Comparisons of some learning models for response bias in signal detection

The effects of variations in signal probability and varying degrees of correct feedback on response bias were studied in a yes-no auditory signal detection task. The main finding was that the bias towards saying yes was an increasing function of the frequency of signal feedback events, but did not depend on the correctness of the feedback. Several learning models coupled with a simple psychophysical and decision model yielded predictions about overall biases and certain sequential statistics. Only one model, which can be decribed as an “informational” model, gave a good account of both observed overall biases and sequential statistics. This model assumes the observer’s response bias is strengthened for the feedback-reinforced response when the observer’s sensory information is ambiguous or is contradicted by the feedback information.     

Incorporating prior biases innetwork models of conceptual rule learning

A series of simulations is reported in which extant formal categorization models are applied to human rule-learning data (Salatas & Bourne, 1974). These data show that there are clear differences in the ease with which humans learn rules, with the conjunctive the easiest and the biconditional the hardest. The original ALCOVE model (an exemplar-based model), a configuralcue model, and two-layer backpropagation models did not fit the rule-learning data. ALCOVE successfully fit the data, however, when prior biases observed in human rule learning were implemented into weights of the network. Thus, current empirical learning models may not fare well in situations in which learners enter the concept-formation situation with preconceived biases regarding the kinds of concepts that are possible, but such biases might nevertheless be captured within these models. By incorporating preexperimental biases, ALCOVE may hold promise as a comprehensive category-learning model.     

Decision criteria do not shift: Reply to Treisman

Recently, we argued that the detection criterion representation of decision-making biases, embedded within the theory of signal detection, is empirically testable and has, in fact, been falsified by empirical results from visual discrimination experiments. Treisman (2002) attempts to show that there is an alternative interpretation of our results that could explain them without dropping the detection criterion construct. In lieu of attempting to fit the data with a model, however, he gives two kinds of theoretical examples, both involving manipulations of the spacing of criteria on a decision axis. The first example correctly predicts that the bias estimate we developed will be zero but does so by assuming zero spacing between some criteria (some rating responses are never used). We did not observe zero spacing between any criteria and did not perform any analyses on responses that never occurred. Moreover, this example does not explain why the upper-bound bias estimates that we obtained by combining results from two criteria placements were also trivially small. His second example predicts that the bias should have been detectable with sufficiently large sample sizes. In our experiments, the sample sizes were, in fact, quite large, large enough for the results to be consistent in 18 different experimental conditions. Finally, all of Treisman’s criteria placement examples also fail to explain the pronounced effects of base rates on the shapes of the rating ROC curves, and his suggestion that there are problems of logical interpretation with our proposed distribution model ignores the predictions of large classes of alternatives to detection theory, including the dynamic models of perception.     

Language Evolution Can Be Shaped by the Structure of the World

Human languages vary in many ways but also show striking cross‐linguistic universals. Why do these universals exist? Recent theoretical results demonstrate that Bayesian learners transmitting language to each other through iterated learning will converge on a distribution of languages that depends only on their prior biases about language and the quantity of data transmitted at each point; the structure of the world being communicated about plays no role (Griffiths & Kalish, 2005 , 2007 ). We revisit these findings and show that when certain assumptions about the relationship between language and the world are abandoned, learners will converge to languages that depend on the structure of the world as well as their prior biases. These theoretical results are supported with a series of experiments showing that when human learners acquire language through iterated learning, the ultimate structure of those languages is shaped by the structure of the meanings to be communicated.     

