David Marr's (1982) three‐level analysis of computational cognition argues for three distinct levels of cognitive information processing—namely, the computational, representational, and implementational levels. But Marr's levels are—and were meant to be—descriptive, rather than interactive and dynamic. For this reason, we suggest that, had Marr been writing today, he might well have gone even farther in his analysis, including the emergence of structure—in particular, explicit structure at the conceptual level—from lower levels, and the effect of explicit emergent structures on the level (...) (or levels) that gave rise to them. The message is that today's cognitive scientists need not only to understand how emergent structures—in particular, explicit emergent structures at the cognitive level—develop but also to understand how they feed back on the sub‐structures from which they emerged. (shrink)

What makes a word easy to learn? Early‐learned words are frequent and tend to name concrete referents. But words typically do not occur in isolation. Some words are predictable from their contexts; others are less so. Here, we investigate whether predictability relates to when children start producing different words (age of acquisition; AoA). We operationalized predictability in terms of a word's surprisal in child‐directed speech, computed using n‐gram and long‐short‐term‐memory (LSTM) language models. Predictability derived from LSTMs was generally a better (...) predictor than predictability derived from n‐gram models. Across five languages, average surprisal was positively correlated with the AoA of predicates and function words but not nouns. Controlling for concreteness and word frequency, more predictable predicates and function words were learned earlier. Differences in predictability between languages were associated with cross‐linguistic differences in AoA: the same word (when it was a predicate) was produced earlier in languages where the word was more predictable. (shrink)

Speech unfolds over time, and the cues for even a single phoneme are rarely available simultaneously. Consequently, to recognize a single phoneme, listeners must integrate material over several hundred milliseconds. Prior work contrasts two accounts: (a) a memory buffer account in which listeners accumulate auditory information in memory and only access higher level representations (i.e., lexical representations) when sufficient information has arrived; and (b) an immediate integration scheme in which lexical representations can be partially activated on the basis of early (...) cues and then updated when more information arises. These studies have uniformly shown evidence for immediate integration for a variety of phonetic distinctions. We attempted to extend this to fricatives, a class of speech sounds which requires not only temporal integration of asynchronous cues (the frication, followed by the formant transitions 150–350 ms later), but also integration across different frequency bands and compensation for contextual factors like coarticulation. Eye movements in the visual world paradigm showed clear evidence for a memory buffer. Results were replicated in five experiments, ruling out methodological factors and tying the release of the buffer to the onset of the vowel. These findings support a general auditory account for speech by suggesting that the acoustic nature of particular speech sounds may have large effects on how they are processed. It also has major implications for theories of auditory and speech perception by raising the possibility of an encapsulated memory buffer in early auditory processing. (shrink)

Speech unfolds over time, and the cues for even a single phoneme are rarely available simultaneously. Consequently, to recognize a single phoneme, listeners must integrate material over several hundred milliseconds. Prior work contrasts two accounts: (a) a memory buffer account in which listeners accumulate auditory information in memory and only access higher level representations (i.e., lexical representations) when sufficient information has arrived; and (b) an immediate integration scheme in which lexical representations can be partially activated on the basis of early (...) cues and then updated when more information arises. These studies have uniformly shown evidence for immediate integration for a variety of phonetic distinctions. We attempted to extend this to fricatives, a class of speech sounds which requires not only temporal integration of asynchronous cues (the frication, followed by the formant transitions 150–350 ms later), but also integration across different frequency bands and compensation for contextual factors like coarticulation. Eye movements in the visual world paradigm showed clear evidence for a memory buffer. Results were replicated in five experiments, ruling out methodological factors and tying the release of the buffer to the onset of the vowel. These findings support a general auditory account for speech by suggesting that the acoustic nature of particular speech sounds may have large effects on how they are processed. It also has major implications for theories of auditory and speech perception by raising the possibility of an encapsulated memory buffer in early auditory processing. (shrink)

