Deep neural networks are not a single hypothesis but a language for expressing computational hypotheses.

An ideal vision model accounts for behavior and neurophysiology in both naturalistic conditions and designed lab experiments. Unlike psychological theories, artificial neural networks (ANNs) actually perform visual tasks and generate testable predictions for arbitrary inputs. These advantages enable ANNs to engage the entire spectrum of the evidence. Failures of particular models drive progress in a vibrant ANN research program of human vision.

The neuroconnectionist research programme.

Artificial neural networks (ANNs) inspired by biology are beginning to be widely used to model behavioural and neural data, an approach we call 'neuroconnectionism'. ANNs have been not only lauded as the current best models of information processing in the brain but also criticized for failing to account for basic cognitive functions. In this Perspective article, we propose that arguing about the successes and failures of a restricted set of current ANNs is the wrong approach to assess the promise of neuroconnectionism for brain science. Instead, we take inspiration from the philosophy of science, and in particular from Lakatos, who showed that the core of a scientific research programme is often not directly falsifiable but should be assessed by its capacity to generate novel insights. Following this view, we present neuroconnectionism as a general research programme centred around ANNs as a computational language for expressing falsifiable theories about brain computation. We describe the core of the programme, the underlying computational framework and its tools for testing specific neuroscientific hypotheses and deriving novel understanding. Taking a longitudinal view, we review past and present neuroconnectionist projects and their responses to challenges and argue that the research programme is highly progressive, generating new and otherwise unreachable insights into the workings of the brain.

First-person experience cannot rescue causal structure theories from the unfolding argument.

We recently put forward an argument, the Unfolding Argument (UA), that integrated information theory (IIT) and other causal structure theories are either already falsified or unfalsifiable, which provoked significant criticism. It seems that we and the critics agree that the main question in this debate is whether first-person experience, independent of third-person data, is a sufficient foundation for theories of consciousness. Here, we argue that pure first-person experience cannot be a scientific foundation for IIT because science relies on taking measurements, and pure first-person experience is not measurable except through reports, brain activity, and the relationship between them. We also argue that pure first-person experience cannot be taken as ground truth because science is about backing up theories with data, not about asserting that we have ground truth independent of data. Lastly, we explain why no experiment based on third-person data can test IIT as a theory of consciousness. IIT may be a good theory of something, but not of consciousness. We conclude by exposing a deeper reason for the above conclusions: IIT's consciousness is by construction fully dissociated from any measurable thing and, for this reason, IIT implies that both the level and content of consciousness are epiphenomenal, with no causal power. IIT and other causal structure theories end up in a form of dissociative epiphenomenalism, in which we cannot even trust reports about first-person experiences. But reports about first-person experiences are taken as ground truth and the foundation for IIT's axioms. Therefore, accepting IIT leads to rejecting its own axioms. We also respond to several other criticisms against the UA.

Global and high-level effects in crowding cannot be predicted by either high-dimensional pooling or target cueing.

In visual crowding, the perception of a target deteriorates in the presence of nearby flankers. Traditionally, target-flanker interactions have been considered as local, mostly deleterious, low-level, and feature specific, occurring when information is pooled along the visual processing hierarchy. Recently, a vast literature of high-level effects in crowding (grouping effects and face-holistic crowding in particular) led to a different understanding of crowding, as a global, complex, and multilevel phenomenon that cannot be captured or explained by simple pooling models. It was recently argued that these high-level effects may still be captured by more sophisticated pooling models, such as the Texture Tiling model (TTM). Unlike simple pooling models, the high-dimensional pooling stage of the TTM preserves rich information about a crowded stimulus and, in principle, this information may be sufficient to drive high-level and global aspects of crowding. In addition, it was proposed that grouping effects in crowding may be explained by post-perceptual target cueing. Here, we extensively tested the predictions of the TTM on the results of six different studies that highlighted high-level effects in crowding. Our results show that the TTM cannot explain any of these high-level effects, and that the behavior of the model is equivalent to a simple pooling model. In addition, we show that grouping effects in crowding cannot be predicted by post-perceptual factors, such as target cueing. Taken together, these results reinforce once more the idea that complex target-flanker interactions determine crowding and that crowding occurs at multiple levels of the visual hierarchy.

A comparative biology approach to DNN modeling of vision: A focus on differences, not similarities.

Deep neural networks (DNNs) have revolutionized computer science and are now widely used for neuroscientific research. A hot debate has ensued about the usefulness of DNNs as neuroscientific models of the human visual system; the debate centers on to what extent certain shortcomings of DNNs are real failures and to what extent they are redeemable. Here, we argue that the main problem is that we often do not understand which human functions need to be modeled and, thus, what counts as a falsification. Hence, not only is there a problem on the DNN side, but there is also one on the brain side (i.e., with the explanandum-the thing to be explained). For example, should DNNs reproduce illusions? We posit that we can make better use of DNNs by adopting an approach of comparative biology by focusing on the differences, rather than the similarities, between DNNs and humans to improve our understanding of visual information processing in general.

