Funded by a gift from Robert J. Glushko, Adjunct Professor in the UC Berkeley School of Information, the Prize is intended to encourage students to pursue senior honors theses in cognitive science. At least two such prizes are awarded each year, in order to insure that excellence can be recognized in both empirical disciplines like cognitive neuroscience, and theoretical disciplines like philosophy. The prize inaugurated in 2006, includes a $500 cash award and a certificate.
A latent variable model for the learning of absence exceptions
Advisor: Tom Griffiths (Cognitive Science and Psychology)
Within language acquisition, little feedback regarding ungrammatical utterances is provided to the learner; instead, noisy positive data make up the majority of linguistic stimuli. From such an impoverished data set, valuable information lies in indirect negative evidence - absences from positive stimuli - which the learner could utilize to improve conceptual learning (without recourse to innate knowledge). Across all domains, the process of acquiring inductive generalizations about the world can be predicated on observations of solitary exceptions to expand a sparse under-generalization, or observations of a string of absences to contract an over-generalization. The cognitive procedure of determining adequate evidence for an absence through observation of repeated absence of evidence should vary in efficiency and accuracy depending on the salience of absences. The mechanism of learning such an 'absence exception' is tested using an experimental framework reminiscent of minesweeper, and probabilistically modeled with a latent variable model. In the first experiment, the efficacy of the experimental framework in manipulating subjects' perceptions of absence salience, reported as subjects' confidences in the appearance of an absence, is determined. In the second experiment, the effect of absence salience on the accuracy of generalizations is tested in a conceptual learning task, and results are compared to a latent variable model with two generative models, one based on sampling from rational distributions and the other on sampling from a user-generated confidence distribution. We find that the degree of sensitivity to absences is a factor in acquisition of inductive generalizations. Our conclusions further the understanding of language acquisition, inductive learning of rule-based concepts and formation of scientific theories by explicating a widely-used indirect inferential process by which hypotheses are acquired, and offering a statistical learning model for such a process.
Effects of learning on courtship behavior in male jumping spiders
Advisor: Lucia Jacobs (Psychology)
Courtship plasticity is a mechanism which can lead to sexual selection and speciation, and though mate choice trials in jumping spiders have provided insights into complex systems of sexual selection, little is known about how jumping spiders learn in this context. Jumping spider learning has been demonstrated in a variety of other contexts. Here, we test the hypothesis that prior experience with females influences the future courtship behavior of male red-backed jumping spiders (Phidippus johnsoni). After training courting males with females unresponsive to their advances, we found evidence that jumping spiders changed their behavior to accommodate learned information. In the experience condition, individual males were exposed to mounted mature females to whom they were unable to gain access over a day-long series of training trials. The control group had no prior experience with females. In a test trial following training trials, male spiders were placed in an arena with a live mature female. Courtship display lengths and attempt numbers were measured, and found to be reduced in males with prior (unsuccessful) courting experience. This study adds a cognitive component to the already immensely complex mating rituals of the jumping spider.
The role of sleep in emotional memory allocation
Advisor: Matthew P. Walker (Psychology)
Sleep after learning supports the consolidation of both neutral and emotional memory, but the influence of sleep on initial encoding remains largely unknown. Here, we test the hypothesis that a preceding period of sleep, relative to an equivalent time of wake, enables the preferential allocation of memory for events most emotionally salient, and that this benefit is mediated by rapid eye movement (REM) sleep. Thirty-three adults performed an encoding task, followed by a memory test, either in the morning following a night of sleep monitored by EEG (n=18, Sleep-group), or in the evening following a day of wake (n=15, Wake-group). During the task, participants viewed 150 pictures of varying emotionality, and rated their subjective emotional response to each image on a 5-point intensity scale. Both groups demonstrated equivalent overall memory, yet the allocation of memory across the ratings was markedly different between groups. Superior encoding of emotional relative to non-emotional items was evident following sleep, demonstrated by a significant increasing linear trend across ratings, but not following wake. Each group was then split into two subgroups, LowREM and HighREM, corresponding to the amount of prior REM sleep received. Within the Wake-group, the HighREM subgroup showed an increasing linear trend across the ratings, indicative of the importance of REM sleep in facilitating emotional memory allocation, a trend absent in the LowREM subgroup. Together, these findings support a role for sleep, especially REM, in restoring the optimal allocation of episodic encoding on the basis of emotional salience. This potentially adaptive evolutionary function of sleep has clinical implications in mood and anxiety disorders where sleep disruption and emotional memory imbalance co-occur.