The Parallel Episodic Processing (PEP) model: Dissociating contingency and conflict adaptation in the item-specific proportion congruent paradigm

The present work introduces a computational model, the Parallel Episodic Processing (PEP) model, which demonstrates that contingency learning achieved via simple storage and retrieval of episodic memories can explain the item-specific proportion congruency effect in the colour-word Stroop paradigm. The current work also presents a new experimental procedure to more directly dissociate contingency biases from conflict adaptation (i.e., proportion congruency). This was done with three different types of incongruent words that allow a comparison of: (a) high versus low contingency while keeping proportion congruency constant, and (b) high versus low proportion congruency while keeping contingency constant. Results demonstrated a significant contingency effect, but no effect of proportion congruence. It was further shown that the proportion congruency associated with the colour does not matter, either. Thus, the results quite directly demonstrate that ISPC effects are not due to conflict adaptation, but instead to contingency learning biases.     

The evolution of frequency distributions: Relating regularization to inductive biases through iterated learning

The regularization of linguistic structures by learners has played a key role in arguments for strong innate constraints on language acquisition, and has important implications for language evolution. However, relating the inductive biases of learners to regularization behavior in laboratory tasks can be challenging without a formal model. In this paper we explore how regular linguistic structures can emerge from language evolution by iterated learning, in which one person’s linguistic output is used to generate the linguistic input provided to the next person. We use a model of iterated learning with Bayesian agents to show that this process can result in regularization when learners have the appropriate inductive biases. We then present three experiments demonstrating that simulating the process of language evolution in the laboratory can reveal biases towards regularization that might not otherwise be obvious, allowing weak biases to have strong effects. The results of these experiments suggest that people tend to regularize inconsistent word-meaning mappings, and that even a weak bias towards regularization can allow regular languages to be produced via language evolution by iterated learning.     

Decision processes in discrimination: fundamental misrepresentations of signal detection theory.

A new approach to studying decision making in discrimination tasks is described that does not depend on the technical assumptions of signal detection theory (e.g., normality of the encoding distributions). In 3 different experiments, results of these new distribution-free tests converge on a single, surprising conclusion: response biases had substantial effects on the encoding distributions but no effect on the decision rule, which was uniformly unbiased in equal and unequal base rate conditions and in symmetric and asymmetric payoff conditions. This seemingly paradoxical result is fundamentally inconsistent with the entire family of signal detection theory models, raising some important questions about the significance of many published results in the human performance literature.     

Are unsupervised neural networks ignorant? Sizing the effect of environmental distributions on unsupervised learning

Learning environmental biases is a rational behavior: by using prior odds, Bayesian networks rapidly became a benchmark in machine learning. Moreover, a growing body of evidence now suggests that humans are using base rate information. Unsupervised connectionist networks are used in computer science for machine learning and in psychology to model human cognition, but it is unclear whether they are sensitive to prior odds. In this paper, we show that hard competitive learners are unable to use environmental biases while recurrent associative memories use frequency of exemplars and categories independently. Hence, it is concluded that recurrent associative memories are more useful than hard competitive networks to model human cognition and have a higher potential in machine learning.     

Consensus effects in categorization decisions

A previous study (Gilat et al., J. Exp. Psychol. Appl. 3 (1997) 83) has shown that the incentive to reach consensus can raise the tendency to rely on base rates in signal detection decisions and can reduce the probability that less likely events will be accurately classified. This phenomenon was named the “consensus effect”. The current study assesses the conditions under which this effect develops and in particular the effects of information about the game and of the incentive structure on the learning process. The results of three experiments show that the learning process slows when participants have information about the actual state of nature. This finding is captured by a reinforcement learning model with the assumption that information narrows the distribution of the initial propensities for choosing among cutoffs. The results are further evidence for the utility of the combination of learning models and analyses of cognitive processes for the prediction of decision making in situations involving multiple players.     

