy between what is known and what is discovered, can elicit activity and exploration of the environment.
What is the relevance to developmental process? Attention to novelty plays a seemingly simple role in learning and development, directing the senses toward what is as yet unknown. Yet, research shows that patterns of attention to novelty are not straightforward, particularly during infancy. There is considerable evidence that attention is sometimes biased toward familiarity, rather than novelty. Unlike our understanding of novelty preference, we struggle to understand when and why familiarity preferences occur. Below I briefly review this area of research and illustrate how this basic aspect of learning continues to puzzle developmental psychologists.
A head camera was used to examine the visual correlates of object name learning by toddlers as they played with novel objects and as the parent spontaneously named those objects. The toddlers' learning of the object names was tested after play, and the visual properties of the head camera images during naming events associated with learned and unlearned object names were analyzed. Naming events associated with learning had a clear visual signature, one in which the visual information itself was clean and visual competition among objects was minimized. Moreover, for learned object names, the visual advantage of the named target over competitors was sustained, both before and after the heard name. The findings are discussed in terms of the visual and cognitive processes that may depend on clean sensory input for learning and also on the sensory-motor, cognitive, and social processes that may create these optimal visual moments for learning.
Human infants, like immature members of any species, must be highly selective in sampling information from their environment to learn efficiently. Failure to be selective would waste precious computational resources on material that is already known (too simple) or unknowable (too complex). In two experiments with 7- and 8-month-olds, we measure infants' visual attention to sequences of events varying in complexity, as determined by an ideal learner model. Infants' probability of looking away was greatest on stimulus items whose complexity (negative log probability) according to the model was either very low or very high. These results suggest a principle of infant attention that may have broad applicability: infants implicitly seek to maintain intermediate rates of information absorption and avoid wasting cognitive resources on overly simple or overly complex events.
Intrinsic motivation, centrally involved in spontaneous exploration and curiosity, is a crucial concept in developmental psychology. It has been argued to be a crucial mechanism for open-ended cognitive development in humans, and as such has gathered a growing interest from developmental roboticists in the recent years. The goal of this paper is threefold. First, it provides a synthesis of the different approaches of intrinsic motivation in psychology. Second, by interpreting these approaches in a computational reinforcement learning framework, we argue that they are not operational and even sometimes inconsistent. Third, we set the ground for a systematic operational study of intrinsic motivation by presenting a formal typology of possible computational approaches. This typology is partly based on existing computational models, but also presents new ways of conceptualizing intrinsic motivation. We argue that this kind of computational typology might be useful for opening new avenues for research both in psychology and developmental robotics.
Disentangling bottom-up and top-down processing in adult category learning is notoriously difficult. Studying category learning in infancy provides a simple way of exploring category learning while minimizing the contribution of top-down information. Three- to 4-month-old infants presented with cat or dog images will form a perceptual category representation for cat that excludes dogs and for dog that includes cats. The authors argue that an inclusion relationship in the distribution of features in the images explains the asymmetry. Using computational modeling and behavioral testing, the authors show that the asymmetry can be reversed or removed by using stimulus images that reverse or remove the inclusion relationship. The findings suggest that categorization of nonhuman animal images by young infants is essentially a bottom-up process.