Dissecting (un)crowding.

In crowding, perception of a target deteriorates in the presence of nearby flankers. Surprisingly, perception can be rescued from crowding if additional flankers are added (uncrowding). Uncrowding is a major challenge for all classic models of crowding and vision in general, because the global configuration of the entire stimulus is crucial. However, it is unclear which characteristics of the configuration impact (un)crowding. Here, we systematically dissected flanker configurations and showed that (un)crowding cannot be easily explained by the effects of the sub-parts or low-level features of the stimulus configuration. Our modeling results suggest that (un)crowding requires global processing. These results are well in line with previous studies showing the importance of global aspects in crowding.

Shrinking Bouma's window: How to model crowding in dense displays.

In crowding, perception of a target deteriorates in the presence of nearby flankers. Traditionally, it is thought that visual crowding obeys Bouma's law, i.e., all elements within a certain distance interfere with the target, and that adding more elements always leads to stronger crowding. Crowding is predominantly studied using sparse displays (a target surrounded by a few flankers). However, many studies have shown that this approach leads to wrong conclusions about human vision. Van der Burg and colleagues proposed a paradigm to measure crowding in dense displays using genetic algorithms. Displays were selected and combined over several generations to maximize human performance. In contrast to Bouma's law, only the target's nearest neighbours affected performance. Here, we tested various models to explain these results. We used the same genetic algorithm, but instead of selecting displays based on human performance we selected displays based on the model's outputs. We found that all models based on the traditional feedforward pooling framework of vision were unable to reproduce human behaviour. In contrast, all models involving a dedicated grouping stage explained the results successfully. We show how traditional models can be improved by adding a grouping stage.

Response to commentaries on 'hard criteria for empirical theories of consciousness'.

In consciousness research, we have a very large number of theories, which exceeds by far the number of theories in other fields. We recently presented a set of criteria for evaluating and comparing theories of consciousness, and then applied the criteria to a number of different theories. Our publication sparked strong responses as evident by the many comments published in Cognitive Neuroscience (this issue). Overall, there seems to be consensus that a theory of consciousness (ToC) needs to have an unconscious alternative, but other criteria sparked controversy. The hottest debate is to what extent consciousness needs to work with purely 1st person data, containing information not available in 3rd person reports. We would like to thank all the commentators for their lively input and we look forward to continued dialog as theories evolve and compete.

Hard criteria for empirical theories of consciousness.

Consciousness is now a well-established field of empirical research. A large body of experimental results has been accumulated and is steadily growing. In parallel, many Theories of Consciousness (ToCs) have been proposed. These theories are diverse in nature, ranging from computational to neurophysiological and quantum theoretical approaches. This contrasts with other fields of natural science, which host a smaller number of competing theories. We suggest that one reason for this abundance of extremely different theories may be the lack of stringent criteria specifying how empirical data constrains ToCs. First, we argue that consciousness is a well-defined topic from an empirical point of view and motivate a purely empirical stance on the quest for consciousness. Second, we present a checklist of criteria that, we propose, empirical ToCs need to cope with. Third, we review 13 of the most influential ToCs and subject them to the criteria. Our analysis helps to situate these different ToCs in the theoretical landscapeand sheds light on their strengths and weaknesses from a strictly empirical point of view.

All in Good Time: Long-Lasting Postdictive Effects Reveal Discrete Perception.

Is consciousness a continuous stream of percepts or is it discrete, occurring only at certain moments in time? This question has puzzled philosophers, psychologists, and neuroscientists for centuries. Both hypotheses have fallen repeatedly in and out of favor. Here, we review recent studies exploring long-lasting postdictive effects and show that the results favor a two-stage discrete model, in which substantial periods of continuous unconscious processing precede discrete conscious percepts. We propose that such a model marries the advantages of both continuous and discrete models and resolves centuries old debates about perception and consciousness.

Capsule networks as recurrent models of grouping and segmentation.

Classically, visual processing is described as a cascade of local feedforward computations. Feedforward Convolutional Neural Networks (ffCNNs) have shown how powerful such models can be. However, using visual crowding as a well-controlled challenge, we previously showed that no classic model of vision, including ffCNNs, can explain human global shape processing. Here, we show that Capsule Neural Networks (CapsNets), combining ffCNNs with recurrent grouping and segmentation, solve this challenge. We also show that ffCNNs and standard recurrent CNNs do not, suggesting that the grouping and segmentation capabilities of CapsNets are crucial. Furthermore, we provide psychophysical evidence that grouping and segmentation are implemented recurrently in humans, and show that CapsNets reproduce these results well. We discuss why recurrence seems needed to implement grouping and segmentation efficiently. Together, we provide mutually reinforcing psychophysical and computational evidence that a recurrent grouping and segmentation process is essential to understand the visual system and create better models that harness global shape computations.

How stable is perception in #TheDress and #TheShoe?

