A major challenge in recent AI literature is understanding why state-of-the-art deep learning models show great success on a range of datasets while they severely degrade in performance when presented with examples which slightly vary from their training distribution. In this proposal, we will examine this question in the context of visual question-answering, a challenging task which requires models to jointly reason over images and text. We will start our exploration with the GQA dataset, which, along with images and text, also includes a rich semantic scene graph, representing the spatial relations between objects in the image, and thus lends itself to probing through high-quality automatic manipulation. In prelminary work we have augmented GQA with examples that vary slightly from the original questions, and shown that here too high-performing models perform much worse on the augmented questions compared to the original ones. Our proposal will analyze our results, exploring the reasons for the drop in performance, and what makes our new questions more challenging. We also plan to to generalize the reasons we find to other datasets of visual question answering, and more broadly to other AI datasets. Given those insights, we will augment the training set with instances that capture model “blind spots”, in an attempt to improve the model’s generalization ability. Our results will improve our understanding of what state-of-the-art AI models know, what they are still missing, and how can we improve them based on this new understanding.

Goal

To achieve a better understanding of the limitations of current state-of-the-art models and datasets, as well as ways to improve them.

Researchers

• Gabriel Stanovsky
• Roy Schwartz

What we did

In progress

Deep machine learning is revolutionizing computer science. Instead of manually creating complex software, engineers now use automatically-generated deep neural networks (DNNs). DNNs are being used in critical financial, medical and autonomous driving systems, obtaining previously unimaginable results. Recently, they are being adopted also in the field of computer systems, replacing hand-crafted algorithms that control key infrastructure, such as adaptive bit rate selection and congestion control algorithms. But despite their remarkable achievements, DNN opacity is a tremendous challenge: DNNs work, but we do not fully understand how or why, and cannot manually prove their correctness. Consequently, there is a crucial need to ensure that DNNs operate correctly. This issue is urgent, as errors have already been observed in modern DNNs: for example, slight perturbations to inputs (“adversarial examples”) that can cause modern DNNs to perform severe misclassification errors. This lack of formal guarantees about DNN behavior is preventing their safe deployment in critical systems.
Recent and exciting developments in the field of formal methods allow us to automatically reason about DNNs. However, this is a nascent technology, and we are just now beginning to tap its full potential. Here, we propose to bring together Prof. Schapira’s expertise in computer networks and Dr. Katz’s expertise in neural network verification, in order to create techniques and tools suitable for the formal verification of learning-based computer networks policies. We will develop novel certification techniques, capable of leveraging common traits of computer network algorithms in order to achieve better flexibility and scalability than general, off-the-shelf verification tools. We will then apply these techniques to state-of-the-art controllers in computer networks. Our research will thus help ensure the safety and correctness of key computer network systems before they are deployed, greatly benefiting programmers, engineers, and users of these systems.

Goal

To combine the recent advances in learning-based controllers for computer systems with those in DNN verification, in order to ensure the robustness and reliability of these controllers.

Researchers

• Michael Schapira
• Guy Katz

What we did

In progress