Search

Abstract reasoning is one of the defining characteristics of human intelligence and can be estimated by visual IQ tests such as Raven's Progressive Matrices. In this paper, we propose using a hierarchical Transformer encoder with structured representation that employs a novel neural network architecture to improve both perception and reasoning in a visual IQ test. For perception, we used object detection models to extract the structured features. For reasoning, we used the Transformer encoder in a hierarchical manner that fits the structure of Raven's Progressive Matrices. Experimental results on the RAVEN dataset, which is one of the major large-scale datasets on Raven's Progressive Matrices, showed that our proposed architecture achieved an overall accuracy of 99.62%, which is an improvement of more than 8% points over CoPINet, the present-day, state-of-the-art neural network model.

The world contains not only objects and features (red apples, glass bowls, wooden tables), but also relations holding between them (apples contained in bowls, bowls supported by tables). Representations of these relations are often developmentally precocious and linguistically privileged; but how does the mind extract them in the first place? Although relations themselves cast no light onto our eyes, a growing body of work suggests that even very sophisticated relations display key signatures of automatic visual processing. Across physical, eventive, and social domains, relations such as support , fit , cause , chase , and even socially interact are extracted rapidly, are impossible to ignore, and influence other perceptual processes. Sophisticated and structured relations are not only judged and understood, but also seen — revealing surprisingly rich content in visual perception itself. The world is more than a 'bag of objects': it contains relations and interactions between entities. How are such relations extracted by the mind? Recent work using the tools of vision science suggests that, just as the visual system computes properties such as an object's color or shape, it also computes an object's relation to its physical and social environment, automatically categorizing configurations of objects into distinct relational types. Across physical, eventive, and social domains, sophisticated and structured relations such as support , fit , cause , and chase are not only judged and considered, but also seen. The perception of such sophisticated relations reveals surprisingly rich content in visual perception itself, with consequences for longstanding debates in the philosophy of perception and computational modeling of human vision. [ABSTRACT FROM AUTHOR]

This paper proposes a novel technique for representing templates and instances of concept classes. A template representation refers to the generic representation that captures the characteristics of an entire class. The proposed technique uses end-to-end deep learning to learn structured and composable representations from input images and discrete labels. The obtained representations are based on distance estimates between the distributions given by the class label and those given by contextual information, which are modeled as environments. We prove that the representations have a clear structure allowing decomposing the representation into factors that represent classes and environments. We evaluate our novel technique on classification and retrieval tasks involving different modalities (visual and language data). In various experiments, we show how the representations can be compressed and how different hyperparameters impact performance.

Many theoretical and empirical contributions to the Predictive Processing account emphasize the important role of precision modulation of prediction errors. Recently it has been proposed that the causal models used in human predictive processing are best formally modeled by categorical probability distributions. Crucially, such distributions assume a well-defined, discrete state space. In this paper we explore the consequences of this formalization. In particular we argue that the level of detail of generative models and predictions modulates prediction error. We show that both increasing the level of detail of the generative models and decreasing the level of detail of the predictions can be suitable mechanisms for lowering prediction errors. Both increase precision, yet come at the price of lowering the amount of information that can be gained by correct predictions. Our theoretical result establishes a key open empirical question to address: How does the brain optimize the trade-off between high precision and information gain when making its predictions? [ABSTRACT FROM AUTHOR]