Connectionist models of development

Toward computational neuroconstructivism: a framework for developmental systems neuroscience

Brain development is underpinned by complex interactions between neural assemblies, driving structural and functional change. This neuroconstructivism (the notion that neural functions are shaped by these interactions) is core to some developmental theories. However, due to their complexity, understanding underlying developmental mechanisms is challenging. Elsewhere in neurobiology, a computational revolution has shown that mathematical models of hidden biological mechanisms can bridge observations with theory building. Can we build a similar computational framework yielding mechanistic insights for brain development? Here, we outline the conceptual and technical challenges of addressing this theory gap, and demonstrate that there is great potential in specifying brain development as mathematically defined processes operating within physical constraints. We provide examples, alongside broader ingredients needed, as the field explores computational explanations of system-wide development.

Efficient neural codes naturally emerge through gradient descent learning

Human sensory systems are more sensitive to common features in the environment than uncommon features. For example, small deviations from the more frequently encountered horizontal orientations can be more easily detected than small deviations from the less frequent diagonal ones. Here we find that artificial neural networks trained to recognize objects also have patterns of sensitivity that match the statistics of features in images. To interpret these findings, we show mathematically that learning with gradient descent in neural networks preferentially creates representations that are more sensitive to common features, a hallmark of efficient coding. This effect occurs in systems with otherwise unconstrained coding resources, and additionally when learning towards both supervised and unsupervised objectives. This result demonstrates that efficient codes can naturally emerge from gradient-like learning.

Machine Learning as a Model for Cultural Learning: Teaching an Algorithm What it Means to be Fat

Public culture is a powerful source of cognitive socialization; for example, media language is full of meanings about body weight. Yet it remains unclear how individuals process meanings in public culture. We suggest that schema learning is a core mechanism by which public culture becomes personal culture. We propose that a burgeoning approach in computational text analysis - neural word embeddings - can be interpreted as a formal model for cultural learning. Embeddings allow us to empirically model schema learning and activation from natural language data. We illustrate our approach by extracting four lower-order schemas from news articles: the gender, moral, health, and class meanings of body weight. Using these lower-order schemas we quantify how words about body weight "fill in the blanks" about gender, morality, health, and class. Our findings reinforce ongoing concerns that machine-learning models (e.g., of natural language) can encode and reproduce harmful human biases.

Braitenberg Vehicles as Developmental Neurosimulation

Connecting brain and behavior is a longstanding issue in the areas of behavioral science, artificial intelligence, and neurobiology. As is standard among models of artificial and biological neural networks, an analogue of the fully mature brain is presented as a blank slate. However, this does not consider the realities of biological development and developmental learning. Our purpose is to model the development of an artificial organism that exhibits complex behaviors. We introduce three alternate approaches to demonstrate how developmental embodied agents can be implemented. The resulting developmental Braitenberg vehicles (dBVs) will generate behaviors ranging from stimulus responses to group behavior that resembles collective motion. We will situate this work in the domain of artificial brain networks along with broader themes such as embodied cognition, feedback, and emergence. Our perspective is exemplified by three software instantiations that demonstrate how a BV-genetic algorithm hybrid model, a multisensory Hebbian learning model, and multi-agent approaches can be used to approach BV development. We introduce use cases such as optimized spatial cognition (vehicle-genetic algorithm hybrid model), hinges connecting behavioral and neural models (multisensory Hebbian learning model), and cumulative classification (multi-agent approaches). In conclusion, we consider future applications of the developmental neurosimulation approach.

Quantitative or qualitative development in decision making?

A key question in the developmental sciences is whether developmental differences are quantitative or qualitative. For example, does age increase the speed in processing a task (quantitative differences) or does age affect the way a task is processed (qualitative differences)? Until now, findings in the domain of decision making have been based on the assumption that developmental differences are either quantitative or qualitative. In the current study, we took a different approach in which we tested whether development is best described as being quantitative or qualitative. We administered a judgment version and a choice version of a decision-making task to a developmental sample (n_judgment = 109 and n_choice = 137; M_age = 12.5 years, age range = 9-18). The task, the so-called Gambling Machine Task, required decisions between two options characterized by constant gains and probabilistic losses; these characteristics were known beforehand and thus did not need to be learned from experience. Data were analyzed by comparing the fit of quantitative and qualitative latent variable models, so-called multiple indicator multiple cause (MIMIC) models. Results indicated that individual differences in both judgment and choice tasks were quantitative and pertained to individual differences in "consideration of gains," that is, to what extent decisions were guided by gains. These differences were affected by age in the judgment version, but not in the choice version, of the task. We discuss implications for theories of decision making and discuss potential limitations and extensions. We also argue that the MIMIC approach is useful in other domains, for example, to test quantitative versus qualitative development of categorization, reasoning, math, and memory.

