Deep learning models are powerful tools for representing the complex learning processes and decision-making strategies used by humans. Such neural network models make fewer assumptions about the underlying mechanisms thus providing experimental flexibility in terms of applicability. However, this comes at the cost of involving a larger number of parameters requiring significantly more data for effective learning. This presents practical challenges given that most cognitive experiments involve relatively small numbers of subjects. Laboratory collaborations are a natural way to increase overall (...) dataset size. However, data sharing barriers between laboratories as necessitated by data protection regulations encourage the search for alternative methods to enable collaborative data science. Distributed learning, especially federated learning, which supports the preservation of data privacy, is a promising method for addressing this issue. To verify the reliability and feasibility of applying FL to train neural networks models used in the characterization of decision making, we conducted experiments on a real-world, many-labs data pool including experiment data-sets from ten independent studies. The performance of single models trained on single laboratory data-sets was poor. This unsurprising finding supports the need for laboratory collaboration to train more reliable models. To that end we evaluated four collaborative approaches. The first approach represents conventional centralized learning and is the optimal approach but requires complete sharing of data which we wish to avoid. The results however establish a benchmark for the other three approaches, federated learning, incremental learning, and cyclic incremental learning. We evaluate these approaches in terms of prediction accuracy and capacity to characterize human decision-making strategies. The FL-based model achieves performance most comparable to that of the CL-based model. This indicates that FL has value in scaling data science methods to data collected in computational modeling contexts when data sharing is not convenient, practical or permissible. (shrink)

Cognitive Science, Volume 46, Issue 4, April 2022.

This article probes the practical ethical implications of AI system design by reconsidering the important topic of bias in the datasets used to train autonomous intelligent systems. The discussion draws on recent work concerning behaviour-guiding technologies, and it adopts a cautious form of technological utopianism by assuming it is potentially beneficial for society at large if AI systems are designed to be comparatively free from the biases that characterise human behaviour. However, the argument presented here critiques the common well-intentioned requirement (...) that, in order to achieve this, all such datasets must be debiased prior to training. By focusing specifically on gender-bias in Neural Machine Translation systems, three automated strategies for the removal of bias are considered – downsampling, upsampling, and counterfactual augmentation – and it is shown that systems trained on datasets debiased using these approaches all achieve general translation performance that is much worse than a baseline system. In addition, most of them also achieve worse performance in relation to metrics that quantify the degree of gender bias in the system outputs. By contrast, it is shown that the technique of domain adaptation can be effectively deployed to debias existing NMT systems after they have been fully trained. This enables them to produce translations that are quantitatively far less biased when analysed using gender-based metrics, but which also achieve state-of-the-art general performance. It is hoped that the discussion presented here will reinvigorate ongoing debates about how and why bias can be most effectively reduced in state-of-the-art AI systems. (shrink)

A basic deep neural network (DNN) is trained to exhibit a large set of input–output dispositions. While being a good model of the way humans perform some tasks automatically, without deliberative reasoning, more is needed to approach human‐like artificial intelligence. Analysing recent additions brings to light a distinction between two fundamentally different styles of computation: content‐specific and non‐content‐specific computation (as first defined here). For example, deep episodic RL networks draw on both. So does human conceptual reasoning. Combining the two takes (...) advantage of the complementary costs and benefits of each. It also offers a better model of human cognitive competence. (shrink)

Why can’t we keep as many items as we want in working memory? It has long been debated whether this resource limitation is a bug or instead a feature. We propose that the resource limitation is a consequence of a useful feature. Specifically, we propose that flexible cognition requires time-based binding, and time-based binding necessarily limits the number of memoranda that can be stored simultaneously. Time-based binding is most naturally instantiated via neural oscillations, for which there exists ample experimental evidence. (...) We report simulations that illustrate this theory and that relate it to empirical data. We also compare the theory to several other resource theories. (shrink)

Interest in computerised central economic planning (CCEP) has seen a resurgence, as there is strong demand for an alternative vision to modern free (or not so free) market liberal capitalism. Given the close links of CCEP with what we would now broadly call artificial intelligence (AI)—e.g. optimisation, game theory, function approximation, machine learning, automated reasoning—it is reasonable to draw direct analogues and perform an analysis that would help identify what commodities and institutions we should see for a CCEP programme to (...) become successful. Following this analysis, we conclude that a CCEP economy would need to have a very different outlook from current market practices, with a focus on producing basic “interlinking” commodities (e.g. tools, processed materials, instruction videos) that consumers can use as a form of collective R &D. On an institutional level, CCEP should strive for the release of basic commodities that empower consumers by having as many alternative uses as possible, but also making sure that a baseline of basic necessities is widely available. (shrink)

