MindSet: Vision. A toolbox for testing DNNs on key psychological experiments

Multiple benchmarks have been developed to assess the alignment between deep neural networks (DNNs) and human vision. In almost all cases these benchmarks are observational in the sense they are composed of behavioural and brain responses to naturalistic images that have not been manipulated to test hypotheses regarding how DNNs or humans perceive and identify objects. Here we introduce the toolbox MindSet: Vision, consisting of a collection of image datasets and related scripts designed to test DNNs on 30 psychological findings. In all experimental conditions, the stimuli are systematically manipulated to test specific hypotheses regarding human visual perception and object recognition. In addition to providing pre-generated datasets of images, we provide code to regenerate these datasets, offering many configurable parameters which greatly extend the dataset versatility for different research contexts, and code to facilitate the testing of DNNs on these image datasets using three different methods (similarity judgments, out-of-distribution classification, and decoder method), accessible at https://github.com/MindSetVision/mindset-vision. We test ResNet-152 on each of these methods as an example of how the toolbox can be used.

Visual Reasoning in Object-Centric Deep Neural Networks: A Comparative Cognition Approach

Achieving visual reasoning is a long-term goal of artificial intelligence. In the last decade, several studies have applied deep neural networks (DNNs) to the task of learning visual relations from images, with modest results in terms of generalization of the relations learned. However, in recent years, object-centric representation learning has been put forward as a way to achieve visual reasoning within the deep learning framework. Object-centric models attempt to model input scenes as compositions of objects and relations between them. To this end, these models use several kinds of attention mechanisms to segregate the individual objects in a scene from the background and from other objects. In this work we tested relation learning and generalization in several object-centric models, as well as a ResNet-50 baseline. In contrast to previous research, which has focused heavily in the same-different task in order to asses relational reasoning in DNNs, we use a set of tasks -- with varying degrees of difficulty -- derived from the comparative cognition literature. Our results show that object-centric models are able to segregate the different objects in a scene, even in many out-of-distribution cases. In our simpler tasks, this improves their capacity to learn and generalize visual relations in comparison to the ResNet-50 baseline. However, object-centric models still struggle in our more difficult tasks and conditions. We conclude that abstract visual reasoning remains an open challenge for DNNs, including object-centric models.

The role of object-centric representations, guided attention, and external memory on generalizing visual relations

Visual reasoning is a long-term goal of vision research. In the last decade, several works have attempted to apply deep neural networks (DNNs) to the task of learning visual relations from images, with modest results in terms of the generalization of the relations learned. In recent years, several innovations in DNNs have been developed in order to enable learning abstract relation from images. In this work, we systematically evaluate a series of DNNs that integrate mechanism such as slot attention, recurrently guided attention, and external memory, in the simplest possible visual reasoning task: deciding whether two objects are the same or different. We found that, although some models performed better than others in generalizing the same-different relation to specific types of images, no model was able to generalize this relation across the board. We conclude that abstract visual reasoning remains largely an unresolved challenge for DNNs.

Do DNNs trained on Natural Images organize visual features into Gestalts?

Gestalt psychologists have identified a range of conditions in which humans organize elements of a scene into a group or whole, and these perceptual grouping principles play an important role in scene perception and object identification. More recently, Deep Neural Networks (DNNs) trained on natural images have been proposed as compelling models of human vision based on reports that they perform well on various brain and behavioral benchmarks. Here we compared human and DNNs responses in discrimination judgments that assess a range of Gestalt organization principles. We found that most networks exhibited a moderate degree of Gestalt grouping for some complex stimuli at the last fully connected layer. However, in contrast with human neural data, this sensitivity vanishes at earlier visual processing layers. In a second experiment, by using simple dots configuration patterns, we found that all networks were only weakly sensitive to the grouping properties of proximity, and completely insensitive to orientation and linearity, three principles that have been shown to have a strong and robust effect on humans. Even top-performing models on the behavioral and brain benchmark Brain-Score miss these fundamental properties of human vision. Our overall conclusion is that, even when exhibiting Gestalt grouping, networks trained on 2D images use perceptual principles fundamentally different than humans.

Successes and critical failures of neural networks in capturing human-like speech recognition

Natural and artificial audition can in principle evolve different solutions to a given problem. The constraints of the task, however, can nudge the cognitive science and engineering of audition to qualitatively converge, suggesting that a closer mutual examination would improve artificial hearing systems and process models of the mind and brain. Speech recognition - an area ripe for such exploration - is inherently robust in humans to a number transformations at various spectrotemporal granularities. To what extent are these robustness profiles accounted for by high-performing neural network systems? We bring together experiments in speech recognition under a single synthesis framework to evaluate state-of-the-art neural networks as stimulus-computable, optimized observers. In a series of experiments, we (1) clarify how influential speech manipulations in the literature relate to each other and to natural speech, (2) show the granularities at which machines exhibit out-of-distribution robustness, reproducing classical perceptual phenomena in humans, (3) identify the specific conditions where model predictions of human performance differ, and (4) demonstrate a crucial failure of all artificial systems to perceptually recover where humans do, suggesting a key specification for theory and model building. These findings encourage a tighter synergy between the cognitive science and engineering of audition.