Sequential action makes up the bulk of human daily activity, and yet much remains unknown about how people learn such actions. In one motor learning paradigm, the serial reaction time (SRT) task, people are taught a consistent sequence of button presses by cueing them with the next target response. However, the SRT task only records keypress response times to a cued target, and thus it cannot reveal the full time‐course of motion, including predictive movements. This paper describes a mouse movement (...) trajectory SRT task in which the cursor must be moved to a cued location. We replicated keypress SRT results, but also found that predictive movement—before the next cue appears—increased during the experiment. Moreover, trajectory analyses revealed that people developed a centering strategy under uncertainty. In a second experiment, we made prediction explicit, no longer cueing targets. Thus, participants had to explore the response alternatives and learn via reinforcement, receiving rewards and penalties for correct and incorrect actions, respectively. Participants were not told whether the sequence of stimuli was deterministic, nor if it would repeat, nor how long it was. Given the difficulty of the task, it is unsurprising that some learners performed poorly. However, many learners performed remarkably well, and some acquired the full 10‐item sequence within 10 repetitions. Comparing the high‐ and low‐performers’ detailed results in this reinforcement learning (RL) task with the first experiment's cued trajectory SRT task, we found similarities between the two tasks, suggesting that the effects in Experiment 1 are due to predictive, rather than reactive processes. Finally, we found that two standard model‐free reinforcement learning models fit the high‐performing participants, while the four low‐performing participants provide better fit with a simple negative recency bias model. (shrink)

Our understanding of the cognitive processes involved in reading has been advanced by computational models that simulate those processes. Unfortunately, most of these models have been developed to explain the reading of English and other alphabetic languages, with relatively fewer efforts to examine whether or not the assumptions of these models also explain what has been learned from other languages and, in particular, non-alphabetic writing systems like Chinese. In this article, we will review those computational models that have been developed (...) to explain the reading of Chinese, with the goal of comparing their theoretical assumptions to those of models that explain the reading of English. Our analysis indicates that there are both points of convergence and divergence between the theoretical assumptions of Chinese versus English models, suggesting that the cognitive systems supporting reading may be differentially influenced by features of the languages and/or writing systems, or that certain theoretical assumptions developed to explain the reading of one language might be adapted to explain the reading of others. (shrink)

From a theoretical perspective, most discussions of statistical learning have focused on the possible “statistical” properties that are the object of learning. Much less attention has been given to defining what “learning” is in the context of “statistical learning.” One major difficulty is that SL research has been monitoring participants’ performance in laboratory settings with a strikingly narrow set of tasks, where learning is typically assessed offline, through a set of two-alternative-forced-choice questions, which follow a brief visual or auditory familiarization (...) stream. Is that all there is to characterizing SL abilities? Here we adopt a novel perspective for investigating the processing of regularities in the visual modality. By tracking online performance in a self-paced SL paradigm, we focus on the trajectory of learning. In a set of three experiments we show that this paradigm provides a reliable and valid signature of SL performance, and it offers important insights for understanding how statistical regularities are perceived and assimilated in the visual modality. This demonstrates the promise of integrating different operational measures to our theory of SL. (shrink)

Ten years ago, researchers using event-related brain potentials to study language comprehension were puzzled by what looked like a Semantic Illusion: Semantically anomalous, but structurally well-formed sentences did not affect the N400 component—traditionally taken to reflect semantic integration—but instead produced a P600 effect, which is generally linked to syntactic processing. This finding led to a considerable amount of debate, and a number of complex processing models have been proposed as an explanation. What these models have in common is that they (...) postulate two or more separate processing streams, in order to reconcile the Semantic Illusion and other semantically induced P600 effects with the traditional interpretations of the N400 and the P600. Recently, however, these multi-stream models have been called into question, and a simpler single-stream model has been proposed. According to this alternative model, the N400 component reflects the retrieval of word meaning from semantic memory, and the P600 component indexes the integration of this meaning into the unfolding utterance interpretation. In the present paper, we provide support for this “Retrieval–Integration ” account by instantiating it as a neurocomputational model. This neurocomputational model is the first to successfully simulate the N400 and P600 amplitude in language comprehension, and simulations with this model provide a proof of concept of the single-stream RI account of semantically induced patterns of N400 and P600 modulations. (shrink)