#TheDress is perceived by some people as black and blue while others perceive it as white and gold. We have previously shown that the first encounter with #TheDress strongly biases its perception. This percept remained stable during the experiment, suggesting a role of one-shot learning. #TheShoe is another image that elicits similar bimodal color percepts. Here, we investigated how percepts change over time in both #TheShoe and #TheDress. First, we show that the important role of one-shot learning, which we found for #TheDress extends to #TheShoe. Similarly to our previous results with the dress, hiding large parts of the image with occluders biased the percept of the shoe. The percept did not change for the majority of observers when the occluders were removed. Second, we investigated if and how percepts switch over a time course of 14 days. We found that although some observers experienced percept switches, the percept was largely stable for most observers.

Feature integration within discrete time windows.

Sensory information must be integrated over time to perceive, for example, motion and melodies. Here, to study temporal integration, we used the sequential metacontrast paradigm in which two expanding streams of lines are presented. When a line in one stream is offset observers perceive all other lines to be offset too, even though they are straight. When more lines are offset the offsets integrate mandatorily, i.e., observers cannot report the individual offsets. We show that mandatory integration lasts for up to 450 ms, depending on the observer. Importantly, integration occurs only when offsets are presented within a discrete window of time. Even stimuli that are in close spatio-temporal proximity do not integrate if they are in different windows. A window of integration starts with stimulus onset and integration in the next window has similar characteristics. We present a two-stage computational model based on discrete time windows that captures these effects.

The unfolding argument: Why IIT and other causal structure theories cannot explain consciousness.

How can we explain consciousness? This question has become a vibrant topic of neuroscience research in recent decades. A large body of empirical results has been accumulated, and many theories have been proposed. Certain theories suggest that consciousness should be explained in terms of brain functions, such as accessing information in a global workspace, applying higher order to lower order representations, or predictive coding. These functions could be realized by a variety of patterns of brain connectivity. Other theories, such as Information Integration Theory (IIT) and Recurrent Processing Theory (RPT), identify causal structure with consciousness. For example, according to these theories, feedforward systems are never conscious, and feedback systems always are. Here, using theorems from the theory of computation, we show that causal structure theories are either false or outside the realm of science.

Beyond Bouma's window: How to explain global aspects of crowding?

In crowding, perception of an object deteriorates in the presence of nearby elements. Although crowding is a ubiquitous phenomenon, since elements are rarely seen in isolation, to date there exists no consensus on how to model it. Previous experiments showed that the global configuration of the entire stimulus must be taken into account. These findings rule out simple pooling or substitution models and favor models sensitive to global spatial aspects. In order to investigate how to incorporate global aspects into models, we tested a large number of models with a database of forty stimuli tailored for the global aspects of crowding. Our results show that incorporating grouping like components strongly improves model performance.

Building perception block by block: a response to Fekete et al.

Is consciousness a continuous stream, or do percepts occur only at certain moments of time? This age-old question is still under debate. Both positions face difficult problems, which we proposed to overcome with a 2-stage model, where unconscious processing continuously integrates information before a discrete, conscious percept occurs. Recently, Fekete et al. criticized our model. Here, we show that, contrary to their proposal, simple sliding windows cannot explain apparent motion and related phenomena within a continuous framework, and that their supervenience argument only holds true for qualia realists, a philosophical position we do not adopt.

The role of one-shot learning in #TheDress.

#TheDress is remarkable in two aspects. First, there is a bimodal split of the population in the perception of the dress's colors (white/gold vs. black/blue). Second, whereas interobserver variance is high, intra-observer variance is low, i.e., the percept rarely switches in a given individual. There are two plausible routes of explanations: either one-shot learning during the first presentation of the image splits observers into two different, stable populations, or the differences are caused by stable traits of observers, such as different visual systems. Here, we hid large parts of the image by white occluders. The majority of naïve participants perceived the dress as black and blue. With black occluders, the majority of observers perceived the dress as white and gold. The percept did not change when we subsequently presented the full image, arguing for a crucial role of one-shot learning. Next, we investigated whether the first fixation determines the perceived color in naïve observers. We found no such effect. It remains thus a puzzling question where the source of variability in the different percepts comes from.

How best to unify crowding?

In crowding, the perception of an object deteriorates in the presence of nearby elements. Obviously, crowding is a ubiquitous phenomenon, as elements are rarely seen in isolation. One of the main characteristics of crowding is that the elements themselves are not rendered invisible, but their features are averaged[1] or substituted[2] with those of neighboring elements. Recently, Harrison and Bex [3] presented "A Unifying Model of Orientation Crowding in Peripheral Vision", which elegantly explains these two characteristics of crowding with one unifying mechanism. They tested their model using a new crowding paradigm and demonstrated an excellent match between human and model results. A key prediction of their model is that a higher number of flankers leads to stronger crowding, simply because more non-target features contribute to the model's output and thus deteriorate performance. However, several recent studies have shown that increasing the number of flankers can actually improve performance [4-9]. Using the same experimental design as Harrison and Bex [3], we report here that adding more flankers can also improve performance in their paradigm, whereas their model predicts the opposite result. We propose that a truly unified model of crowding must include a grouping stage.