Exploratory activities seem to be intrinsically rewarding for children and crucial for their cognitive development. Can a machine be endowed with such an intrinsic motivation system? This is the question we study in this paper, presenting a number of computational systems that try to capture this drive towards novel or curious situations. After discussing related research coming from developmental psychology, neuroscience, developmental robotics, and active learning, this paper presents the mechanism of Intelligent Adaptive Curiosity, an intrinsic motivation system which pushes a robot towards situations in which it maximizes its learning progress. This drive makes the robot focus on situations which are neither too predictable nor too unpredictable, thus permitting autonomous mental development. The complexity of the robot's activities autonomously increases and complex developmental sequences self-organize without being constructed in a supervised manner. Two experiments are presented illustrating the stage-like organization emerging with this mechanism. In one of them, a physical robot is placed on a baby play mat with objects that it can learn to manipulate. Experimental results show that the robot first spends time in situations which are easy to learn, then shifts its attention progressively to situations of increasing difficulty, avoiding situations in which nothing can be learned. Finally, these various results are discussed in relation to more complex forms of behavioral organization and data coming from developmental psychology
This paper presents a connectionist model of correlation based categorization by 10month -old infants (Younger, 1985). Simple autoencoder networks were exposed to the same stimuli used to test 10-month-olds. The familiarisation regime was kept as close as possible to that used with the infants. The model's performance matched that of the infants. Both infants and networks used co-variation information (when available) to segregate items into separate categories. The model provides a mechanistic account of category learning with a test session. It demonstrates how categorization arises as the product of an inextricable interaction between the subject (the infant) and the environment (the stimuli). The computational characteristics of both subject and environment must be considered in conjunction to understand the observed behaviors. Mechanisms of Categorization in Infancy The ability to categorize underlies much of cognition. It is a way of reducing the load on memory and oth...
Infants' own activities create and actively select their learning experiences. Here we review recent models of embodied information seeking and curiosity-driven learning and show that these mechanisms have deep implications for development and evolution. We discuss how these mechanisms yield self-organized epigenesis with emergent ordered behavioral and cognitive developmental stages. We describe a robotic experiment that explored the hypothesis that progress in learning, in and for itself, generates intrinsic rewards: The robot learners probabilistically selected experiences according to their potential for reducing uncertainty. In these experiments, curiosity-driven learning led the robot learner to successively discover object affordances and vocal interaction with its peers. We explain how a learning curriculum adapted to the current constraints of the learning system automatically formed, constraining learning and shaping the developmental trajectory. The observed trajectories in the robot experiment share many properties with those in infant development, including a mixture of regularities and diversities in the developmental patterns. Finally, we argue that such emergent developmental structures can guide and constrain evolution, in particular with regard to the origins of language.
I postulate that human or other intelligent agents function or should function as follows. They store all sensory observations as they come—the data is ‘holy.’ At any time, given some agent’s current coding capabilities, part of the data is compressible by a short and hopefully fast program / description / explanation / world model. In the agent’s subjective eyes, such data is more regular and more beautiful than other data. It is well-known that knowledge of regularity and repeatability may improve the agent’s ability to plan actions leading to external rewards. In absence of such rewards, however, known beauty is boring. Then interestingness becomes the first derivative of subjective beauty: as the learning agent improves its compression algorithm, formerly apparently random data parts become subjectively more regular and beautiful. Such progress in data compression is measured and maximized by the curiosity drive: create action sequences that extend the observation history and yield previously unknown / unpredictable but quickly learnable algorithmic regularity. I discuss how all of the above can be naturally implemented on computers, through an extension of passive unsupervised learning to the case of active data selection: we reward a general reinforcement learner (with access to the adaptive compressor) for actions that improve the subjective compressibility of the growing data. An unusually large compression breakthrough deserves the name discovery. The creativity of artists, dancers, musicians, pure mathematicians can be viewed as a by-product of this principle. Several qualitative examples support this hypothesis.
The Formal Theory of Fun and Creativity (1990–2010) [Schmidhuber, J.: Formal theory of creativity, fun, and intrinsic motivation (1990–2010). IEEE Trans. Auton. Mental Dev. 2(3), 230–247 (2010b)] describes principles of a curious and creative agent that never stops generating nontrivial and novel and surprising tasks and data. Two modules are needed: a data encoder and a data creator. The former encodes the growing history of sensory data as the agent is interacting with its environment; the latter executes actions shaping the history. Both learn. The encoder continually tries to encode the created data more efficiently, by discovering new regularities in it. Its learning progress is the wow-effect or fun or intrinsic reward of the creator, which maximizes future expected reward, being motivated to invent skills leading to interesting data that the encoder does not yet know but can easily learn with little computational effort. I have argued that this simple formal principle explains science and art and music and humor.