Revealing culture-specific melodic expectations through iterated learning
Advisor: Tom Griffiths (Cognitive Science and Psychology)
Psychological theories of music appreciation assert that listeners form expectations based on the notes they previously heard. Recently, statistical analyses of music have identified the distributions of pitches and intervals in European folk melodies (Krumhansl, Toiviainen, Jarvinen, & Eerola, 1999) and demonstrated that Bayesian models using these distributions can infer structural properties of the melodies, such as their key (Temperley, 2008). This raises the question of whether people make similar inferences by internalizing these "natural statistics" of their musical environment. To explore this hypothesis, we used the pure iterated learning method (Kalish, Griffiths, and Lewandowsky, 2007) to estimate the prior distribution of participants over melodies. In this method, we simulate the process of cultural transmission of melodies. Each participant hears a set of melodies, and is asked to replace a muted note in each melody. The resulting melodies are then given to the next participant, with another note muted at random. If people fill in these notes according to Bayesian inference based on the notes they hear and their prior distribution, then this process will eventually converge on a set of melodies that are sampled from their prior (Griffiths & Kalish, 2005). The statistics of the melodies produced by our participants converge towards those of folk melodies over time, suggesting that people have internalized the structure of the music of their specific cultural milieu. Our results also indicate how cultural transmission of music can shape music to match people's expectations.
Layered sparse associative network for soft pattern classification and contextual pattern completion
Advisor: Fritz Sommer (Redwood Center for Theoretical Neuroscience)
Traditional models of associative networks have impressive pattern completion capabilities but are not useful as pattern classifiers beyond simple toy examples where the classes form well-separated clusters in pattern space. Here we propose a neurobiologically inspired sparse associative memory that is able to 1) be trained to recognize real-world patterns and 2) use memory content for reconstructing noisy input in a context-sensitive fashion. The network is built with a layered architecture and is trained with labeled handwritten digits. The input layers use random connectivity to form sparse representations of the visual and semantic input. Further, the network contains two synaptic matrices with Hebbian-type plasticity. One synaptic matrix learns maps between the sparse representation and the visual pattern itself (Matrix C). The other matrix uses a non-standard learning rule to associate between the sparse representations of visual inputs and labels (Matrix W). Activity flow between the two internal representations (through W) can be used to differentiate ambiguous visual input and to constrain image reconstruction to a particular semantic class. We used the MNIST database of handwritten digits collected by the U.S. Post Office to test our memory models on real world data. The dataset includes 60,000 training digits and 10,000 test digits, each a normalized and centered 28x28 pixel grayscale image.
The Effects of Pedagogy on the Fidelity of Intergenerational Knowledge Transmission
Advisor: Tom Griffiths (Cognitive Science and Psychology)
Cumulative cultural evolution is a process by which humans accumulate knowledge over multiple generations. The uniquely human ability to exploit pedagogical reasoning is hypothesized to account for the success of this process. We examined whether pedagogical reasoning meets the minimal preconditions to support cumulative cultural evolution. Knowledge accumulation essentially requires high fidelity knowledge transmission, and our goal was to show that pedagogy improves fidelity. We conducted a behavioral study using a function learning task within an iterated learning framework that simulates intergenerational transfer. This framework allows us to provide empirical evidence to support developing formal models. We found that pedagogy improves the fidelity of transfer between generations. This supports our prediction for the role pedagogy may play in cultural transmission. This result validates future computational models and motivates further research in uncovering the conditions under which cumulative cultural evolution occurs. This has strong implications for the ways that computers can potentially mediate the process of cultural evolution.
