With the rapid improvement of machine learning (ML) models, cognitive scientists are increasingly asking about their alignment with how humans think. Here, we ask this question for computer vision models and human sensitivity to geometric and topological (GT) concepts. Under the core knowledge account, these concepts are innate and supported by dedicated neural circuitry. In this work, we investigate an alternative explanation, that GT concepts are learned ``for free'' through everyday interaction with the environment. We do so using computer visions models, which are trained on large image datasets. We build on prior studies to investigate the overall performance and human alignment of three classes of models -- convolutional neural networks (CNNs), transformer-based models, and vision-language models -- on an odd-one-out task testing 43 GT concepts spanning seven classes. Transformer-based models achieve the highest overall accuracy, surpassing that of young children. They also show strong alignment with children's performance, finding the same classes of concepts easy vs. difficult. By contrast, vision-language models underperform their vision-only counterparts and deviate further from human profiles, indicating that na\"ive multimodality might compromise abstract geometric sensitivity. These findings support the use of computer vision models to evaluate the sufficiency of the learning account for explaining human sensitivity to GT concepts, while also suggesting that integrating linguistic and visual representations might have unpredicted deleterious consequences.
Object detection methods trained on a fixed set of known classes struggle to detect objects of unknown classes in the open-world setting. Current fixes involve adding approximate supervision with pseudo-labels corresponding to candidate locations of objects, typically obtained in a class-agnostic manner. While previous approaches mainly rely on the appearance of objects, we find that geometric cues improve unknown recall. Although additional supervision from pseudo-labels helps to detect unknown objects, it also introduces confusion for known classes. We observed a notable decline in the model's performance for detecting known objects in the presence of noisy pseudo-labels. Drawing inspiration from studies on human cognition, we propose to group known classes into superclasses. By identifying similarities between classes within a superclass, we can identify unknown classes through an odd-one-out scoring mechanism. Our experiments on open-world detection benchmarks demonstrate significant improvements in unknown recall, consistently across all tasks. Crucially, we achieve this without compromising known performance, thanks to better partitioning of the feature space with superclasses.
This paper introduces a novel anomaly detection (AD) problem that focuses on identifying `odd-looking' objects relative to the other instances in a given scene. In contrast to the traditional AD benchmarks, anomalies in our task are scene-specific, defined by the regular instances that make up the majority. Since object instances may be only partly visible from a single viewpoint, our setting employs multiple views of each scene as input. To provide a testbed for future research in this task, we introduce two benchmarks, ToysAD-8K and PartsAD-15K. We propose a novel method that constructs 3D object-centric representations from multiple 2D views for each instance and detects the anomalous ones through a cross-instance comparison. We rigorously analyze our method quantitatively and qualitatively on the presented benchmarks.
Diffusion models, trained on large amounts of data, showed remarkable performance for image synthesis. They have high error consistency with humans and low texture bias when used for classification. Furthermore, prior work demonstrated the decomposability of their bottleneck layer representations into semantic directions. In this work, we analyze how well such representations are aligned to human responses on a triplet odd-one-out task. We find that despite the aforementioned observations: I) The representational alignment with humans is comparable to that of models trained only on ImageNet-1k. II) The most aligned layers of the denoiser U-Net are intermediate layers and not the bottleneck. III) Text conditioning greatly improves alignment at high noise levels, hinting at the importance of abstract textual information, especially in the early stage of generation.
Endowing machines with abstract reasoning ability has been a long-term research topic in artificial intelligence. Raven's Progressive Matrix (RPM) is widely used to probe abstract visual reasoning in machine intelligence, where models will analyze the underlying rules and select one image from candidates to complete the image matrix. Participators of RPM tests can show powerful reasoning ability by inferring and combining attribute-changing rules and imagining the missing images at arbitrary positions of a matrix. However, existing solvers can hardly manifest such an ability in realistic RPM tests. In this paper, we propose a deep latent variable model for answer generation problems through Rule AbstractIon and SElection (RAISE). RAISE can encode image attributes into latent concepts and abstract atomic rules that act on the latent concepts. When generating answers, RAISE selects one atomic rule out of the global knowledge set for each latent concept to constitute the underlying rule of an RPM. In the experiments of bottom-right and arbitrary-position answer generation, RAISE outperforms the compared solvers in most configurations of realistic RPM datasets. In the odd-one-out task and two held-out configurations, RAISE can leverage acquired latent concepts and atomic rules to find the rule-breaking image in a matrix and handle problems with unseen combinations of rules and attributes.