Optimum Integration Procedure for Connectionist and Dynamic Field Equations

Connectionist and dynamic field models consist of a set of coupled first-order differential equations describing the evolution in time of different units. We compare three numerical methods for the integration of these equations: the Euler method, and two methods we have developed and present here: a modified version of the fourth-order Runge Kutta method, and one semi-analytical method. We apply them to solve a well-known nonlinear connectionist model of retrieval in single-digit multiplication, and show that, in many regimes, the semi-analytical and modified Runge Kutta methods outperform the Euler method, in some regimes by more than three orders of magnitude. Given the outstanding difference in execution time of the methods, and that the EM is widely used, we conclude that the researchers in the field can greatly benefit from our analysis and developed methods.

Looking represents choosing in toddlers: Exploring the equivalence between multimodal measures in forced-choice tasks

In the two-alternative forced-choice (2AFC) paradigm, manual responses such as pointing have been widely used as measures to estimate cognitive abilities. While pointing measurements can be easily collected, coded, analyzed, and interpreted, absent responses are often observed particularly when adopting these measures for toddler studies, which leads to an increase of missing data. Although looking responses such as preferential looking can be available as alternative measures in such cases, it is unknown how well looking measurements can be interpreted as equivalent to manual ones. This study aimed to answer this question by investigating how accurately pointing responses (i.e., left or right) could be predicted from concurrent preferential looking. Using pre-existing videos of toddlers aged 18-23 months engaged in an intermodal word comprehension task, we developed models predicting manual from looking responses. Results showed substantial prediction accuracy for both the Simple Majority Vote and Machine Learning-Based classifiers, which indicates that looking responses would be reasonable alternative measures of manual ones. However, the further exploratory analysis revealed that when applying the created models for data of toddlers who did not produce clear pointing responses, the estimation agreement of missing pointing between the models and the human coders slightly dropped. This indicates that looking responses without pointing were qualitatively different from those with pointing. Bridging two measurements in forced-choice tasks would help researchers avoid wasting collected data due to the absence of manual responses and interpret results from different modalities comprehensively.

Partial knowledge in the development of number word understanding

A common measure of number word understanding is the give-N task. Traditionally, to receive credit for understanding a number, N, children must understand that N does not apply to other set sizes (e.g. a child who gives three when asked for 'three' but also when asked for 'four' would not be credited with knowing 'three'). However, it is possible that children who correctly provide the set size directly above their knower level but also provide that number for other number words ('N + 1 givers') may be in a partial, transitional knowledge state. In an integrative analysis including 191 preschoolers, subset knowers who correctly gave N + 1 at pretest performed better at posttest than did those who did not correctly give N + 1. This performance was not reflective of 'full' knowledge of N + 1, as N + 1 givers performed worse than traditionally coded knowers of that set size on separate measures of number word understanding within a given timepoint. Results support the idea of graded representations (Munakata, Trends in Cognitive Sciences, 5, 309-315, 2001.) in number word development and suggest traditional approaches to coding the give-N task may not completely capture children's knowledge.

Goldilocks Forgetting in Cross-Situational Learning

Given that there is referential uncertainty (noise) when learning words, to what extent can forgetting filter some of that noise out, and be an aid to learning? Using a Cross Situational Learning model we find a U-shaped function of errors indicative of a "Goldilocks" zone of forgetting: an optimum store-loss ratio that is neither too aggressive nor too weak, but just the right amount to produce better learning outcomes. Forgetting acts as a high-pass filter that actively deletes (part of) the referential ambiguity noise, retains intended referents, and effectively amplifies the signal. The model achieves this performance without incorporating any specific cognitive biases of the type proposed in the constraints and principles account, and without any prescribed developmental changes in the underlying learning mechanism. Instead we interpret the model performance as more of a by-product of exposure to input, where the associative strengths in the lexicon grow as a function of linguistic experience in combination with memory limitations. The result adds a mechanistic explanation for the experimental evidence on spaced learning and, more generally, advocates integrating domain-general aspects of cognition, such as memory, into the language acquisition process.