Previous research indicates learning words facilitates categorisation. The current study explores how categorisation affects word learning. In the current study, we investigated whether learning about a category facilitates retention of newly learned words by presenting 2-year-old children with multiple referent selection trials to the same object category. In Experiment 1, children either encountered the same exemplar repeatedly or encountered multiple exemplars across trials. All children did very well on the initial task; however, only children who encountered multiple exemplars retained these mappings after a short delay. Experiment 2 replicated and extended this finding by exploring the effect of within-category variability on children's word retention. Children encountered either narrow or broad exemplars across trials. Again, all children did very well on the initial task; however, only children who encountered narrow exemplars retained mappings after a short delay. Overall, these data offer strong evidence that providing children with the opportunity to compare across exemplars during fast mapping facilitates retention.
Computational models are tools for testing mechanistic theories of learning and development. Formal models allow us to instantiate theories of cognitive development in computer simulations. Model behavior can then be compared to real performance. Connectionist models, loosely based on neural information processing, have been successful in capturing a range of developmental phenomena, in particular on-line within-task category learning by young infants. Here we describe two new models. One demonstrates how age dependent changes in neural receptive field sizes can explain observed changes in on-line category learning between 3 and 10 months of age. The other aims to reconcile two conflicting views of infant categorization by focusing on the different task requirements of preferential looking and manual exploration studies. A dual-memory hypothesis posits that within-task category learning that drives looking time behaviors is based on a fast-learning memory system, whereas categorization based on background experience and assessed by paradigms requiring complex motor behavior relies on a second, slow-learning system. The models demonstrate how emphasizing the mechanistic causes of behaviors leads to discovery of deeper, more explanatory accounts of learning and development.
The topics that are to be treated in this book were unduly neglected by psychology for many years but are now beginning to come to the fore. My own researches into attention and exploratory behavior began in 1947, and at about the same time several other psychologists became independently impressed with the importance of these matters and started to study them experimentally. It is interesting that those were also the years when information theory was making its appearance and when the reticular formation of the brain stem was first attracting the notice of neurophysiologists. During the last ten years, the tempo of research into exploratory behavior and related phenomena has been steadily quickening. The book is prompted by the feeling that it is now time to pause and take stock: to review relevant data contributed by several different specialties, to consider what conclusions, whether firm or tentative, are justified at the present juncture, and to clarify what remains to be done. The primary aim of the book is, in fact, to raise problems. The book is intended as a contribution to behavior theory, i.e., to psychology conceived as a branch of science with the circumscribed objective of explaining and predicting behavior. But interest in attention and exploratory behavior and in other topics indissociably bound up with them, such as art, humor and thinking, has by no means been confined to professional psychologists. The book has two features that would have surprised me when I first set out to plan it. One is that it ends up sketching a highly modified form of drive-reduction theory. Drive-reduction theory has appeared more and more to be full of shortcomings, even for the phenomena that it was originally designed to handle. The second surprising feature is the prominence of neurophysiology. (PsycINFO Database Record (c) 2012 APA, all rights reserved)
Three experiments utilizing the familiarization/novelty-preference procedure were conducted to examine the form categorization abilities of newborn and 3- to 4- month-old infants. In the first two experiments, newborn infants discriminated between individual exemplars chosen from different form categories (Expt 1) and from within a form category (Expt 2). In Expt 3, older infants provided evidence of having formed individuated categorical representations for circles, crosses, squares and triangles, whereas newborn infants did not. However, newborn performance was consistent with the formation of broader categorical representations for open versus closed classes of form (i.e. crosses vs. circles, squares and triangles). The results are discussed in terms of a possible differentiation-based developmental change from broad-to-narrow (or global-to-basic) in the perceptual category representations formed by young infants.
This paper considers possible problems researchers might face when interpreting the results of studies that employ variants of the preference procedure. Infants show a tendency to shift their preference from familiar to novel stimuli with increasing exposure to the familiar stimulus, a behaviour that is exploited by the habituation paradigm. This change in attentional preference with exposure leads us to suggest that researchers interested in infants' pre-experimental or spontaneous preferences should beware of the potentially confounding effects of exposing infants to familiarization trials prior to employing the preference procedure. The notion that infant attentional preference is dynamic also calls into question the use of the direction of post-familiarization preference per se when interpreting the knowledge or strategies available to infants. We look into the results of a cross-modal word learning study to show how the interpretation of results may be difficult when infants exhibit a significant preference in an unexpected direction. As a possible solution to this problem we propose that significant preferences in both directions should be sought at multiple intervals over time. Copyright © 2004 John Wiley & Sons, Ltd.