Abstract Visual Reasoning (AVR) comprises a wide selection of various problems similar to those used in human IQ tests. Recent years have brought dynamic progress in solving particular AVR tasks, however, in the contemporary literature AVR problems are largely dealt with in isolation, leading to highly specialized task-specific methods. With the aim of developing universal learning systems in the AVR domain, we propose the unified model for solving Single-Choice Abstract visual Reasoning tasks (SCAR), capable of solving various single-choice AVR tasks, without making any a priori assumptions about the task structure, in particular the number and location of panels. The proposed model relies on a novel Structure-Aware dynamic Layer (SAL), which adapts its weights to the structure of the considered AVR problem. Experiments conducted on Raven's Progressive Matrices, Visual Analogy Problems, and Odd One Out problems show that SCAR (SAL-based models, in general) effectively solves diverse AVR tasks, and its performance is on par with the state-of-the-art task-specific baselines. What is more, SCAR demonstrates effective knowledge reuse in multi-task and transfer learning settings. To our knowledge, this work is the first successful attempt to construct a general single-choice AVR solver relying on self-configurable architecture and unified solving method. With this work we aim to stimulate and foster progress on task-independent research paths in the AVR domain, with the long-term goal of development of a general AVR solver.
Despite the huge recent breakthroughs in neural networks (NNs) for artificial intelligence (specifically deep convolutional networks) such NNs do not achieve human-level performance: they can be hacked by images that would fool no human and lack `common sense'. It has been argued that a basis of human-level intelligence is mankind's ability to perform relational reasoning: the comparison of different objects, measuring similarity, grasping of relations between objects and the converse, figuring out the odd one out in a set of objects. Mankind can even do this with objects they have never seen before. Here we show how ClusterFlow, a semi-supervised hierarchical clustering framework can operate on trained NNs utilising the rich multi-dimensional class and feature data found at the pre-SoftMax layer to build a hyperspacial map of classes/features and this adds more human-like functionality to modern deep convolutional neural networks. We demonstrate this with 3 tasks. 1. the statistical learning based `mistakes' made by infants when attending to images of cats and dogs. 2. improving both the resilience to hacking images and the accurate measure of certainty in deep-NNs. 3. Relational reasoning over sets of images, including those not known to the NN nor seen before. We also demonstrate that ClusterFlow can work on non-NN data and deal with missing data by testing it on a Chemistry dataset. This work suggests that modern deep NNs can be made more human-like without re-training of the NNs. As it is known that some methods used in deep and convolutional NNs are not biologically plausible or perhaps even the best approach: the ClusterFlow framework can sit on top of any NN and will be a useful tool to add as NNs are improved in this regard.
This study presents an adversarial method for anomaly detection in real-world applications, leveraging the power of generative adversarial neural networks (GANs) through cycle consistency in reconstruction error. Previous methods suffer from the high variance between class-wise accuracy which leads to not being applicable for all types of anomalies. The proposed method named RCALAD tries to solve this problem by introducing a novel discriminator to the structure, which results in a more efficient training process. Additionally, RCALAD employs a supplementary distribution in the input space to steer reconstructions toward the normal data distribution, effectively separating anomalous samples from their reconstructions and facilitating more accurate anomaly detection. To further enhance the performance of the model, two novel anomaly scores are introduced. The proposed model has been thoroughly evaluated through extensive experiments on six various datasets, yielding results that demonstrate its superiority over existing state-of-the-art models. The code is readily available to the research community at https://github.com/zahraDehghanian97/RCALAD.