Curiosity-based learning in infants: a neurocomputational approach

Infants are curious learners who drive their own cognitive development by imposing structure on their learning environment as they explore. Understanding the mechanisms by which infants structure their own learning is therefore critical to our understanding of development. Here we propose an explicit mechanism for intrinsically motivated information selection that maximizes learning. We first present a neurocomputational model of infant visual category learning, capturing existing empirical data on the role of environmental complexity on learning. Next we "set the model free", allowing it to select its own stimuli based on a formalization of curiosity and three alternative selection mechanisms. We demonstrate that maximal learning emerges when the model is able to maximize stimulus novelty relative to its internal states, depending on the interaction across learning between the structure of the environment and the plasticity in the learner itself. We discuss the implications of this new curiosity mechanism for both existing computational models of reinforcement learning and for our understanding of this fundamental mechanism in early development.

Input and Age-Dependent Variation in Second Language Learning: A Connectionist Account

Language learning requires linguistic input, but several studies have found that knowledge of second language (L2) rules does not seem to improve with more language exposure (e.g., Johnson & Newport, 1989). One reason for this is that previous studies did not factor out variation due to the different rules tested. To examine this issue, we reanalyzed grammaticality judgment scores in Flege, Yeni-Komshian, and Liu's (1999) study of L2 learners using rule-related predictors and found that, in addition to the overall drop in performance due to a sensitive period, L2 knowledge increased with years of input. Knowledge of different grammar rules was negatively associated with input frequency of those rules. To better understand these effects, we modeled the results using a connectionist model that was trained using Korean as a first language (L1) and then English as an L2. To explain the sensitive period in L2 learning, the model's learning rate was reduced in an age-related manner. By assigning different learning rates for syntax and lexical learning, we were able to model the difference between early and late L2 learners in input sensitivity. The model's learning mechanism allowed transfer between the L1 and L2, and this helped to explain the differences between different rules in the grammaticality judgment task. This work demonstrates that an L1 model of learning and processing can be adapted to provide an explicit account of how the input and the sensitive period interact in L2 learning.

Assessing a continuum of lexical-semantic knowledge in the second year of life: A multimodal approach

Behavioral dissociations in young children's visual and haptic responses have been taken as evidence that word knowledge is not all-or-none but instead exists on a continuum from absence of knowledge, to partial knowledge, to robust knowledge. This longitudinal study tested a group of 16- to 18-month-olds, 6months after their initial visit, to replicate results of partial understanding as shown by visual-haptic dissociations and to determine whether partial knowledge of word-referent relations can be leveraged for future word recognition. Results show that, like 16-month-olds, 22-month-olds demonstrate behavioral dissociations exhibited by rapid visual reaction times to a named referent but incorrect haptic responses. Furthermore, results suggest that partial word knowledge at one time predicts the degree to which that word will be understood in the future.

Beyond What Develops When

There is no single methodology that can fully explain the nature of human development and learning. Yet, headway is being made on how cognitive milestones are achieved during development with the use of magnetic resonance imaging (MRI) technology. With this methodology, it is possible to assess changes in brain structure, function, and connectivity. Recent findings suggest that both progressive and regressive processes—as opposed to simple linear patterns of change—underlie changes in cognitive abilities. Functional MRI studies suggest that both biological maturation and learning correspond to a fine-tuning of neural systems with enhanced recruitment of task-relevant regions. This fine-tuning of cortical systems corresponds with their enhanced connectivity with cortical and subcortical circuitry. In sum, imaging has helped to move the field of cognitive development beyond questions of what develops and when, to how these changes may occur.

The acquisition process of musical tonal schema: implications from connectionist modeling

Using connectionist modeling, we address fundamental questions concerning the acquisition process of musical tonal schema of listeners. Compared to models of previous studies, our connectionist model (Learning Network for Tonal Schema, LeNTS) was better equipped to fulfill three basic requirements. Specifically, LeNTS was equipped with a learning mechanism, bound by culture-general properties, and trained by sufficient melody materials. When exposed to Western music, LeNTS acquired musical ‘scale’ sensitivity early and ‘harmony’ sensitivity later. The order of acquisition of scale and harmony sensitivities shown by LeNTS was consistent with the culture-specific acquisition order shown by musically westernized children. The implications of these results for the acquisition process of a tonal schema of listeners are as follows: (a) the acquisition process may entail small and incremental changes, rather than large and stage-like changes, in corresponding neural circuits; (b) the speed of schema acquisition may mainly depend on musical experiences rather than maturation; and (c) the learning principles of schema acquisition may be culturally invariant while the acquired tonal schemas are varied with exposed culture-specific music.