(from the journal abstract) Despite a large body of research demonstrating the kinds of categories to which infants respond, few studies have directly assessed how infants' categorization unfolds over time. Four experiments used a visual familiarization task to evaluate 10-month-old infants' (N = 98) learning of exemplars characterized by commonalities in appearance or function. When learning exemplars with a common function, infants initially responded to the common feature, apparently forming a category, and only learned the individual features with more extensive familiarization. When learning exemplars with a common appearance, infants initially learned the individual features and apparently only formed a category with more extensive familiarization. The results are discussed in terms of models of category learning. (PsycINFO Database Record (c) 2005 APA, all rights reserved).
Two experiments compared infants' attention to the categorical distinction between people and animals in object-examining and sequential-touching tasks. In Experiment 1, 10- and 13-month-old infants distinguished between animals and people in an object-examining task. In this task, infants are familiarized with individual exemplars from one category, and then their response to exemplars from another category is measured. In Experiment 2, 13- and 16-month-old infants, but not 10-month-old infants, attended to the same distinction in a sequential-touching task. In this task, infants are presented with several exemplars from two categories simultaneously, and the order in which they touch those objects is assessed. Evaluation of infants' touching behavior in Experiment 2 also revealed developmental changes in how they approached this task. The combined results of these two experiments confirm the general trend reported in the literature and begin to provide insight into developmental changes that contribute to infants' ability to apply their categorization skills in different task contexts.
The use of mobile robots for inspection tasks is an attractive idea. A robot can travel through environments that humans cannot, and can be trained to identify sensor perceptions that signify potential or actual problems without requiring human intervention. However, in many cases, the appearance of a problem can vary widely, and ensuring that the robot does not miss any possible appearance of the problem (false negatives) is virtually impossible using conventional methods.This paper presents an alternative methodology using novelty detection. A neural network is trained to ignore normal perceptions that do not suggest any problems, so that anything that the robot has not sensed before is highlighted as a possible fault. This makes the incidence of false negatives less likely.We propose a novelty filter that can operate on-line, so that each new input is evaluated for novelty with respect to the data seen so far. The novelty filter learns to ignore inputs that have been sensed previously, or where similar inputs have been perceived. We demonstrate the use of the novelty filter on a series of simple inspection tasks using a mobile robot. The robot highlights those parts of an environment that are novel in some way, that is they are not part of the model acquired during exploration of a different environment. We show the effectiveness of the method using inputs from both sonar sensors and a monochrome camera.
Statistical learning - implicit learning of statistical regularities within sensory input - is a way of acquiring structure within continuous sensory environments. Statistics computation, initially shown to be involved in word segmentation, has been demonstrated to be a general mechanism that operates across domains, across time and space, and across species. Recently, statistical leaning has been reported to be present even at birth when newborns were tested with a speech stream. The aim of the present study was to extend this finding, by investigating whether newborns' ability to extract statistics operates in multiple modalities, as found for older infants and adults. Using the habituation procedure, two experiments were carried out in which visual sequences were presented. Results demonstrate that statistical learning is a general mechanism that extracts statistics across domain since the onset of sensory experience. Intriguingly, present data reveal that newborn learner's limited cognitive resources constrain the functioning of statistical learning, narrowing the range of what can be learned.
How does variability between members of a category influence infants' category learning? We explore the impact of the order in which different items are sampled on category formation. Two groups of 10-months-olds were presented with a series of exemplars to be organized into a single category. In a low distance group, the order of presentation minimized the perceptual distance between consecutive exemplars. In a high distance group, the order of presentation maximized the distance between successive exemplars. At test, only infants in the High Distance condition reliably discriminated between the category prototype and an atypical exemplar. Hence, the order in which infants learnt about the exemplars impacted their categorization performance. Our findings demonstrate the importance of moment-to-moment variations in similarity during infants' category learning.