Biological Confounds De-Value Psychological Explanations
Advisor: Tania Lombrozo (Psychology)
Previous research has documented distinct types of explanations and the different ways that these explanations are interpreted and understood. We are interested in whether biological explanations de-value psychological and environmental explanations for a given phenomenon. We investigated this question in two experiments. Participants were presented with two sham studies. The initial sham study found support for a potential explanation for a phenomenon. Participants were then presented with a second sham study that supplied evidence for a different explanation for the same phenomenon. To measure the degree that different types of confounds interfered with the original explanation, we analyzed the shift in causal ratings of the initial explanation. In general, our results suggest that biological confounds have more undermining power than other types of confounds. However, no clear pattern emerged regarding differences between psychological and environmental explanations. Moreover, the question remains as to why biological confounds may have more undermining power than others. In addition, if biological confounds actually have more power, how should people compensate for this bias?
A Cognitive Linguistic Analysis of Parenting
Advisor: George Lakoff
Parenting philosophies in popular culture do not always reflect the findings of the latest empirical studies of cognitive development. These studies, some of which date back forty years, clearly demonstrate the positive effect of parent responsiveness, displays of affection, mutual respect, and firm discipline. The conclusions of these studies are relatively straightforward, and some psychologists have indeed attempted to relay these findings in books for parents. With such strong results, why do overly strict or overly lenient parenting attitudes still persist? To answer this question, I use tools from Cognitive Linguistics to analyze the deep linguistic structures beyond the three main parenting models: the Nurturant Parent, the Indulgent Parent, and the Strict Father. Contested concepts underlie our perceptions of morality. Image schemas and conceptual metaphors have a rich mapping of goal-oriented approaches. Additionally, we have cultural bias toward understanding all causation in terms of direct causation. This combination mutually inhibits both systemic causation and the process-oriented approach to parenting advocated by the empirical research. The latest developments in Cognitive Science illuminate in profound ways our innate drive to nurture and empower our children. If we construct the right image schemas and adapt the correct conceptual metaphors, then these exciting developments can be communicated in such a way that they gain traction in the popular consciousness.
Fearing the Unknown: Economic Preference Parameters Predict Aversion to Ambiguity
Advisor: Mark D'Esposito
The canonical, economic concepts of choice under risk, including assumptions of introspection and rationality, are being reexamined as evidence for automatic and affective systems in the brain are implicated in uncertain decision-making. Classical utility theory is sufficient for modeling risky decisions, where well-defined choices have explicit probabilities; however, in the absence of well-defined probabilities, or ambiguity, unique probabilities are not always calculated (Ellsberg, 1961). People often demonstrate aversion to ambiguity and will sacrifice potential rewards for the sake of certainty. In this present study, we examined the hypothesis that aversion to ambiguity may be driven by a perception of whether something is knowable or not. We created an Ellsberg-type gambling task, where subjects were asked to choose between two alternatives that varied in degree of risk or ambiguity (the amount of missing information) and calculated parameters that indexed each subjects’ tolerance for risky and ambiguous decisions. Critically, two kinds of ambiguous gambles were presented with either a fixed or variable probability, and separate parameters were calibrated using an a-MaxMin utility function. Even though the fixed and variable gambles were identical in composition, subjects were more likely to bet on the fixed probabilities, suggesting aversion to ambiguity is modulated by a perception of whether something can be learned (or ever known) and provides further support for the role of affective systems in economic decision-making tasks.
The Effect of Scarcity on Children’s Decisions
Advisor: Lori Markson
The present studies were motivated by empirical and anecdotal evidence suggesting the existence of a human preference for scarce items. A recent study found that 3- to 5-year-old children were more likely to choose a more scarce good over an abundant one (Markson & Padoa-Schioppa, submitted). In contrast, a similar study found that adults chose at chance levels when presented with the same two options. Here, we further explored the notion of a scarcity bias in children and adults. In one study, children failed to exhibit a scarcity bias when an unequal amount of goods were represented in the form of auditory stimuli. A second study with adults utilized a verbal shadowing task to limit their use of reasoning when making decisions. This study found that they continued to choose at chance levels. Taken together, the results highlight the need for further research concerning the origin of the scarcity bias and decision-making in children more generally.