The article examines the question of how learning multiple tasks interacts with neural architectures and the flow of information through those architectures. It approaches the question by using the idealization of an artificial neural network where it is possible to ask more precise questions about the effects of modular versus nonmodular architectures as well as the effects of sequential versus simultaneous learning of tasks. A prior work has demonstrated a clear advantage of modular architectures when the two tasks must be (...) learned at the same time from the start, but this advantage may disappear when one task is first learned to a criterion before the second task is undertaken. Indeed, in some cases of sequential learning, nonmodular networks achieve success levels comparable to those of modular networks. In particular, if a nonmodular network is to learn two tasks of different difficulty and the more difficult task is presented first and learned to a criterion, then the network will learn the second, easier one without permanent degradation of the first one. In contrast, if the easier task is learned first, a nonmodular task may perform significantly less well than a modular one. It seems that the reason for this difference has to do with the fact that the sequential presentation of the more difficult task first minimizes interference between the two tasks. More broadly, the studies summarized in this article seem to imply that no single learning architecture is optimal for all situations. (shrink)

The default view in the epistemology of forgetting is that human memory would be epistemically better if we were not so susceptible to forgetting—that forgetting is in general a cognitive vice. In this paper, I argue for the opposed view: normal human forgetting—the pattern of forgetting characteristic of cognitively normal adult human beings—approximates a virtue located at the mean between the opposed cognitive vices of forgetting too much and remembering too much. I argue, first, that, for any finite cognizer, a (...) certain pattern of forgetting is necessary if her memory is to perform its function well. I argue, second, that, by eliminating clutter from her memory store, this pattern of forgetting improves the overall shape of the subject’s total doxastic state. I conclude by reviewing work in psychology which suggests that normal human forgetting approximates this virtuous pattern of forgetting. (shrink)

According to the age-of-acquisition hypothesis, words acquired early in life are processed faster and more accurately than words acquired later. Connectionist models have begun to explore the influence of the age/order of acquisition of items (and also their frequency of encounter). This study attempts to reconcile two different methodological and theoretical approaches (proposed by Lambon Ralph & Ehsan, 2006 and Zevin & Seidenberg, 2002) to age-limited learning effects. The current simulations extend the findings reported by Zevin and Seidenberg (2002) that (...) have shown that frequency trajectories (FTs) have limited and specific effects on word-reading tasks. Using the methodological framework proposed by Lambon Ralph and Ehsan (2006), which makes it possible to compare word-reading and picture-naming tasks in connectionist networks, we were able to show that FT has a considerable influence on age-limited learning effects in a picture naming task. The findings show that when the input–output mappings are arbitrary (simulating picture naming tasks), the links formed by the network become entrenched as a result of early experience and that subsequent variations in frequency of exposure of the items have only a minor impact. In contrast, when the mappings between input-output are quasi-systematic or systematic (simulating word-reading tasks), the training of new items was generalized and resulted in the suppression of age-limited learning effects. At a theoretical level, we suggest that FT, which simultaneously takes account of time and the level of exposure across time, represents a more precise and modulated measure compared with the order of introduction of the items and may lead to innovative hypotheses in the field of age-limited learning effects. (shrink)

Three‐ to 4‐month‐old infants show asymmetric exclusivity in the acquisition of cat and dog perceptual categories. The cat perceptual category excludes dog exemplars, but the dog perceptual category does not exclude cat exemplars. We describe a connectionist autoencoder model of perceptual categorization that shows the same asymmetries as infants. The model predicts the presence of asymmetric retroactive interference when infants acquire cat and dog categories sequentially. A subsequent experiment conducted with 3‐ to 4‐month‐olds verifies the predicted pattern of looking time (...) behaviors. We argue that bottom‐up, associative learning systems with distributed representations are appropriate for modeling the operation of short‐term visual memory in early perceptual category learning. (shrink)

A substantial body of experimental evidence has demonstrated that labels have an impact on infant categorization processes. Yet little is known regarding the nature of the mechanisms by which this effect is achieved. We distinguish between two competing accounts: supervised name‐based categorization and unsupervised feature‐based categorization. We describe a neurocomputational model of infant visual categorization, based on self‐organizing maps, that implements the unsupervised feature‐based approach. The model successfully reproduces experiments demonstrating the impact of labeling on infant visual categorization reported in (...) Plunkett, Hu, and Cohen (2008). It mimics infant behavior in both the familiarization and testing phases of the procedure, using a training regime that involves only single presentations of each stimulus and using just 24 participant networks per experiment. The model predicts that the observed behavior in infants is due to a transient form of learning that might lead to the emergence of hierarchically organized categorical structure and that the impact of labels on categorization is influenced by the perceived similarity and the sequence in which the objects are presented. The results suggest that early in development, say before 12 months old, labels need not act as invitations to form categories nor highlight the commonalities between objects, but they may play a more mundane but nevertheless powerful role as additional features that are processed in the same fashion as other features that characterize objects and object categories. (shrink)