Human toddlers learn about objects through second-by-second, minute-by-minute sensory-motor interactions. In an effort to understand how toddlers' bodily actions structure the visual learning environment, mini-video cameras were placed low on the foreheads of toddlers, and for comparison also on the foreheads of their parents, as they jointly played with toys. Analyses of the head camera views indicate visual experiences with profoundly different dynamic structures. The toddler view often consists of a single dominating object that is close to the sensors and thus that blocks the view of other objects such that individual objects go in and out of view. The adult view, in contrast, is broad and stable, with all potential targets continually in view. These differences may arise for several developmentally relevant reasons, including the small visuo-motor workspace of the toddler (short arms) and the engagement of the whole body when actively handling objects.
Infants do not readily organize using form similarity: 6- to 7-month-olds familiarized with horizontal or vertical bars (filled rectangles) do not display a subsequent preference for a novel column versus row organization of X-O elements (Quinn and Bhatt, 2006 Journal of Experimental Psychology: Human Perception and Performance 32 1221-1230). In experiment 1, infants were familiarized with more complex bars composed of beads or crosshatches, and performance was again unsuccessful. In experiment 2, all three bar types were presented during familiarization and infants performed successfully, indicating that variability in pattern information depicting an invariant structure enhances the learning of perceptual organization. In experiment 3, we examined whether the manner in which variability is experienced (simultaneous versus sequential contrast) impacts this learning. One group of infants was familiarized with a single pattern containing the three different bar types (within-trial variability), and another was presented with the same three bar types, but with each appearing on a different trial (across-trial variability). Only the across-trial variability group performed successfully, suggesting that trial-to-trial change in local element information induced by sequential presentation is a significant factor in facilitating the learning of perceptual organization.
2 experiments were conducted investigating infants' use of structural relations (i.e., correlated attribute values) in dividing or segregating items into categories. Rosch argued that attribute values do not occur in all possible combinations in the real world. Even though values of attributes may vary continuously across objects, some combinations are more likely to occur than others, forming breaks or discontinuities between clusters of correlated values. Within a well-controlled laboratory context, using artificial categories and a standard infant-recognition memory procedure, the present experiments demonstrated 10-month-old infants' sensitivity to structural information like that proposed by Rosch to exist in the real world, and their ability to segregate items into categories on the basis of clusters of correlated attribute values.
The paired-preference procedure was used in a series of experiments to explore the abilities of infants aged 3 and 4 months to categorize photographic exemplars from natural (adult-defined) basic-level categories. The question of whether the categorical representations that were evidenced excluded members of a related, perceptually similar category was also investigated. Experiments 1-3 revealed that infants could form categorical representations for dogs and cats that excluded birds. Experiment 4 showed that the representation for cats also excluded dogs, but that the representation for dogs did not exclude cats. However, a supplementary experiment showed that the representation for dogs did exclude cats when the variability of the dog exemplars was reduced to match that of the cat exemplars. The results are discussed in terms of abilities necessary for the formation of more complex categorical representations.
We investigated how exposure to pairs of different items (as compared with pairs of identical items) influences 10-month-olds' (n=79) categorization of horses versus dogs in an object-examining task. Infants responded to an exclusive category when familiarized with pairs of different items but not when familiarized with pairs of identical items (Experiment 1), even when the frequency of exposure to each item was controlled (Experiment 2). When familiarized with pairs of identical items, infants failed to show evidence of memory for the individual exemplars (Experiment 3). Reducing the retention interval between presentations of different items in the identical pairs condition facilitated infants' recognition of an exclusive categorical distinction (Experiment 4). These results are discussed in terms of how exposure to collections of different items, and how opportunities to compare items, influences infants' categorization.