The Role of Production in L2 Word Learning
Advisor: Carla Hudson-Kam
In this study we seek to broaden our understanding of the role of production in language learning, asking specifically if hard-to-pronounce words are more difficult to learn and if pronouncing a word out-loud has an effect on one’s ability to learn it. In Experiment 1, we taught adults to associate Polish words with novel objects, testing the prediction that producing words out loud would facilitate learning easy-to-pronounce words, but would make difficult-to-pronounce words harder to learn. On a test of recognition, easy-to-pronounce words did not improve with out-loud repetition, but difficult words did worsen compared to those who did not produce the words out-loud. Experiments 2 and 3 tested whether this effect resulted from actual articulation or from the distraction caused by drawing more attention to the difficult words, testing the effects of subvocal repetition and hearing another English speaker repeat the words. It found that those who repeated subvocally performed significantly better than those who repeated words out loud. Experiment 4 tested how interfering with the ability to mentally rehearse words affects learning by engaging subjects in articulatory suppression, with the surprising result that it did not negatively affect learning. In every experiment, pronunciation difficulty proved to have a significant effect on learning. These results suggest that knowing how to pronounce a word may have a key role in forming an initial mental representation.
Electroencephalogram Correlates of Momentary Mindfulness
Advisor: Eleanor Rosch
Despite widespread interest, few experiments have studied neural correlates of momentary mindfulness. To address this, electroencephalography with respirometry was collected from two groups trained in “focused attention”: mindfulness meditators and rowing athletes. Participants engaged in 1) a 50-minute task involving sustained attention to and counting of breath and 2) a “mind-wandering” control condition. Randomly-spaced tones interrupted the tasks and prompted participants to report their momentary attentional state (i.e., either focused on the breath or “mind wandering”). Reported breath count from trials was compared to respirometer data, allowing classification of trials into three conservative conditions for analysis: a) focused attention on breath with correct respiration count b) “mind wandering” from breath; and c) “mind wandering” during control condition. Analysis of pre-tone spectral power data suggests that momentary mindfulness is associated with alterations in alpha power over parietal electrodes, with different participants displaying different directions of power change.
Reference Frames in Figure-Ground Organization
Advisor: Stephen Palmer
Within what reference frames do orientation-related factors of figure-ground organization operate: retinal, environmental/gravitational, or object-based? We studied all known figure-ground cues that have an orientation component -- lower-region, wider base, shape familiarity, symmetry, and horizontal-vertical orientation – for their sensitivity to head tilt. Observers indicated which region appeared to be figure with their heads upright or tilted (45, 90, or 180 degrees from upright, depending on the particular factor) to dissociate different reference frames. The results indicate that retinal directions clearly dominate for lower region and wider base, but object-based directions dominate for shape familiarity. Symmetry showed only weak retinal and gravitational effects, and no clear directional preference was found for horizontal-vertical orientation. The paucity of evidence for gravitational reference frames is interesting not only because it conflicts with previous findings on shape perception, but because the ecological rationale for orientation sensitive figure-ground cues is based on gravitational considerations. The results are discussed as indicating that at least some figure-ground organization cues (e.g., lower region and wider base) operate before orientation constancy and that retinal directions may provide an evolutionarily useful surrogate for gravitational directions.