Rumelhart and McClelland's chapter about learning the past tense created a degree of controversy extraordinary even in the adversarial culture of modern science. It also stimulated a vast amount of research that advanced the understanding of the past tense, inflectional morphology in English and other languages, the nature of linguistic representations, relations between language and other phenomena such as reading and object recognition, the properties of artificial neural networks, and other topics. We examine the impact of the Rumelhart and McClelland (...) model with the benefit of 25 years of hindsight. It is not clear who “won” the debate. It is clear, however, that the core ideas that the model instantiated have been assimilated into many areas in the study of language, changing the focus of research from abstract characterizations of linguistic competence to an emphasis on the role of the statistical structure of language in acquisition and processing. (shrink)

This paper introduces a special issue of Cognitive Science initiated on the 25th anniversary of the publication of Parallel Distributed Processing (PDP), a two-volume work that introduced the use of neural network models as vehicles for understanding cognition. The collection surveys the core commitments of the PDP framework, the key issues the framework has addressed, and the debates the framework has spawned, and presents viewpoints on the current status of these issues. The articles focus on both historical roots and contemporary (...) developments in learning, optimality theory, perception, memory, language, conceptual knowledge, cognitive control, and consciousness. Here we consider the approach more generally, reviewing the original motivations, the resulting framework, and the central tenets of the underlying theory. We then evaluate the impact of PDP both on the field at large and within specific subdomains of cognitive science and consider the current role of PDP models within the broader landscape of contemporary theoretical frameworks in cognitive science. Looking to the future, we consider the implications for cognitive science of the recent success of machine learning systems called “deep networks”—systems that build on key ideas presented in the PDP volumes. (shrink)

Consciousness remains a mystery—“a phenomenon that people do not know how to think about—yet” (Dennett, , p. 21). Here, I consider how the connectionist perspective on information processing may help us progress toward the goal of understanding the computational principles through which conscious and unconscious processing differ. I begin by delineating the conceptual challenges associated with classical approaches to cognition insofar as understanding unconscious information processing is concerned, and to highlight several contrasting computational principles that are constitutive of the connectionist (...) approach. This leads me to suggest that conscious and unconscious processing are fundamentally connected, that is, rooted in the very same computational principles. I further develop a perspective according to which the brain continuously and unconsciously learns to redescribe its own activity itself based on constant interaction with itself, with the world, and with other minds. The outcome of such interactions is the emergence of internal models that are metacognitive in nature and that function so as to make it possible for an agent to develop a (limited, implicit, practical) understanding of itself. In this light, plasticity and learning are constitutive of what makes us conscious, for it is in virtue of our own experiences with ourselves and with other people that our mental life acquires its subjective character. The connectionist framework continues to be uniquely positioned in the Cognitive Sciences to address the challenge of identifying what one could call the “computational correlates of consciousness” (Mathis & Mozer, ) because it makes it possible to focus on the mechanisms through which information processing takes place. (shrink)

It is possible to learn multiple layers of non-linear features by backpropagating error derivatives through a feedforward neural network. This is a very effective learning procedure when there is a huge amount of labeled training data, but for many learning tasks very few labeled examples are available. In an effort to overcome the need for labeled data, several different generative models were developed that learned interesting features by modeling the higher order statistical structure of a set of input vectors. One (...) of these generative models, the restricted Boltzmann machine (RBM), has no connections between its hidden units and this makes perceptual inference and learning much simpler. More significantly, after a layer of hidden features has been learned, the activities of these features can be used as training data for another RBM. By applying this idea recursively, it is possible to learn a deep hierarchy of progressively more complicated features without requiring any labeled data. This deep hierarchy can then be treated as a feedforward neural network which can be discriminatively fine-tuned using backpropagation. Using a stack of RBMs to initialize the weights of a feedforward neural network allows backpropagation to work effectively in much deeper networks and it leads to much better generalization. A stack of RBMs can also be used to initialize a deep Boltzmann machine that has many hidden layers. Combining this initialization method with a new method for fine-tuning the weights finally leads to the first efficient way of training Boltzmann machines with many hidden layers and millions of weights. (shrink)

