We can learn as much about language models from what they say as we learn from their performance on targeted benchmarks. Sampling is a promising bottom-up method for probing, but generating samples from successful models like BERT remains challenging. Taking inspiration from theories of iterated learning in cognitive science, we explore the use of serial reproduction chains to probe BERT's priors. Although the masked language modeling objective does not guarantee a consistent joint distribution, we observe that a unique and consistent estimator of the ground-truth joint distribution may be obtained by a GSN sampler, which randomly selects which word to mask and reconstruct on each step. We compare the lexical and syntactic statistics of sentences from the resulting prior distribution against those of the ground-truth corpus distribution and elicit a large empirical sample of naturalness judgments to investigate how, exactly, the model deviates from human speakers. Our findings suggest the need to move beyond top-down evaluation methods toward bottom-up probing to capture the full richness of what has been learned about language.
Similarity judgments provide a well-established method for accessing mental representations, with applications in psychology, neuroscience and machine learning. However, collecting similarity judgments can be prohibitively expensive for naturalistic datasets as the number of comparisons grows quadratically in the number of stimuli. One way to tackle this problem is to construct approximation procedures that rely on more accessible proxies for predicting similarity. Here we leverage recent advances in language models and online recruitment, proposing an efficient domain-general procedure for predicting human similarity judgments based on text descriptions. Intuitively, similar stimuli are likely to evoke similar descriptions, allowing us to use description similarity to predict pairwise similarity judgments. Crucially, the number of descriptions required grows only linearly with the number of stimuli, drastically reducing the amount of data required. We test this procedure on six datasets of naturalistic images and show that our models outperform previous approaches based on visual information.
Being able to learn from small amounts of data is a key characteristic of human intelligence, but exactly {\em how} small? In this paper, we introduce a novel experimental paradigm that allows us to examine classification in an extremely data-scarce setting, asking whether humans can learn more categories than they have exemplars (i.e., can humans do "less-than-one shot" learning?). An experiment conducted using this paradigm reveals that people are capable of learning in such settings, and provides several insights into underlying mechanisms. First, people can accurately infer and represent high-dimensional feature spaces from very little data. Second, having inferred the relevant spaces, people use a form of prototype-based categorization (as opposed to exemplar-based) to make categorical inferences. Finally, systematic, machine-learnable patterns in responses indicate that people may have efficient inductive biases for dealing with this class of data-scarce problems.
Language is not only used to inform. We often seek to persuade by arguing in favor of a particular view. Persuasion raises a number of challenges for classical accounts of belief updating, as information cannot be taken at face value. How should listeners account for a speaker's "hidden agenda" when incorporating new information? Here, we extend recent probabilistic models of recursive social reasoning to allow for persuasive goals and show that our model provides a new pragmatic explanation for why weakly favorable arguments may backfire, a phenomenon known as the weak evidence effect. Critically, our model predicts a relationship between belief updating and speaker expectations: weak evidence should only backfire when speakers are expected to act under persuasive goals, implying the absence of stronger evidence. We introduce a simple experimental paradigm called the Stick Contest to measure the extent to which the weak evidence effect depends on speaker expectations, and show that a pragmatic listener model accounts for the empirical data better than alternative models. Our findings suggest potential avenues for rational models of social reasoning to further illuminate decision-making phenomena.
Despite the increasing scale of datasets in machine learning, generalization to unseen regions of the data distribution remains crucial. Such extrapolation is by definition underdetermined and is dictated by a learner's inductive biases. Machine learning systems often do not share the same inductive biases as humans and, as a result, extrapolate in ways that are inconsistent with our expectations. We investigate two distinct such inductive biases: feature-level bias (differences in which features are more readily learned) and exemplar-vs-rule bias (differences in how these learned features are used for generalization). Exemplar- vs. rule-based generalization has been studied extensively in cognitive psychology, and, in this work, we present a protocol inspired by these experimental approaches for directly probing this trade-off in learning systems. The measures we propose characterize changes in extrapolation behavior when feature coverage is manipulated in a combinatorial setting. We present empirical results across a range of models and across both expository and real-world image and language domains. We demonstrate that measuring the exemplar-rule trade-off while controlling for feature-level bias provides a more complete picture of extrapolation behavior than existing formalisms. We find that most standard neural network models have a propensity towards exemplar-based extrapolation and discuss the implications of these findings for research on data augmentation, fairness, and systematic generalization.
Those designing autonomous systems that interact with humans will invariably face questions about how humans think and make decisions. Fortunately, computational cognitive science offers insight into human decision-making using tools that will be familiar to those with backgrounds in optimization and control (e.g., probability theory, statistical machine learning, and reinforcement learning). Here, we review some of this work, focusing on how cognitive science can provide forward models of human decision-making and inverse models of how humans think about others' decision-making. We highlight relevant recent developments, including approaches that synthesize blackbox and theory-driven modeling, accounts that recast heuristics and biases as forms of bounded optimality, and models that characterize human theory of mind and communication in decision-theoretic terms. In doing so, we aim to provide readers with a glimpse of the range of frameworks, methodologies, and actionable insights that lie at the intersection of cognitive science and control research.
Many transfer problems require re-using previously optimal decisions for solving new tasks, which suggests the need for learning algorithms that can modify the mechanisms for choosing certain actions independently of those for choosing others. However, there is currently no formalism nor theory for how to achieve this kind of modular credit assignment. To answer this question, we define modular credit assignment as a constraint on minimizing the algorithmic mutual information among feedback signals for different decisions. We introduce what we call the modularity criterion for testing whether a learning algorithm satisfies this constraint by performing causal analysis on the algorithm itself. We generalize the recently proposed societal decision-making framework as a more granular formalism than the Markov decision process to prove that for decision sequences that do not contain cycles, certain single-step temporal difference action-value methods meet this criterion while all policy-gradient methods do not. Empirical evidence suggests that such action-value methods are more sample efficient than policy-gradient methods on transfer problems that require only sparse changes to a sequence of previously optimal decisions.
Developments in machine learning interpretability techniques over the past decade have provided new tools to observe the image regions that are most informative for classification and localization in artificial neural networks (ANNs). Are the same regions similarly informative to human observers? Using data from 78 new experiments and 6,610 participants, we show that passive attention techniques reveal a significant overlap with human visual selectivity estimates derived from 6 distinct behavioral tasks including visual discrimination, spatial localization, recognizability, free-viewing, cued-object search, and saliency search fixations. We find that input visualizations derived from relatively simple ANN architectures probed using guided backpropagation methods are the best predictors of a shared component in the joint variability of the human measures. We validate these correlational results with causal manipulations using recognition experiments. We show that images masked with ANN attention maps were easier for humans to classify than control masks in a speeded recognition experiment. Similarly, we find that recognition performance in the same ANN models was likewise influenced by masking input images using human visual selectivity maps. This work contributes a new approach to evaluating the biological and psychological validity of leading ANNs as models of human vision: by examining their similarities and differences in terms of their visual selectivity to the information contained in images.