Today's computer vision models achieve human or near-human level performance across a wide variety of vision tasks. However, their architectures, data, and learning algorithms differ in numerous ways from those that give rise to human vision. In this paper, we investigate the factors that affect the alignment between the representations learned by neural networks and human mental representations inferred from behavioral responses. We find that model scale and architecture have essentially no effect on the alignment with human behavioral responses, whereas the training dataset and objective function both have a much larger impact. These findings are consistent across three datasets of human similarity judgments collected using two different tasks. Linear transformations of neural network representations learned from behavioral responses from one dataset substantially improve alignment with human similarity judgments on the other two datasets. In addition, we find that some human concepts such as food and animals are well-represented by neural networks whereas others such as royal or sports-related objects are not. Overall, although models trained on larger, more diverse datasets achieve better alignment with humans than models trained on ImageNet alone, our results indicate that scaling alone is unlikely to be sufficient to train neural networks with conceptual representations that match those used by humans.
Recognition of geometrical patterns seems to be an important aspect of human intelligence. Geometric pattern recognition is used in many intelligence tests, including Dehaene's odd-one-out test of Core Geometry (CG)) based on intuitive geometrical concepts (Dehaene et al., 2006). Earlier work has developed a symmetry-based cognitive model of Dehaene's test and demonstrated performance comparable to that of humans. In this work, we further investigate the role of symmetry in geometrical intuition and build a cognitive model for the 2-Alternative Forced Choice (2-AFC) variation of the CG test (Marupudi & Varma 2021). In contrast to Dehaene's test, 2-AFC leaves almost no space for cognitive models based on generalization over multiple examples. Our symmetry-based model achieves an accuracy comparable to the human average on the 2-AFC test and appears to capture an essential part of intuitive geometry.
A central goal in the cognitive sciences is the development of numerical models for mental representations of object concepts. This paper introduces Variational Interpretable Concept Embeddings (VICE), an approximate Bayesian method for embedding object concepts in a vector space using data collected from humans in a triplet odd-one-out task. VICE uses variational inference to obtain sparse, non-negative representations of object concepts with uncertainty estimates for the embedding values. These estimates are used to automatically select the dimensions that best explain the data. We derive a PAC learning bound for VICE that can be used to estimate generalization performance or determine a sufficient sample size for experimental design. VICE rivals or outperforms its predecessor, SPoSE, at predicting human behavior in the triplet odd-one-out task. Furthermore, VICE's object representations are more reproducible and consistent across random initializations, highlighting the unique advantage of using VICE for deriving interpretable embeddings from human behavior.
We investigate the optimal model size and number of tokens for training a transformer language model under a given compute budget. We find that current large language models are significantly undertrained, a consequence of the recent focus on scaling language models whilst keeping the amount of training data constant. By training over \nummodels language models ranging from 70 million to over 16 billion parameters on 5 to 500 billion tokens, we find that for compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled. We test this hypothesis by training a predicted compute-optimal model, \chinchilla, that uses the same compute budget as \gopher but with 70B parameters and 4$\times$ more more data. \chinchilla uniformly and significantly outperforms \Gopher (280B), GPT-3 (175B), Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks. This also means that \chinchilla uses substantially less compute for fine-tuning and inference, greatly facilitating downstream usage. As a highlight, \chinchilla reaches a state-of-the-art average accuracy of 67.5\% on the MMLU benchmark, greater than a 7\% improvement over \gopher.
Language modelling provides a step towards intelligent communication systems by harnessing large repositories of written human knowledge to better predict and understand the world. In this paper, we present an analysis of Transformer-based language model performance across a wide range of model scales -- from models with tens of millions of parameters up to a 280 billion parameter model called Gopher. These models are evaluated on 152 diverse tasks, achieving state-of-the-art performance across the majority. Gains from scale are largest in areas such as reading comprehension, fact-checking, and the identification of toxic language, but logical and mathematical reasoning see less benefit. We provide a holistic analysis of the training dataset and model's behaviour, covering the intersection of model scale with bias and toxicity. Finally we discuss the application of language models to AI safety and the mitigation of downstream harms.
