Same-different visual reasoning is a basic skill central to abstract combinatorial thought. This fact has lead neural networks researchers to test same-different classification on deep convolutional neural networks (DCNNs), which has resulted in a controversy regarding whether this skill is within the capacity of these models. However, most tests of same-different classification rely on testing on images that come from the same pixel-level distribution as the testing images, yielding the results inconclusive. In this study we tested relational same-different reasoning DCNNs. In a series of simulations we show that DCNNs are capable of visual same-different classification, but only when the test images are similar to the training images at the pixel-level. In contrast, even when there are only subtle differences between the testing and training images, the performance of DCNNs could drop to chance levels. This is true even when DCNNs' training regime included a wide distribution of images or when they were trained in a multi-task setup in which training included an additional relational task with test images from the same pixel-level distribution.

Over the last 20 years, causal-model theory has produced muchknowledge about causal-based categorization. However, per-sistent violations to the normative causal-model theory areprevalent. In particular, violations to the Markov conditionhave been repeatedly found. These violations have receiveddifferent explanations. Here, we develop a model that startsfrom generally accepted cognitive phenomena (e.g., process-ing limitations, the relevance of inference in cognitive process-ing) and assumes that people are not fully causal nor fully asso-ciative when performing causal-based categorization, offeringa new explanation for Markov violations.

This work provides an alternative account for deviations in human causal reasoning from normative predictions based on Causal Bayesian Networks (CBNs). We highlight violations of the Markov condition (Screening Off) and insufficient Explaining Away. Different from other accounts, our model does not assume that people fail to honor normative predictions due to reliance on heuristics, hidden nodes and links or cognitive limitations. Instead, we propose that people are rationally uncertain about the received causal model they are asked to reason with. We fitted the model to published data from two experiments where people were asked to make probability estimates on inferences of interest within a causal model. We find that the model is able to i) reproduce deviations from normative predictions, and ii) predict changes in the magnitude of these deviations across contexts. We conclude that assuming that people, in order to be rational, will always fully believe in the information they receive about a causal model may be too strong an assumption.

Humans readily generalize, applying prior knowledge to novelsituations and stimuli. Advances in machine learning have be-gun to approximate and even surpass human performance, butthese systems struggle to generalize what they have learnedto untrained situations. We present a model based on well-established neurocomputational principles that demonstrateshuman-level generalisation. This model is trained to play onevideo game (Breakout) and performs one-shot generalisationto a new game (Pong) with different characteristics. The modelgeneralizes because it learns structured representations that arefunctionally symbolic (viz., a role-filler binding calculus) fromunstructured training data. It does so without feedback, andwithout requiring that structured representations are specifieda priori. Specifically, the model uses neural co-activation todiscover which characteristics of the input are invariant and tolearn relational predicates, and oscillatory regularities in net-work firing to bind predicates to arguments. To our knowledge,this is the first demonstration of human-like generalisation ina machine system that does not assume structured representa-tions to begin with.

Responding to similarity, difference, and relative magnitude isubiquitous in the animal kingdom. However, humans seemunique in the ability to represent relative magnitude andsimilarity as abstract relations that take arguments (e.g.,greater-than (x,y)). While many models use structuredrelational representations of magnitude and similarity, littleprogress has been made on how these representations arise.Models that use these representations assume access tocomputations of similarity and magnitude a priori. We detail amechanism for producing invariant responses to “same”,“different”, “more”, and “less” which can be exploited tocompute similarity and magnitude as an evaluation operator.Using DORA (Doumas, Hummel, & Sandhofer, 2008), theseinvariant responses can serve to learn structured relationalrepresentations of relative magnitude and similarity from pixelimages of simple shapes.