We identify a number of principles with respect to prediction that, we argue, underpin adult language comprehension: (a) comprehension consists in realizing a mapping between the unfolding sentence and the event representation corresponding to the real‐world event being described; (b) the realization of this mapping manifests as the ability to predict both how the language will unfold, and how the real‐world event would unfold if it were being experienced directly; (c) concurrent linguistic and nonlinguistic inputs, and the prior internal states (...) of the system, each drive the predictive process; (d) the representation of prior internal states across a representational substrate common to the linguistic and nonlinguistic domains enables the predictive process to operate over variable time frames and variable levels of representational abstraction. We review empirical data exemplifying the operation of these principles and discuss the relationship between prediction, event structure, thematic role assignment, and incrementality. (shrink)

Although for many years a sharp distinction has been made in language research between rules and words—with primary interest on rules—this distinction is now blurred in many theories. If anything, the focus of attention has shifted in recent years in favor of words. Results from many different areas of language research suggest that the lexicon is representationally rich, that it is the source of much productive behavior, and that lexically specific information plays a critical and early role in the interpretation (...) of grammatical structure. But how much information can or should be placed in the lexicon? This is the question I address here. I review a set of studies whose results indicate that event knowledge plays a significant role in early stages of sentence processing and structural analysis. This poses a conundrum for traditional views of the lexicon. Either the lexicon must be expanded to include factors that do not plausibly seem to belong there; or else virtually all information about word meaning is removed, leaving the lexicon impoverished. I suggest a third alternative, which provides a way to account for lexical knowledge without a mental lexicon. (shrink)

How does a child rapidly acquire and develop a structured mental organization for the vast number of words in the first years of life? How does a bilingual individual deal with the even more complicated task of learning and organizing two lexicons? It is only until recently have we started to examine the lexicon as a dynamical system with regard to its acquisition, representation, and organization. In this article, I outline a proposal based on our research that takes the dynamical (...) approach to the lexicon, and I discuss how this proposal can be applied to account for lexical organization, structural representation, and competition within and between languages. In particular, I provide computational evidence based on the DevLex model, a self‐organizing neural network model, and neuroimaging evidence based on functional magnetic resonance imaging (fMRI) studies, to illustrate how children and adults learn and represent the lexicon in their first and second languages. In the computational research, our goal has been to identify, through linguistically and developmentally realistic models, detailed cognitive mechanisms underlying the dynamic self‐organizing processes in monolingual and bilingual lexical development; in the neuroimaging research, our goal has been to identify the neural substrates that subserve lexical organization and competition in the monolingual and the bilingual brain. In both cases, our findings lead to a better understanding of the interactive dynamics involved in the acquisition and representation of one or multiple languages. (shrink)

One of the themes that has been widely addressed in both the implicit learning and statistical learning literatures is that of rule learning. While it is widely agreed that the extraction of regularities from the environment is a fundamental facet of cognition, there is still debate about the nature of rule learning. Rey and colleagues show that the comparison between human and non‐human primates can contribute important insights to this debate.

Butz, Achimova, Bilkey, and Knott provide a topic overview and discuss whether the special issue contributions may imply that event‐predictive abilities constitute a root for conceptual human thought, because they enable complex, mutually beneficial, but also intricately competitive, social interactions and language communication.