Inferring the abstract relational and causal structure of the world is a major challenge for reinforcement-learning (RL) agents. For humans, language--particularly in the form of explanations--plays a considerable role in overcoming this challenge. Here, we show that language can play a similar role for deep RL agents in complex environments. While agents typically struggle to acquire relational and causal knowledge, augmenting their experience by training them to predict language descriptions and explanations can overcome these limitations. We show that language can help agents learn challenging relational tasks, and examine which aspects of language contribute to its benefits. We then show that explanations can help agents to infer not only relational but also causal structure. Language can shape the way that agents to generalize out-of-distribution from ambiguous, causally-confounded training, and explanations even allow agents to learn to perform experimental interventions to identify causal relationships. Our results suggest that language description and explanation may be powerful tools for improving agent learning and generalization.
The effective application of representation learning to real-world problems requires both techniques for learning useful representations, and also robust ways to evaluate properties of representations. Recent work in disentangled representation learning has shown that unsupervised representation learning approaches rely on fully supervised disentanglement metrics, which assume access to labels for ground-truth factors of variation. In many real-world cases ground-truth factors are expensive to collect, or difficult to model, such as for perception. Here we empirically show that a weakly-supervised downstream task based on odd-one-out observations is suitable for model selection by observing high correlation on a difficult downstream abstract visual reasoning task. We also show that a bespoke metric-learning VAE model which performs highly on this task also out-performs other standard unsupervised and a weakly-supervised disentanglement model across several metrics.
Identifying common patterns among events is a key ability in human and machine perception, as it underlies intelligent decision making. We propose an approach for learning semantic relational set abstractions on videos, inspired by human learning. We combine visual features with natural language supervision to generate high-level representations of similarities across a set of videos. This allows our model to perform cognitive tasks such as set abstraction (which general concept is in common among a set of videos?), set completion (which new video goes well with the set?), and odd one out detection (which video does not belong to the set?). Experiments on two video benchmarks, Kinetics and Multi-Moments in Time, show that robust and versatile representations emerge when learning to recognize commonalities among sets. We compare our model to several baseline algorithms and show that significant improvements result from explicitly learning relational abstractions with semantic supervision.
Recent advances in the field of saliency have concentrated on fixation prediction, with benchmarks reaching saturation. However, there is an extensive body of works in psychology and neuroscience that describe aspects of human visual attention that might not be adequately captured by current approaches. Here, we investigate singleton detection, which can be thought of as a canonical example of salience. We introduce two novel datasets, one with psychophysical patterns and one with natural odd-one-out stimuli. Using these datasets we demonstrate through extensive experimentation that nearly all saliency algorithms do not adequately respond to singleton targets in synthetic and natural images. Furthermore, we investigate the effect of training state-of-the-art CNN-based saliency models on these types of stimuli and conclude that the additional training data does not lead to a significant improvement of their ability to find odd-one-out targets. Datasets are available at http://data.nvision2.eecs.yorku.ca/P3O3/.
We propose a new self-supervised CNN pre-training technique based on a novel auxiliary task called "odd-one-out learning". In this task, the machine is asked to identify the unrelated or odd element from a set of otherwise related elements. We apply this technique to self-supervised video representation learning where we sample subsequences from videos and ask the network to learn to predict the odd video subsequence. The odd video subsequence is sampled such that it has wrong temporal order of frames while the even ones have the correct temporal order. Therefore, to generate a odd-one-out question no manual annotation is required. Our learning machine is implemented as multi-stream convolutional neural network, which is learned end-to-end. Using odd-one-out networks, we learn temporal representations for videos that generalizes to other related tasks such as action recognition. On action classification, our method obtains 60.3\% on the UCF101 dataset using only UCF101 data for training which is approximately 10% better than current state-of-the-art self-supervised learning methods. Similarly, on HMDB51 dataset we outperform self-supervised state-of-the art methods by 12.7% on action classification task.