Lost in Latent Space: Disentangled Models and the Challenge of Combinatorial Generalisation

Recent research has shown that generative models with highly disentangled representations fail to generalise to unseen combination of generative factor values. These findings contradict earlier research which showed improved performance in out-of-training distribution settings when compared to entangled representations. Additionally, it is not clear if the reported failures are due to (a) encoders failing to map novel combinations to the proper regions of the latent space or (b) novel combinations being mapped correctly but the decoder/downstream process is unable to render the correct output for the unseen combinations. We investigate these alternatives by testing several models on a range of datasets and training settings. We find that (i) when models fail, their encoders also fail to map unseen combinations to correct regions of the latent space and (ii) when models succeed, it is either because the test conditions do not exclude enough examples, or because excluded generative factors determine independent parts of the output image. Based on these results, we argue that to generalise properly, models not only need to capture factors of variation, but also understand how to invert the generative process that was used to generate the data.

Convolutional Neural Networks Are Not Invariant to Translation, but They Can Learn to Be

When seeing a new object, humans can immediately recognize it across different retinal locations: the internal object representation is invariant to translation. It is commonly believed that Convolutional Neural Networks (CNNs) are architecturally invariant to translation thanks to the convolution and/or pooling operations they are endowed with. In fact, several studies have found that these networks systematically fail to recognise new objects on untrained locations. In this work, we test a wide variety of CNNs architectures showing how, apart from DenseNet-121, none of the models tested was architecturally invariant to translation. Nevertheless, all of them could learn to be invariant to translation. We show how this can be achieved by pretraining on ImageNet, and it is sometimes possible with much simpler data sets when all the items are fully translated across the input canvas. At the same time, this invariance can be disrupted by further training due to catastrophic forgetting/interference. These experiments show how pretraining a network on an environment with the right `latent' characteristics (a more naturalistic environment) can result in the network learning deep perceptual rules which would dramatically improve subsequent generalization.

Learning Online Visual Invariances for Novel Objects via Supervised and Self-Supervised Training

Humans can identify objects following various spatial transformations such as scale and viewpoint. This extends to novel objects, after a single presentation at a single pose, sometimes referred to as online invariance. CNNs have been proposed as a compelling model of human vision, but their ability to identify objects across transformations is typically tested on held-out samples of trained categories after extensive data augmentation. This paper assesses whether standard CNNs can support human-like online invariance by training models to recognize images of synthetic 3D objects that undergo several transformations: rotation, scaling, translation, brightness, contrast, and viewpoint. Through the analysis of models' internal representations, we show that standard supervised CNNs trained on transformed objects can acquire strong invariances on novel classes even when trained with as few as 50 objects taken from 10 classes. This extended to a different dataset of photographs of real objects. We also show that these invariances can be acquired in a self-supervised way, through solving the same/different task. We suggest that this latter approach may be similar to how humans acquire invariances.

Generalisation in Neural Networks Does not Require Feature Overlap

That shared features between train and test data are required for generalisation in artificial neural networks has been a common assumption of both proponents and critics of these models. Here, we show that convolutional architectures avoid this limitation by applying them to two well known challenges, based on learning the identity function and learning rules governing sequences of words. In each case, successful performance on the test set requires generalising to features that were not present in the training data, which is typically not feasible for standard connectionist models. However, our experiments demonstrate that neural networks can succeed on such problems when they incorporate the weight sharing employed by convolutional architectures. In the image processing domain, such architectures are intended to reflect the symmetry under spatial translations of the natural world that such images depict. We discuss the role of symmetry in the two tasks and its connection to generalisation.

Are there any 'object detectors' in the hidden layers of CNNs trained to identify objects or scenes?

Various methods of measuring unit selectivity have been developed with the aim of better understanding how neural networks work. But the different measures provide divergent estimates of selectivity, and this has led to different conclusions regarding the conditions in which selective object representations are learned and the functional relevance of these representations. In an attempt to better characterize object selectivity, we undertake a comparison of various selectivity measures on a large set of units in AlexNet, including localist selectivity, precision, class-conditional mean activity selectivity (CCMAS), network dissection,the human interpretation of activation maximization (AM) images, and standard signal-detection measures. We find that the different measures provide different estimates of object selectivity, with precision and CCMAS measures providing misleadingly high estimates. Indeed, the most selective units had a poor hit-rate or a high false-alarm rate (or both) in object classification, making them poor object detectors. We fail to find any units that are even remotely as selective as the 'grandmother cell' units reported in recurrent neural networks. In order to generalize these results, we compared selectivity measures on units in VGG-16 and GoogLeNet trained on the ImageNet or Places-365 datasets that have been described as 'object detectors'. Again, we find poor hit-rates and high false-alarm rates for object classification. We conclude that signal-detection measures provide a better assessment of single-unit selectivity compared to common alternative approaches, and that deep convolutional networks of image classification do not learn object detectors in their hidden layers.