In typical statistical learning studies, researchers define sequences in terms of the probability of the next item in the sequence given the current item (or items), and they show that high probability sequences are treated as more familiar than low probability sequences. Existing accounts of these phenomena all assume that participants represent statistical regularities more or less as they are defined by the experimenters—as sequential probabilities of symbols in a string. Here we offer an alternative, or possibly supplementary, hypothesis. Specifically, (...) rather than identifying or labeling individual stimuli discretely in order to predict the next item in a sequence, we need only assume that the participant is able to represent the stimuli as evincing particular similarity relations to one another, with sequences represented as trajectories through this similarity space. We present experiments in which this hypothesis makes sharply different predictions from hypotheses based on the assumption that sequences are learned over discrete, labeled stimuli. We also present a series of simulation models that encode stimuli as positions in a continuous two‐dimensional space, and predict the next location from the current location. Although no model captures all of the data presented here, the results of three critical experiments are more consistent with the view that participants represent trajectories through similarity space rather than sequences of discrete labels under particular conditions. (shrink)

Children show a remarkable degree of consistency in learning some words earlier than others. What patterns of word usage predict variations among words in age of acquisition? We use distributional analysis of a naturalistic corpus of child‐directed speech to create quantitative features representing natural variability in word contexts. We evaluate two sets of features: One set is generated from the distribution of words into frames defined by the two adjacent words. These features primarily encode syntactic aspects of word usage. The (...) other set is generated from non‐adjacent co‐occurrences between words. These features encode complementary thematic aspects of word usage. Regression models using these distributional features to predict age of acquisition of 656 early‐acquired English words indicate that both types of features improve predictions over simpler models based on frequency and appearance in salient or simple utterance contexts. Syntactic features were stronger predictors of children's production than comprehension, whereas thematic features were stronger predictors of comprehension. Overall, earlier acquisition was predicted by features representing frames that select for nouns and verbs, and by thematic content related to food and face‐to‐face play topics; later acquisition was predicted by features representing frames that select for pronouns and question words, and by content related to narratives and object play. (shrink)

The computations involved in statistical learning have long been debated. Here, we build on work suggesting that a basic memory process, chunking, may account for the processing of statistical regularities into larger units. Drawing on methods from the memory literature, we developed a novel paradigm to test statistical learning by leveraging a robust phenomenon observed in serial recall tasks: that short‐term memory is fundamentally shaped by long‐term distributional learning. In the statistically induced chunking recall (SICR) task, participants are exposed to (...) an artificial language, using a standard statistical learning exposure phase. Afterward, they recall strings of syllables that either follow the statistics of the artificial language or comprise the same syllables presented in a random order. We hypothesized that if individuals had chunked the artificial language into word‐like units, then the statistically structured items would be more accurately recalled relative to the random controls. Our results demonstrate that SICR effectively captures learning in both the auditory and visual modalities, with participants displaying significantly improved recall of the statistically structured items, and even recall specific trigram chunks from the input. SICR also exhibits greater test–retest reliability in the auditory modality and sensitivity to individual differences in both modalities than the standard two‐alternative forced‐choice task. These results thereby provide key empirical support to the chunking account of statistical learning and contribute a valuable new tool to the literature. (shrink)

The pervasive use of distributional semantic models or word embeddings for both cognitive modeling and practical application is because of their remarkable ability to represent the meanings of words. However, relatively little effort has been made to explore what types of information are encoded in distributional word vectors. Knowing the internal knowledge embedded in word vectors is important for cognitive modeling using distributional semantic models. Therefore, in this paper, we attempt to identify the knowledge encoded in word vectors by conducting (...) a computational experiment using Binder et al.'s (2016) featural conceptual representations based on neurobiologically motivated attributes. In an experiment, these conceptual vectors are predicted from text‐based word vectors using a neural network and linear transformation, and prediction performance is compared among various types of information. The analysis demonstrates that abstract information is generally predicted more accurately by word vectors than perceptual and spatiotemporal information, and specifically, the prediction accuracy of cognitive and social information is higher. Emotional information is also found to be successfully predicted for abstract words. These results indicate that language can be a major source of knowledge about abstract attributes, and they support the recent view that emphasizes the importance of language for abstract concepts. Furthermore, we show that word vectors can capture some types of perceptual and spatiotemporal information about concrete concepts and some relevant word categories. This suggests that language statistics can encode more perceptual knowledge than often expected. (shrink)