A Bayesian Model of Biases in Artificial Language Learning: The Case of a Word‐Order Universal

In this article, we develop a hierarchical Bayesian model of learning in a general type of artificial language‐learning experiment in which learners are exposed to a mixture of grammars representing the variation present in real learners’ input, particularly at times of language change. The modeling goal is to formalize and quantify hypothesized learning biases. The test case is an experiment ( Culbertson, Smolensky, & Legendre, 2012 ) targeting the learning of word‐order patterns in the nominal domain. The model identifies internal biases of the experimental participants, providing evidence that learners impose (possibly arbitrary) properties on the grammars they learn, potentially resulting in the cross‐linguistic regularities known as typological universals. Learners exposed to mixtures of artificial grammars tended to shift those mixtures in certain ways rather than others; the model reveals how learners’ inferences are systematically affected by specific prior biases. These biases are in line with a typological generalization—Greenberg's Universal 18—which bans a particular word‐order pattern relating nouns, adjectives, and numerals.     

Call me Alix,not Elix: vowels are more important than consonants in own‐name recognition at 5 months

Consonants and vowels differ acoustically and articulatorily, but also functionally: Consonants are more relevant for lexical processing, and vowels for prosodic/syntactic processing. These functional biases could be powerful bootstrapping mechanisms for learning language, but their developmental origin remains unclear. The relative importance of consonants and vowels at the onset of lexical acquisition was assessed in French‐learning 5‐month‐olds by testing sensitivity to minimal phonetic changes in their own name. Infants’ reactions to mispronunciations revealed sensitivity to vowel but not consonant changes. Vowels were also more salient (on duration and intensity) but less distinct (on spectrally based measures) than consonants. Lastly, vowel (but not consonant) mispronunciation detection was modulated by acoustic factors, in particular spectrally based distance. These results establish that consonant changes do not affect lexical recognition at 5 months, while vowel changes do; the consonant bias observed later in development does not emerge until after 5 months through additional language exposure.     

The detection model of recognition using know and remember judgments

The signal detection model forknow andremember recognition judgments was tested in two experiments. In Experiment 1, two predictions of the model were tested: (1) that measures of memory sensitivity,A′, are equivalent in value when based on either the recognition (know or remember) criterion or on the remember criterion; and (2) that there is a positive correlation between recognition bias and the proportion of know judgments that are hits, but no correlation between recognition bias and proportion of remember hits (Donaldson, 1996). Both predictions were supported by the data. In Experiment 2, the context of test items was manipulated to make it more or less similar to learning context. The detection model requires that memory sensitivity be the same for both recognition and remember judgments, regardless of test context. Alternatively, if remember judgments reflect only the retrieval of episodic information from memory, the two measures of memory sensitivity should become more disparate in value as learning and test context are made more similar. Memory sensitivity was generally the same in value for recognition and remember criteria but different across context conditions, thus supporting the detection model. The nature of the memory continuum used in detection theory is also discussed.     

Testing a neurocomputational model of recollection, familiarity, and source recognition

The authors assess whether the complementary learning systems model of the medial temporal lobes (Norman & O'Reilly, 2003) is able to account for source recognition receiver operating characteristics (ROCs). The model assumes that recognition reflects the contribution of a hippocampally mediated recollection process and a cortically mediated familiarity process. The hippocampal process is found to produce threshold output functions that lead to U-shaped zROCs, whereas the cortical process produces Gaussian signal detection functions and linear zROCs. The model is consistent with several dual process theories of recognition and is capable of producing the types of zROCs observed in studies of item and source recognition. In addition, the model makes the novel prediction that as the level of feature similarity across items increases, the ability of the hippocampus to encode distinct representations for each stimulus will diminish, and the threshold nature of recollection will break down, leading source zROCs to become more linear. The authors conducted 3 new behavioral source experiments that confirmed the model's prediction. The results demonstrate that the model provides a viable account of item and source recognition performance.     

Performance biases and recognition memory for semantic and formal changes in connected discourse

In an experiment comparing memory for formal and semantic information, the confounding effects of attentional and response biases were controlled using an adaptation of Sachs' (1967, 1974) method. Subjects attempted to recognize semantic or formal changes in test sentences following short passages of connected discourse. Attentional biases were controlled by using a single type of change in an experimental session, and response biases were controlled with methods from signal detection theory. Semantic recognition scores were consistently above formal scores, both within subjects and within passages, indicating that superiority of semantic performance is attributable to differences in memorability rather than to biases favoring semantic performance. However, formal scores were above chance, suggesting that the poor memory for formal information, as reported previously, may have been due to performance factors.     