Putting Education in “Educational” Apps

Children are in the midst of a vast, unplanned experiment, surrounded by digital technologies that were not available but 5 years ago. At the apex of this boom is the introduction of applications (“apps”) for tablets and smartphones. However, there is simply not the time, money, or resources available to evaluate each app as it enters the market. Thus, “educational” apps—the number of which, as of January 2015, stood at 80,000 in Apple’s App Store (Apple, 2015)—are largely unregulated and untested. This article offers a way to define the potential educational impact of current and future apps. We build upon decades of work on the Science of Learning, which has examined how children learn best. From this work, we abstract a set of principles for two ultimate goals. First, we aim to guide researchers, educators, and designers in evidence-based app development. Second, by creating an evidence-based guide, we hope to set a new standard for evaluating and selecting the most effective existing children’s apps. In short, we will show how the design and use of educational apps aligns with known processes of children’s learning and development and offer a framework that can be used by parents and designers alike. Apps designed to promote active, engaged, meaningful, and socially interactive learning—four “pillars” of learning—within the context of a supported learning goal are considered educational.

Transfer of conflict and cooperation from experienced games to new games: a connectionist model of learning

The question of whether, and if so how, learning can be transfered from previously experienced games to novel games has recently attracted the attention of the experimental game theory literature. Existing research presumes that learning operates over actions, beliefs or decision rules. This study instead uses a connectionist approach that learns a direct mapping from game payoffs to a probability distribution over own actions. Learning is operationalized as a backpropagation rule that adjusts the weights of feedforward neural networks in the direction of increasing the probability of an agent playing a myopic best response to the last game played. One advantage of this approach is that it expands the scope of the model to any possible n × n normal-form game allowing for a comprehensive model of transfer of learning. Agents are exposed to games drawn from one of seven classes of games with significantly different strategic characteristics and then forced to play games from previously unseen classes. I find significant transfer of learning, i.e., behavior that is path-dependent, or conditional on the previously seen games. Cooperation is more pronounced in new games when agents are previously exposed to games where the incentive to cooperate is stronger than the incentive to compete, i.e., when individual incentives are aligned. Prior exposure to Prisoner's dilemma, zero-sum and discoordination games led to a significant decrease in realized payoffs for all the game classes under investigation. A distinction is made between superficial and deep transfer of learning both—the former is driven by superficial payoff similarities between games, the latter by differences in the incentive structures or strategic implications of the games. I examine whether agents learn to play the Nash equilibria of games, how they select amongst multiple equilibria, and whether they transfer Nash equilibrium behavior to unseen games. Sufficient exposure to a strategically heterogeneous set of games is found to be a necessary condition for deep learning (and transfer) across game classes. Paradoxically, superficial transfer of learning is shown to lead to better outcomes than deep transfer for a wide range of game classes. The simulation results corroborate important experimental findings with human subjects, and make several novel predictions that can be tested experimentally.

Looking and touching: what extant approaches reveal about the structure of early word knowledge

The goal of the current study is to assess the temporal dynamics of vision and action to evaluate the underlying word representations that guide infants' responses. Sixteen-month-old infants participated in a two-alternative forced-choice word-picture matching task. We conducted a moment-by-moment analysis of looking and reaching behaviors as they occurred in tandem to assess the speed with which a prompted word was processed (visual reaction time) as a function of the type of haptic response: Target, Distractor, or No Touch. Visual reaction times (visual RTs) were significantly slower during No Touches compared to Distractor and Target Touches, which were statistically indistinguishable. The finding that visual RTs were significantly faster during Distractor Touches compared to No Touches suggests that incorrect and absent haptic responses appear to index distinct knowledge states: incorrect responses are associated with partial knowledge whereas absent responses appear to reflect a true failure to map lexical items to their target referents. Further, we found that those children who were faster at processing words were also those children who exhibited better haptic performance. This research provides a methodological clarification on knowledge measured by the visual and haptic modalities and new evidence for a continuum of word knowledge in the second year of life.