Development in any domain is often characterized by increasingly abstract representations. Recent evidence in the domain of shape recognition provides one example; between 18 and 24 months children appear to build increasingly abstract representations of object shape [Smith, L. B. (2003). Learning to recognize objects. Psychological Science, 14, 244-250]. Abstraction is in part simplification because it requires the removal of irrelevant information. At the same time, part of generalization is ignoring irrelevant differences. The resulting prediction is this: simplification may enable generalization. Four experiments asked whether simple training instances could shortcut the process of abstraction and directly promote appropriate generalization. Toddlers were taught novel object categories with either simple or complex training exemplars. We found that children who learned with simple objects were able to generalize according to shape similarity, typically relevant for early object categories, better than those who learned with complex objects. Abstraction is the product of learning; using simplified - already abstracted instances - can short-cut that learning, leading to robust generalization.
I postulate that human or other intelligent agents function or should function as follows. They store all sensory observations as they come - the data is holy. At any time, given some agent's current coding capabilities, part of the data is compressible by a short and hopefully fast program / description / explanation / world model. In the agent's subjective eyes, such data is more regular and more "beautiful" than other data. It is well-known that knowledge of regularity and repeatability may improve the agent's ability to plan actions leading to external rewards. In absence of such rewards, however, known beauty is boring. Then "interestingness" becomes the first derivative of subjective beauty: as the learning agent improves its compression algorithm, formerly apparently random data parts become subjectively more regular and beautiful. Such progress in compressibility is measured and maximized by the curiosity drive: create action sequences that extend the observation history and yield previously unknown / unpredictable but quickly learnable algorithmic regularity. We discuss how all of the above can be naturally implemented on computers, through an extension of passive unsupervised learning to the case of active data selection: we reward a general reinforcement learner (with access to the adaptive compressor) for actions that improve the subjective compressibility of the growing data. An unusually large breakthrough in compressibility deserves the name "discovery". The "creativity" of artists, dancers, musicians, pure mathematicians can be viewed as a by-product of this principle. Several qualitative examples support this hypothesis.
Same items, different order: Effects of temporal variability on infant categorization A bottom-up view of toddler word learning
- E Mather
- K ] A F Plunkett
- L B Pereira
- C Smith
- Yu
E. Mather and K. Plunkett, " Same items, different order: Effects of
temporal variability on infant categorization, " Cognition, vol. 119, no.
3, pp. 438–447, 2011.
[2]
A. F. Pereira, L. B. Smith, and C. Yu, " A bottom-up view of toddler
word learning, " Psychon. Bull. Rev., vol. 21, no. 1, pp. 178–185,
2014.
[3]
Novelty, attention, and challenges for developmental psychology Not your mother's view: The dynamics of toddler visual experience
- E B Mather4 ] L
- C Smith
- A F Yu
- Pereira
E. Mather, " Novelty, attention, and challenges for developmental
psychology, " Front. Psychol., vol. 4, Aug. 2013.
[4]
L. B. Smith, C. Yu, and A. F. Pereira, " Not your mother's view: The
dynamics of toddler visual experience, " Dev. Sci., vol. 14, no. 1, pp.
9–17, 2011.
[5]
Robust active binocular vision through intrinsically motivated learning On-line novelty detection for autonomous mobile robots Emergent structured transition from variation to repetition in a biologically-plausible model of learning in basal ganglia
- L Lonini
- S Forestier
- C Teulière
- Y Zhao
- B E Shi
- J S Triesch
- U Marsland
- J Nehmzow
- A Shapiro
- K N Shah
- Gurney
L. Lonini, S. Forestier, C. Teulière, Y. Zhao, B. E. Shi, and J. Triesch,
" Robust active binocular vision through intrinsically motivated
learning, " Front. Neurorobotics, vol. 7, p. 20, 2013.
[7]
S. Marsland, U. Nehmzow, and J. Shapiro, " On-line novelty detection
for autonomous mobile robots, " Robot. Auton. Syst., vol. 51, no. 2, pp.
191–206, 2005.
[8]
A. Shah and K. N. Gurney, " Emergent structured transition from
variation to repetition in a biologically-plausible model of learning in
basal ganglia, " Front. Psychol., vol. 5, Feb. 2014.
[9]