Characterizing perceptual learning with external noise

Performance in perceptual tasks often improves with practice. This effect is known as ‘perceptual learning,’ and it has been the source of a great deal of interest and debate over the course of the last century. Here, we consider the effects of perceptual learning within the context of signal detection theory. According to signal detection theory, the improvements that take place with perceptual learning can be due to increases in internal signal strength or decreases in internal noise. We used a combination of psychophysical techniques (external noise masking and double-pass response consistency) that involve corrupting stimuli with externally added noise to discriminate between the effects of changes in signal and noise as observers learned to identify sets of unfamiliar visual patterns. Although practice reduced thresholds by as much as a factor of 14, internal noise remained virtually fixed throughout training, indicating learning served to predominantly increase the strength of the internal signal. We further examined the specific nature of the changes that took place in signal strength by correlating the externally added noise with observer’s decisions across trials (response classification). This technique allowed us to visualize some of the changes that took place in the linear templates used by the observers as learning occurred, as well as test the predictions of a linear template-matching model. Taken together, the results of our experiments offer important new theoretical constraints on models of perceptual learning.     

Learning biases predict a word order universal

How recurrent typological patterns, or universals, emerge from the extensive diversity found across the world's languages constitutes a central question for linguistics and cognitive science. Recent challenges to a fundamental assumption of generative linguistics-that universal properties of the human language acquisition faculty constrain the types of grammatical systems which can occur-suggest the need for new types of empirical evidence connecting typology to biases of learners. Using an artificial language learning paradigm in which adult subjects are exposed to a mix of grammatical systems (similar to a period of linguistic change), we show that learners' biases mirror a word-order universal, first proposed by Joseph Greenberg, which constrains typological patterns of adjective, numeral, and noun ordering. We briefly summarize the results of a probabilistic model of the hypothesized biases and their effect on learning, and discuss the broader implications of the results for current theories of the origins of cross-linguistic word-order preferences.     

Induction of combination rules in two-dimensional function learning

Previous studies have typically found that when people learn to combine two dimensions of a stimulus to select a response, they learn additive combination rules more easily than nonadditive (e.g., multiplicative) ones. The present experiments demonstrate that in some situations people can learn multiplicative rules more easily than other (e.g., additive) rules. Subjects learned to produce specified response durations when presented with stimulus lines varying in length and angle of orientation. When stimuliand correct responses were related by a multiplicative combination of power functions, learning was relatively easy (Experiment 1). In contrast, systematic response biases occurred during the early phases of learning an additive combination of linear functions (Experiment 2) and a more complex (nonadditive and nonmultiplicative) combination of linear functions (Experiment3), suggesting that people have a tendency to induce a multiplicative combination of power functions. However, the initial biases decreased with practice. These results are explained in terms of a revised adaptive regression model of function learning originally proposed by Koh and Meyer (1991). Differences between the present results and previous results in the literature are discussed.     

Revealing human inductive biases for category learning by simulating cultural transmission

We explored people’s inductive biases in category learning—that is, the factors that make learning category structures easy or hard—using iterated learning. This method uses the responses of one participant to train the next, simulating cultural transmission and converging on category structures that people find easy to learn. We applied this method to four different stimulus sets, varying in the identifiability of their underlying dimensions. The results of iterated learning provide an unusually clear picture of people’s inductive biases. The category structures that emerge often correspond to a linear boundary on a single dimension, when such a dimension can be identified. However, other kinds of category structures also appear, depending on the nature of the stimuli. The results from this single experiment are consistent with previous empirical findings that were gleaned from decades of research into human category learning.