A Charming Little Cabernet: Effects of Label Design on Purchase Intent and Brand Personality of Wine
Advisor: Stephen Palmer
This thesis, completed for an Interdisciplinary Studies Field Major in Cognitive design in the Marketplace, examined the impact of brand personality on purchase intent, and the influence of three elements of packaging design as an antecedent of brand personality. In the study, 262 subjects made brand personality judgments and rated their purchase intent for a total of 90 experimental wine labels, which varied along three dimensions chosen on the basis of a pilot study: color (six colors), imagery (picture of a chateau or vineyard; grape motif; coat-of-arms; an elk, a traditional animal; or a platypus, an unusual animal) with or without a picture of a deer) and design layout (traditional with white background, traditional with full color, or modern with half unprinted, half color background). Brand personality explained nearly half of the variance in purchase intent, with the facets successful, charming, spirited, and up-to-date most strongly correlated with purchase intent. Of the three dimensions of visual design studied, the illustration used on the label had the greatest impact on both purchase intent and perceptions of brand personality.
Independent Component Analysis and Bayesian Networks: A New Approach to Non-Invasive Brain Machine Interfaces
Advisor: Prof. Richard Ivry
This paper described a new method of non-invasive brain-machine interfaces (BMIs), also known as brain-computer interfaces (BCIs), based on combining independent component (IC) analysis and Bayesian networks. Cohen tested his new method on a simple finger-tapping task from the BCI Competition III; his results were comparable to the best published. He then tested his approach on a continuous movement task in which he predicted the time series of the ICs of a set of electromyography (EMG) signals over a specific trial period using only the ICs derived from electroencephalography (EEG) signals. Finally, he used his methodology to investigate whether the spinal cord is transmitting motor signals from the brain to the peripheral muscles in a simple linear manner, as a wire transmits electricity, or if the spinal cord is performing complex non-linear transformations on the motor signals coming from the brain to the peripheral muscles.
Cross-Modal Differences in Response Inhibition
Advisor: Mark D'Esposito
The ability to inhibit unwanted or irrelevant thoughts and actions is central to cognition. Often, inhibition is conceptualized as a unitary, central faculty, independent of what cues signal its initiation or what process or action needs to be inhibited. Deriving from this perspective, recent neuroimaging and neuropsychological evidence implicating the right inferior prefrontal cortex in the inhibition of an already initiated manual response (the stop-signal task) has led to the proposal that the right inferior prefrontal cortex is the site of a general capacity inhibitory system. Critically, however, the assumption that the stop-signal task or other inhibition-based tasks are reflective of a central inhibition faculty remains to be tested. This project sought to test this assumption directly. In the stop-signal task, regular responses (go trials) are interrupted by periodic cues to stop an already initiated go response (stop trials). This procedure ca yield an estimate of a subject's stop-signal reaction time (SSRT). The SSRT is thought to be directly reflective of the central inhibitory mechanism. This experiment tested whether SSRTs would be consistent regardless of the output modality (manual or verbal) or the type of go decision (spatial or semantic). Output and go decision were crossed factorially, producing four stop-signal tasks. Subjects made either button press (manual) or voice key (verbal) go responses, indicating either the direction of arrow stimuli (spatial) or the typical size of pictured objects (semantic). An SSRT was computed for each of the four tasks. SSRTs correlated within output domain, in that the SSRTs for manual conditions were correlated whether the go decision was spatial or semantic; likewise within the verbal conditions. Strikingly, however, there was no such correlation across output domain, even holding decision content constant. These results cast some doubt on the assumption that all inhibition is due to a central faculty, mediated by the right prefrontal cortex. Rather, these data may indicate that multiple inhibitory mechanisms exist depending on the nature of the response to be inhibited.
Robert J. Glushko, Adjunct Professor at the UC Berkeley School of Information, and Director of its Center for Document Engineering, specializes in information management, electronic publishing, Internet commerce, and human factors in computing systems. After receiving his BA in experimental psychology from Stanford in 1974, and his PhD in Cognitive Science from UC San Diego in 1979, he went on to found three companies, and pioneered the use of the XML language for business-to-business transactions. Prof Glushko is also President of the Robert J. Glushko and Pamela Samuelson Foundation, which sponsors the Rumelhart Prize in Cognitive Science awarded annually by the Cognitive Science Society. He is the author (with Tim McGrath) of Document Engineering (2005), among other works.