Much previous work has suggested that word order preferences across languages can be explained by the dependency distance minimization constraint (Ferrer‐i Cancho, 2008, 2015; Hawkins, 1994). Consistent with this claim, corpus studies have shown that the average distance between a head (e.g., verb) and its dependent (e.g., noun) tends to be short cross‐linguistically (Ferrer‐i Cancho, 2014; Futrell, Mahowald, & Gibson, 2015; Liu, Xu, & Liang, 2017). This implies that on average languages avoid inefficient or complex structures for simpler structures. But (...) a number of studies in psycholinguistics (Konieczny, 2000; Levy & Keller, 2013; Vasishth, Suckow, Lewis, & Kern, 2010) show that the comprehension system can adapt to the typological properties of a language, for example, verb‐final order, leading to more complex structures, for example, having longer linear distance between a head and its dependent. In this paper, we conduct a corpus study for a group of 38 languages, which were either Subject–Verb–Object (SVO) or Subject–Object–Verb (SOV), in order to investigate the role of word order typology in determining syntactic complexity. We present results aggregated across all dependency types, as well as for specific verbal (objects, indirect objects, and adjuncts) and nonverbal (nominal, adjectival, and adverbial) dependencies. The results suggest that dependency distance in a language is determined by the default word order of a language, and crucially, the direction of a dependency (whether the head precedes the dependent or follows it; e.g., whether the noun precedes the verb or follows it). Particularly we show that in SOV languages (e.g., Hindi, Korean) as well as SVO languages (e.g., English, Spanish), longer linear distance (measured as number of words) between head and dependent arises in structures when they mirror the default word order of the language. In addition to showing results on linear distance, we also investigate the influence of word order typology on hierarchical distance (HD; measured as number of heads between head and dependent). The results for HD are similar to that of linear distance. At the same time, in comparison to linear distance, the influence of adaptability on HD seems less strong. In particular, the results show that most languages tend to avoid greater structural depth. Together, these results show evidence for “limited adaptability” to the default word order preferences in a language. Our results support a large body of work in the processing literature that highlights the importance of linguistic exposure and its interaction with working memory constraints in determining sentence complexity. Our results also point to the possible role of other factors such as the morphological richness of a language and a multifactor account of sentence complexity remains a promising area for future investigation. (shrink)

Since its inception, Shannon's information theory has attracted interest for the study of language and music. Recently, a wide range of converging studies have shown how efficient communication pervades language, from phonetics to syntax. Efficient principles imply that more resources should be assigned to highly informative items. For instance, average information content was shown to be a better predictor of word length than frequency, revisiting the famous Zipf's law. However, in spite of the success of the efficient communication framework in (...) the study of language and speech, very little work has investigated its relevance in the analysis of music. Here, we examine the organization of harmonic information in two large corpora of Western music, one made of MIDI files directly sequenced from scores, and the other made of MIDI recordings of live performances of highly skilled piano players. We show that there is a clear positive relationship between (contextual) information content of harmonic sequences and two essential musical properties, namely duration and loudness: the more unexpected a harmonic event is, the longer and the louder it is. (shrink)

The ability to combine words into novel sentences has been used to argue that humans have symbolic language production abilities. Critiques of connectionist models of language often center on the inability of these models to generalize symbolically (Fodor & Pylyshyn, 1988; Marcus, 1998). To address these issues, a connectionist model of sentence production was developed. The model had variables (role‐concept bindings) that were inspired by spatial representations (Landau & Jackendoff, 1993). In order to take advantage of these variables, a novel (...) dual‐pathway architecture with event semantics is proposed and shown to be better at symbolic generalization than several variants. This architecture has one pathway for mapping message content to words and a separate pathway that enforces sequencing constraints. Analysis of the model's hidden units demonstrated that the model learned different types of information in each pathway, and that the model's compositional behavior arose from the combination of these two pathways. The model's ability to balance symbolic and statistical behavior in syntax acquisition and to model aphasic double dissociations provided independent support for the dual‐pathway architecture. (shrink)

