Self-supervised representation learning has made significant leaps fueled by progress in contrastive learning, which seeks to learn transformations that embed positive input pairs nearby, while pushing negative pairs far apart. While positive pairs can be generated reliably (e.g., as different views of the same image), it is difficult to accurately establish negative pairs, defined as samples from different images regardless of their semantic content or visual features. A fundamental problem in contrastive learning is mitigating the effects of false negatives. Contrasting false negatives induces two critical issues in representation learning: discarding semantic information and slow convergence. In this paper, we propose novel approaches to identify false negatives, as well as two strategies to mitigate their effect, i.e. false negative elimination and attraction, while systematically performing rigorous evaluations to study this problem in detail. Our method exhibits consistent improvements over existing contrastive learning-based methods. Without labels, we identify false negatives with ~40% accuracy among 1000 semantic classes on ImageNet, and achieve 5.8% absolute improvement in top-1 accuracy over the previous state-of-the-art when finetuning with 1% labels. Our code is available at https://github.com/google-research/fnc

We introduce a novel approach to endowing neural networks with emergent, long-term, large-scale memory. Distinct from strategies that connect neural networks to external memory banks via intricately crafted controllers and hand-designed attentional mechanisms, our memory is internal, distributed, co-located alongside computation, and implicitly addressed, while being drastically simpler than prior efforts. Architecting networks with multigrid structure and connectivity, while distributing memory cells alongside computation throughout this topology, we observe the emergence of coherent memory subsystems. Our hierarchical spatial organization, parameterized convolutionally, permits efficient instantiation of large-capacity memories, while multigrid topology provides short internal routing pathways, allowing convolutional networks to efficiently approximate the behavior of fully connected networks. Such networks have an implicit capacity for internal attention; augmented with memory, they learn to read and write specific memory locations in a dynamic data-dependent manner. We demonstrate these capabilities on exploration and mapping tasks, where our network is able to self-organize and retain long-term memory for trajectories of thousands of time steps. On tasks decoupled from any notion of spatial geometry: sorting, associative recall, and question answering, our design functions as a truly generic memory and yields excellent results.

Autonomous driving carries the promise to drastically reduce the number of car accidents; however, recently reported fatal crashes involving self-driving cars show that such an important goal is not yet achieved. This calls for better testing of the software controlling self-driving cars, which is difficult because it requires producing challenging driving scenarios. To better test self-driving car soft- ware, we propose to specifically test car crash scenarios, which are critical par excellence. Since real car crashes are difficult to test in field operation, we recreate them as physically accurate simulations in an environment that can be used for testing self-driving car software. To cope with the scarcity of sensory data collected during real car crashes which does not enable a full reproduction, we extract the information to recreate real car crashes from the police reports which document them. Our extensive evaluation, consisting of a user study involving 34 participants and a quantitative analysis of the quality of the generated tests, shows that we can generate accurate simulations of car crashes in a matter of minutes. Compared to tests which implement non critical driving scenarios, our tests effectively stressed the test subject in different ways and exposed several shortcomings in its implementation.

While deep learning has achieved great success in building vision applications for mainstream users, there is relatively less work for the blind and visually impaired to have a personal, on-device visual assistant for their daily life. Unlike mainstream applications, vision system for the blind must be robust, reliable and safe-to-use. In this paper, we propose a fine-grained currency recognizer based on CONGAS, which significantly surpasses other popular local features by a large margin. In addition, we introduce an effective and light-weight coarse classifier that gates the fine-grained recognizer on resource-constrained mobile devices. The coarse-to-fine approach is orchestrated to provide an extensible mobile-vision architecture, that demonstrates how the benefits of coordinating deep learning and local feature based methods can help in resolving a challenging problem for the blind and visually impaired. The proposed system runs in real-time with ~150ms latency on a Pixel device, and achieved 98% precision and 97% recall on a challenging evaluation set.

Computed tomography (CT) imaging is an essential tool in various clinical diagnoses and radiotherapy treatment planning. Since CT image intensities are directly related to positron emission tomography (PET) attenuation coefficients, they are indispensable for attenuation correction (AC) of the PET images. However, due to the relatively high dose of radiation exposure in CT scan, it is advised to limit the acquisition of CT images. In addition, in the new PET and magnetic resonance (MR) imaging scanner, only MR images are available, which are unfortunately not directly applicable to AC. These issues greatly motivate the development of methods for reliable estimate of CT image from its corresponding MR image of the same subject. In this paper, we propose a learning-based method to tackle this challenging problem. Specifically, we first partition a given MR image into a set of patches. Then, for each patch, we use the structured random forest to directly predict a CT patch as a structured output, where a new ensemble model is also used to ensure the robust prediction. Image features are innovatively crafted to achieve multi-level sensitivity, with spatial information integrated through only rigid-body alignment to help avoiding the error-prone inter-subject deformable registration. Moreover, we use an auto-context model to iteratively refine the prediction. Finally, we combine all of the predicted CT patches to obtain the final prediction for the given MR image. We demonstrate the efficacy of our method on two datasets: human brain and prostate images. Experimental results show that our method can accurately predict CT images in various scenarios, even for the images undergoing large shape variation, and also outperforms two state-of-the-art methods.