Explore the vast range of titles available.

Distinct scientific theories can make similar predictions. To adjudicate between theories, we must design experiments for which the theories make distinct predictions. Here we consider the problem of comparing deep neural networks as models of human visual recognition. To efficiently compare models’ ability to predict human responses, we synthesize controversial stimuli: images for which different models produce distinct responses.We applied this approach to two visual recognition tasks, handwritten digits (MNIST) and objects in small natural images (CIFAR-10). For each task, we synthesized controversial stimuli to maximize the disagreement among models which employed different architectures and recognition algorithms. Human subjects viewed hundreds of these stimuli, as well as natural examples, and judged the probability of presence of each digit/object category in each image. We quantified how accurately each model predicted the human judgments. The best-performing models were a generative analysis-by-synthesis model (based on variational autoencoders) for MNIST and a hybrid discriminative–generative joint energy model for CIFAR-10. These deep neural networks (DNNs), which model the distribution of images, performed better than purely discriminative DNNs, which learn only to map images to labels. None of the candidate models fully explained the human responses. Controversial stimuli generalize the concept of adversarial examples, obviating the need to assume a ground-truth model. Unlike natural images, controversial stimuli are not constrained to the stimulus distribution models are trained on, thus providing severe out-of-distribution tests that reveal the models’ inductive biases. Controversial stimuli therefore provide powerful probes of discrepancies between models and human perception.

Reconstructing natural images and decoding their semantic category from fMRI brain recordings is challenging. Acquiring sufficient pairs of images and their corresponding fMRI responses, which span the huge space of natural images, is prohibitive. We present a novel self-supervised approach that goes well beyond the scarce paired data, for achieving both: (i) state-of-the art fMRI-to-image reconstruction, and (ii) first-ever large-scale semantic classification from fMRI responses. By imposing cycle consistency between a pair of deep neural networks (from image-to-fMRI & from fMRI-to-image), we train our image reconstruction network on a large number of “unpaired” natural images (images without fMRI recordings) from many novel semantic categories. This enables to adapt our reconstruction network to a very rich semantic coverage without requiring any explicit semantic supervision. Specifically, we find that combining our self-supervised training with high-level perceptual losses, gives rise to new reconstruction & classification capabilities. In particular, this perceptual training enables to classify well fMRIs of never-before-seen semantic classes, without requiring any class labels during training. This gives rise to: (i) Unprecedented image-reconstruction from fMRI of never-before-seen images (evaluated by image metrics and human testing), and (ii) Large-scale semantic classification of categories that were never-before-seen during network training. Such large-scale (1000-way) semantic classification from fMRI recordings has never been demonstrated before. Finally, we provide evidence for the biological consistency of our learned model.

This paper recounts the environmental history of a main waterway in Northern Israel—the Kishon, and deploys this history to examine the evolution of Israel water policy as it struggled to bridge the growing gap between its ambitions of development and the realities of its limited water supply. The first part of the paper describes the decay of the Kishon since the early 1950s, and the multiple scientific, political and legal attempts to alleviate its misfortunes, and discusses the reasons for their failings. Some of these reasons were administrative by nature, but the paper suggests a deeper reason, rooted in the ideological core of the infant state that was overwhelmingly concerned with the development of its infrastructure, and invited the pioneering Israeli society to consider the demise of the Kishon as a necessary sacrifice for progress. The second part of the paper describes the late-20th century developments that allowed for the recovery of the ailing river. Changing social mores, the growing importance of environmental politics, the advance of Israel’s water technologies, and an environmental scandal that endowed the rehabilitation of the Kishon with a new political and moral meaning, have all contributed to the rehabilitation of the river. Once a testament for the sacrifices involved in a struggle to create a viable state, the Kishon has become a theater for a confident society that has triumphed in its struggle against nature.

Primates can recognize objects despite 3D geometric variations such as in-depth rotations. The computational mechanisms that give rise to such invariances are yet to be fully understood. A curious case of partial invariance occurs in the macaque face-patch AL and in fully connected layers of deep convolutional networks in which neurons respond similarly to mirror-symmetric views (e.g. left and right profiles). Why does this tuning develop? Here, we propose a simple learning-driven explanation for mirror-symmetric viewpoint tuning. We show that mirror-symmetric viewpoint tuning for faces emerges in the fully connected layers of convolutional deep neural networks trained on object recognition tasks, even when the training dataset does not include faces. First, using 3D objects rendered from multiple views as test stimuli, we demonstrate that mirror-symmetric viewpoint tuning in convolutional neural network models is not unique to faces: it emerges for multiple object categories with bilateral symmetry. Second, we show why this invariance emerges in the models. Learning to discriminate among bilaterally symmetric object categories induces reflection-equivariant intermediate representations. AL-like mirror-symmetric tuning is achieved when such equivariant responses are spatially pooled by downstream units with sufficiently large receptive fields. These results explain how mirror-symmetric viewpoint tuning can emerge in neural networks, providing a theory of how they might emerge in the primate brain. Our theory predicts that mirror-symmetric viewpoint tuning can emerge as a consequence of exposure to bilaterally symmetric objects beyond the category of faces, and that it can generalize beyond previously experienced object categories.