Syllogisms (e.g. “All A are B; All B are C; What is true about A and C?”) are a long‐studied area of human reasoning. Riesterer, Brand, and Ragni compare a variety of models to human performance and show that not only do current models have a lot of room for improvement, but more importantly a large part of this improvement must come from examining individual differences in performance.

Zuidema et al. illustrate how empirical AGL studies can benefit from computational models and techniques. Computational models can help clarifying theories, and thus in delineating research questions, but also in facilitating experimental design, stimulus generation, and data analysis. The authors show, with a series of examples, how computational modeling can be integrated with empirical AGL approaches, and how model selection techniques can indicate the most likely model to explain experimental outcomes.

This paper introduces coupled clustering—a novel computational framework for detecting corresponding themes in unstructured data. Gaining its inspiration from the structure mapping theory, our framework utilizes unsupervised statistical learning tools for automatic construction of aligned representations reflecting the context of the particular mapping being made. The coupled clustering algorithm is demonstrated and evaluated through detecting conceptual correspondences in textual corpora. In its current phase, the method is primarily oriented towards context-dependent feature-based similarity. However, it is preliminary demonstrated how it could (...) be utilized for identification of relational commonalities, as well. (shrink)

This paper discusses different approaches incognitive science and artificial intelligenceresearch from the perspective of radicalconstructivism, addressing especially theirrelation to the biologically based theories ofvon Uexküll, Piaget as well as Maturana andVarela. In particular recent work in New AI and adaptive robotics on situated and embodiedintelligence is examined, and we discuss indetail the role of constructive processes asthe basis of situatedness in both robots andliving organisms.

Many of the computing systems programmed using Machine Learning are opaque: it is difficult to know why they do what they do or how they work. Explainable Artificial Intelligence aims to develop analytic techniques that render opaque computing systems transparent, but lacks a normative framework with which to evaluate these techniques’ explanatory successes. The aim of the present discussion is to develop such a framework, paying particular attention to different stakeholders’ distinct explanatory requirements. Building on an analysis of “opacity” from (...) philosophy of science, this framework is modeled after accounts of explanation in cognitive science. The framework distinguishes between the explanation-seeking questions that are likely to be asked by different stakeholders, and specifies the general ways in which these questions should be answered so as to allow these stakeholders to perform their roles in the Machine Learning ecosystem. By applying the normative framework to recently developed techniques such as input heatmapping, feature-detector visualization, and diagnostic classification, it is possible to determine whether and to what extent techniques from Explainable Artificial Intelligence can be used to render opaque computing systems transparent and, thus, whether they can be used to solve the Black Box Problem. (shrink)

Many of the computing systems programmed using Machine Learning are opaque: it is difficult to know why they do what they do or how they work. Explainable Artificial Intelligence aims to develop analytic techniques that render opaque computing systems transparent, but lacks a normative framework with which to evaluate these techniques’ explanatory successes. The aim of the present discussion is to develop such a framework, paying particular attention to different stakeholders’ distinct explanatory requirements. Building on an analysis of “opacity” from (...) philosophy of science, this framework is modeled after accounts of explanation in cognitive science. The framework distinguishes between the explanation-seeking questions that are likely to be asked by different stakeholders, and specifies the general ways in which these questions should be answered so as to allow these stakeholders to perform their roles in the Machine Learning ecosystem. By applying the normative framework to recently developed techniques such as input heatmapping, feature-detector visualization, and diagnostic classification, it is possible to determine whether and to what extent techniques from Explainable Artificial Intelligence can be used to render opaque computing systems transparent and, thus, whether they can be used to solve the Black Box Problem. (shrink)