Learning to read as the formation of a dynamic system: evidence for dynamic stability in phonological recoding

Two aspects of dynamic systems approaches that are pertinent to developmental models of reading are the emergence of a system with self-organizing characteristics, and its evolution over time to a stable state that is not easily modified or perturbed. The effects of dynamic stability may be seen in the differences obtained in the processing of print by beginner readers taught by different approaches to reading (phonics and text-centered), and more long-term effects on adults, consistent with these differences. However, there is little direct evidence collected over time for the same participants. In this study, lexicalized (implicit) phonological processing, and explicit phonological and letter-sound skills are further examined in a precocious reader whose early development at 3 and 5 years has been extensively described (Cognition, 2000, 2004). At ages 10 and 14 years, comparisons were made with these earlier reports and skilled adult readers, using the same tasks for evidence of changes in reading processes. The results showed that along with an increase of reading accuracy and speed, her pattern of lexicalized phonological responses for reading did not change over time. Neither did her pattern of explicit phonological and letter-sound skills, aspects of which were inferior to her lexicalized phonological processing, and word reading. These results suggest dynamic stability of the word reading system. The early emergence of this system with minimal explicit skill development calls into question developmental reading theories that require such skills for learning to read. Currently, only the Knowledge Sources theory of reading acquisition can account for such findings. Consideration of these aspects of dynamic systems raise theoretical issues that could result in a paradigm shift with regard to best practice and intervention.

The amygdala: an agent of change in adolescent neural networks

This article is part of a Special Issue "Puberty and Adolescence". A unique component of adolescent development is the need to master new developmental tasks in which peer interactions become primary (for the purposes of becoming autonomous from parents, forming intimate friendships, and romantic/sexual partnerships). Previously, it has been suggested that the ability to master these tasks requires an important re-organization in the relation between perceptual, motivational, affective, and cognitive systems in a very general and broad way that is fundamentally influenced by the infusion of sex hormones during pubertal development (Scherf et al., 2012). Herein, we extend this argument to suggest that the amygdala, which is vastly connected with cortical and subcortical regions and contains sex hormone receptors, may lie at the heart of this re-organization. We propose that during adolescent development there is a shift in the attribution of relevance to existing stimuli and contexts that is mediated by the amygdala (e.g., heightened relevance of peer faces, reduced relevance of physical distance from parents). As a result, amygdala inputs to existing stable neural networks are re-weighted (increased or decreased), which destabilizes the functional interactions among regions within these networks and allows for a critical restructuring of the network functional organization. This process of network re-organization enables processing of qualitatively new kinds of social information and the emergence of novel behaviors that support mastery of adolescent-specific developmental tasks.

Prelinguistic communication development in children with childhood apraxia of speech: a retrospective analysis

In a retrospective study of prelinguistic communication development, clinically referred preschool children (n = 9) aged 3-4 years, who as infants had failed a community-based screening program, were evaluated for features of childhood apraxia of speech (CAS). Four children showed no features and either delayed or normal language, five had from three-to-seven CAS features and all exhibited delayed language. These children were matched by age with 21 children with typically-developing (TD) speech and language skills. Case-control comparisons of retrospective data from 9 months of age for two participants with more severe features of CAS at preschool age showed a dissociated pattern with low expressive quotients on the Receptive-Expressive Emergent Language Assessment-Second Edition (REEL-2) and records of infrequent babbling, but normal receptive quotients. However, other profiles were observed. Two children with milder CAS features showed poor receptive and expressive development similar to other clinically referred children with no CAS features, and one child with severe CAS features showed poor receptive but normal expressive developmental milestones at 9 months and records of frequent babbling. Results suggest some but not all children with features of suspected CAS have a selective deficit originating within speech motor development.

DEVELOPMENTAL ROBOTICS: THEORY AND EXPERIMENTS

A hand-designed internal representation of the world cannot deal with unknown or uncontrolled environments. Motivated by human cognitive and behavioral development, this paper presents a theory, an architecture, and some experimental results for developmental robotics. By a developmental robot, we mean that the robot generates its "brain" (or "central nervous system," including the information processor and controller) through online, real-time interactions with its environment (including humans). A new Self-Aware Self-Effecting (SASE) agent concept is proposed, based on our SAIL and Dav developmental robots. The manual and autonomous development paradigms are formulated along with a theory of representation suited for autonomous development. Unlike traditional robot learning, the tasks that a developmental robot ends up learning are unknown during the programming time so that the task-specific representation must be generated and updated through real-time "living" experiences. Experimental results with SAIL and Dav developmental robots are presented, including visual attention selection, autonomous navigation, developmental speech learning, range-based obstacle avoidance, and scaffolding through transfer and chaining.