Neural network models of memory are notorious for catastrophic interference: Old items are forgotten as new items are memorized (French, 1999; McCloskey & Cohen, 1989). While working memory (WM) in human adults shows severe capacity limitations, these capacity limitations do not reflect neural network style catastrophic interference. However, our ability to quickly apprehend the numerosity of small sets of objects (i.e., subitizing) does show catastrophic capacity limitations, and this subitizing capacity and WM might reflect a common capacity. Accordingly, computational investigations (...) (Knops, Piazza, Sengupta, Eger & Melcher, 2014; Sengupta, Surampudi & Melcher, 2014) suggest that mutual inhibition among neurons can explain both kinds of capacity limitations as well as why our ability to estimate the numerosity of larger sets is limited according to a Weber ratio signature. Based on simulations with a saliency map‐like network and mathematical proofs, we provide three results. First, mutual inhibition among neurons leads to catastrophic interference when items are presented simultaneously. The network can remember a limited number of items, but when more items are presented, the network forgets all of them. Second, if memory items are presented sequentially rather than simultaneously, the network remembers the most recent items rather than forgetting all of them. Hence, the tendency in WM tasks to sequentially attend even to simultaneously presented items might not only reflect attentional limitations, but also an adaptive strategy to avoid catastrophic interference. Third, the mean activation level in the network can be used to estimate the number of items in small sets, but it does not accurately reflect the number of items in larger sets. Rather, we suggest that the Weber ratio signature of large number discrimination emerges naturally from the interaction between the limited precision of a numeric estimation system and a multiplicative gain control mechanism. (shrink)

As modern deep networks become more complex, and get closer to human‐like capabilities in certain domains, the question arises as to how the representations and decision rules they learn compare to the ones in humans. In this work, we study representations of sentences in one such artificial system for natural language processing. We first present a diagnostic test dataset to examine the degree of abstract composable structure represented. Analyzing performance on these diagnostic tests indicates a lack of systematicity in representations (...) and decision rules, and reveals a set of heuristic strategies. We then investigate the effect of training distribution on learning these heuristic strategies, and we study changes in these representations with various augmentations to the training set. Our results reveal parallels to the analogous representations in people. We find that these systems can learn abstract rules and generalize them to new contexts under certain circumstances—similar to human zero‐shot reasoning. However, we also note some shortcomings in this generalization behavior—similar to human judgment errors like belief bias. Studying these parallels suggests new ways to understand psychological phenomena in humans as well as informs best strategies for building artificial intelligence with human‐like language understanding. (shrink)

Memory is fleeting. New material rapidly obliterates previous material. How, then, can the brain deal successfully with the continual deluge of linguistic input? We argue that, to deal with this “Now-or-Never” bottleneck, the brain must compress and recode linguistic input as rapidly as possible. This observation has strong implications for the nature of language processing: (1) the language system must “eagerly” recode and compress linguistic input; (2) as the bottleneck recurs at each new representational level, the language system must build (...) a multilevel linguistic representation; and (3) the language system must deploy all available information predictively to ensure that local linguistic ambiguities are dealt with “Right-First-Time”; once the original input is lost, there is no way for the language system to recover. This is “Chunk-and-Pass” processing. Similarly, language learning must also occur in the here and now, which implies that language acquisition is learning to process, rather than inducing, a grammar. Moreover, this perspective provides a cognitive foundation for grammaticalization and other aspects of language change. Chunk-and-Pass processing also helps explain a variety of core properties of language, including its multilevel representational structure and duality of patterning. This approach promises to create a direct relationship between psycholinguistics and linguistic theory. More generally, we outline a framework within which to integrate often disconnected inquiries into language processing, language acquisition, and language change and evolution. (shrink)

If human language must be squeezed through a narrow cognitive bottleneck, what are the implications for language processing, acquisition, change, and structure? In our target article, we suggested that the implications are far-reaching and form the basis of an integrated account of many apparently unconnected aspects of language and language processing, as well as suggesting revision of many existing theoretical accounts. With some exceptions, commentators were generally supportive both of the existence of the bottleneck and its potential implications. Many commentators (...) suggested additional theoretical and linguistic nuances and extensions, links with prior work, and relevant computational and neuroscientific considerations; some argued for related but distinct viewpoints; a few, though, felt traditional perspectives were being abandoned too readily. Our response attempts to build on the many suggestions raised by the commentators and to engage constructively with challenges to our approach. (shrink)