Some argue the common practice of inferring multiple processes or systems from a dissociation is flawed. One proposed solution is state-trace analysis, which involves plotting, across two or more conditions of interest, performance measured by either two dependent variables, or two conditions of the same dependent measure. The resulting analysis is considered to provide evidence that either a single process underlies performance or there is evidence for more than one process. This article reports simulations using the simple recurrent network in (...) which changes to the learning rate produced state-trace plots with multiple functions. We also report simulations using a single-layer error-correcting network that generate plots with a single function. We argue that the presence of different functions on a state-trace plot does not necessarily support a dual-system account, at least as typically defined ; it can also indicate variation in a single parameter within theories generally considered to be single-system accounts. (shrink)

Figure 1: A pr ototyp ical exa mple of a three-layer feed forward network, used by Plunkett and M archm an (1 991 ) to simulate learning the past-tense of En glish verbs. The inpu t units encode representations of the three phonemes of the present tense of the artificial words used in this simulation. Th e netwo rk is trained to produce a representation of the phonemes employed in the past tense form and the suffix (/d/, /ed/, or /t/) (...) used on regular verbs. To run the network, each input unit is assigned an activation value o f 0 or 1 , dep ending on whethe r the featu re is present or not. Eac h input unit is conne cted to each of the 30 hidden units by a we ighted conn ection and p rovid es an inp ut to each hidden unit equal to the product of the input unit's activation and the weight. Each hidd en unit's activation is then determined by summing ov er the va lues co ming fro m each inp ut unit to deter mine a netinput, and then applying a non-linear function (e.g., the logistic function 1/(1+e^netinput). Th is whole proced ure is. (shrink)

Neural network models have recently made striking progress in natural language processing, but they are typically trained on orders of magnitude more language input than children receive. What can these neural networks, which are primarily distributional learners, learn from a naturalistic subset of a single child's experience? We examine this question using a recent longitudinal dataset collected from a single child, consisting of egocentric visual data paired with text transcripts. We train both language-only and vision-and-language neural networks and analyze the (...) linguistic knowledge they acquire. In parallel with findings from Jeffrey Elman's seminal work, the neural networks form emergent clusters of words corresponding to syntactic (nouns, transitive and intransitive verbs) and semantic categories (e.g., animals and clothing), based solely on one child's linguistic input. The networks also acquire sensitivity to acceptability contrasts from linguistic phenomena, such as determiner-noun agreement and argument structure. We find that incorporating visual information produces an incremental gain in predicting words in context, especially for syntactic categories that are comparatively more easily grounded, such as nouns and verbs, but the underlying linguistic representations are not fundamentally altered. Our findings demonstrate which kinds of linguistic knowledge are learnable from a snapshot of a single child's real developmental experience. (shrink)

With a focus in the financial market, stock market dynamics forecasting has received much attention. Predicting stock market fluctuations is usually challenging due to the nonlinear and nonstationary time series of stock prices. The Elman recurrent network is renowned for its capability of dealing with dynamic information, which has made it a successful application to predicting. We developed a hybrid approach which combined Elman recurrent network with factorization machine technique, i.e., the FM-Elman neural network, to predict stock market volatility. In (...) this paper, the Standard & Poor’s 500 Composite Stock Price index, the Dow Jones industrial average index, the Shanghai Stock Exchange Composite index, and the Shenzhen Securities Component Index were used to demonstrate the validity of our proposed FM-Elman model in time-series prediction. The results were compared with predictions obtained from the other two models which are basic BP neural network and the Elman neural network. Some experiments showed that the FM-Elman model outperforms others through different accuracy measures. Furthermore, the effects of volatility degree on prediction performance from different stock indexes were investigated. An interesting phenomenon had been found through some numerical experiments on the effects of different user-specified dimensions on the proposed FM-Elman neural network. (shrink)