Acute kidney injury (AKI) is a frequent complication among hospitalized children, particularly during critical illness, and is associated with higher mortality rates and prolonged hospitalization. Accumulating evidence indicates that the consequences of pediatric AKI extend beyond the acute hospitalization and have strong associations with subsequent adverse kidney outcomes and hypertension. A review of the pediatric literature reveals a paucity of data defining the optimal timing, frequency, and components of post-AKI follow-up in children to mitigate long-term morbidity. In this review, we will address the potential sequelae of AKI among hospitalized children, including the temporal spectrum of acute kidney disease and chronic kidney disease. We will further discuss risk factors associated with adverse long-term outcomes and recommended follow-up surveillance after childhood AKI. In addition, we will review emerging biomarkers that may help identify children at higher risk for adverse outcomes following AKI.

We hardly notice our eye blinks, yet an externally generated retinal interruption of a similar duration is perceptually salient. We examined the neural correlates of this perceptual distinction using intracranially measured ECoG signals from the human visual cortex in 14 patients. In early visual areas (V1 and V2), the disappearance of the stimulus due to either invisible blinks or salient blank video frames ('gaps') led to a similar drop in activity level, followed by a positive overshoot beyond baseline, triggered by stimulus reappearance. Ascending the visual hierarchy, the reappearance-related overshoot gradually subsided for blinks but not for gaps. By contrast, the disappearance-related drop did not follow the perceptual distinction – it was actually slightly more pronounced for blinks than for gaps. These findings suggest that blinks' limited visibility compared with gaps is correlated with suppression of blink-related visual activity transients, rather than with \"filling-in\" of the occluded content during blinks. The average person blinks once every few seconds, each time shutting off their view of the world for about a tenth of a second. Nevertheless, we rarely notice a blink. By contrast, we readily notice a single blank frame in a movie, even if the frame lasts far less than a blink. The fact that we do not usually notice our spontaneous blinks is a striking example of the discrepancy between the images we perceive versus the information that enters our eyes. This dissociation between the information that the eyes receive and what we perceive raises a number of questions. First, which brain areas represent the actual information from the eyes, and at what point do brain areas start to represent our subjective perception instead? Second, how does the brain \"stabilize\" our perception of vision despite the frequent interruptions that occur whenever we blink? In short, does the brain \"fill in\" the missing images or “edit out” the gaps? To answer these questions, Golan et al. turned to human patients who were undergoing a surgical procedure related to the treatment of epilepsy. In the course of such procedures, and strictly for diagnosis purposes, electrodes are temporarily placed directly on the surface of the brain – the cortex – making it possible to monitor the activity of individual cortical areas. Towards the back of the brain, where cortical processing of visual signals begins, neurons responded in a way that was consistent with the physical information the eye actually received rather than the perception of vision. Thus, neurons showed the same responses to easily seen blank frames in a movie as to unnoticeable blinks. However, as the signals streamed forward to down-stream brain regions involved in vision, neurons in successive areas were increasingly likely to distinguish between the perceptually visible blank frames versus the invisible blinks. Unexpectedly, Golan et al. found no evidence that the brain fills in the missing picture during blinks. Instead, it seems that the brain generates a continuous perception by actively \"deleting\" the brief neural signals that are turned on when our visual input has been shut off. The brain only does this for blinks but not for artificial interruptions – such as blank movie frames – which explains why we notice the latter but not the former. A future challenge will be to isolate the pathway that leads from the brain regions that generate blinks to the regions that deal with vision, and that enables us to tell blinks from blanks.

A key hallmark of visual perceptual awareness is robustness to instabilities arising from unnoticeable eye and eyelid movements. In previous human intracranial (iEEG) work (Golan et al., 2016) we found that excitatory broadband high-frequency activity transients, driven by eye blinks, are suppressed in higher-level but not early visual cortex. Here, we utilized the broad anatomical coverage of iEEG recordings in 12 eye-tracked neurosurgical patients to test whether a similar stabilizing mechanism operates following small saccades. We compared saccades (1.3°−3.7°) initiated during inspection of large individual visual objects with similarly-sized external stimulus displacements. Early visual cortex sites responded with positive transients to both conditions. In contrast, in both dorsal and ventral higher-level sites the response to saccades (but not to external displacements) was suppressed. These findings indicate that early visual cortex is highly unstable compared to higher-level visual regions which apparently constitute the main target of stabilizing extra-retinal oculomotor influences.

Are scientific expert witnesses partisans, or spokesmen for objective science? This ambiguity has troubled the relations between scientists and the legal system for more than 200 years. Modern expert testimony first appeared in the late eighteenth century, and while its use steadily increased throughout the nineteenth century, in cases involving everything from patents to X-rays, the respect paid to it steadily declined, inside and outside of the courtroom. With deep learning and wry humor, Tal Golan tells stories of courtroom drama and confusion and media jeering on both sides of the Atlantic, until the start of the twenty-first century, as the courts still search for ways that will allow them to distinguish between good and bad science.

Neural network language models appear to be increasingly aligned with how humans process and generate language, but identifying their weaknesses through adversarial examples is challenging due to the discrete nature of language and the complexity of human language perception. We bypass these limitations by turning the models against each other. We generate controversial sentence pairs where two language models disagree about which sentence is more likely to occur. Considering nine language models (including n -gram, recurrent neural networks and transformers), we created hundreds of controversial sentence pairs through synthetic optimization or by selecting sentences from a corpus. Controversial sentence pairs proved highly effective at revealing model failures and identifying models that aligned most closely with human judgements of which sentence is more likely. The most human-consistent model tested was GPT-2, although experiments also revealed substantial shortcomings in its alignment with human perception. With the advances in neural language models, the question arises if some models align better with human processing than others. Golan et al. identify sentences that language models disagree about and use them to compare the shortcomings of different language models.