For an overview of its proof, see Appendix B. Lemma A.1. In the following lemma, we use Lemma A.1 in order to show RSA T -hardness of By Assumption 2.1, there is K such that CSP K literals in the clause are satisfied by ψ, and otherwise null z, w null 1 . A.3 Hardness of learning random fully-connected neural networks Let n = ( n Let M be a diagonal-blocks matrix. By Lemma A.3, we have s By Lemma A.4, we have with probability 1 o Finally, Theorem 3.1 follows immediately from Theorem A.1 and the following lemma. By Lemma A.6, we have that By Theorem A.1, we need to show that SCAT We say that a distribution is isotropic if it has mean zero and its covariance matrix is the identity.

Despite their promising performance across various natural language processing (NLP) tasks, current NLP systems are vulnerable to textual adversarial attacks. To defend against these attacks, most existing methods apply adversarial training by incorporating adversarial examples. However, these methods have to rely on ground-truth labels to generate adversarial examples, rendering it impractical for large-scale model pre-training which is commonly used nowadays for NLP and many other tasks. In this paper, we propose a novel learning framework called SCAT (Self-supervised Contrastive Learning via Adversarial Training), which can learn robust representations without requiring labeled data. Specifically, SCAT modifies random augmentations of the data in a fully labelfree manner to generate adversarial examples. Adversarial training is achieved by minimizing the contrastive loss between the augmentations and their adversarial counterparts. We evaluate SCAT on two text classification datasets using two state-of-the-art attack schemes proposed recently. Our results show that SCAT can not only train robust language models from scratch, but it can also significantly improve the robustness of existing pre-trained language models. Moreover, to demonstrate its flexibility, we show that SCAT can also be combined with supervised adversarial training to further enhance model robustness.

At the end of my Melbourne street there's a new system being installed for traffic management. I hadn't even noticed the extra cameras, vehicle and pedestrian sensors, LiDAR and radar on the intersection, but these tools are all part of a larger system, with researchers hoping that a 2.5km section of Nicholson Street, in Carlton, will eventually be run by an artificial intelligence (AI). This might sound a little nerve-wracking to the average commuter, but these "smart corridors" are popping up around the world – systems that promise to provide us with less traffic and better safety. "Many cities around the world have dedicated corridors or smart motorways that are equipped with sensors, CCTV cameras and AI for predicting the traffic flow, speed, or occupancy at a specific moment in time," says Dr Adriana-Simona Mihaita, an AI infrastructure researcher at the University of Technology Sydney, who was not involved in the research. "Accurate predictions will provide transport operators with the means to make informed decisions and apply new control plans, or adjust the current ones according to ongoing traffic or eventual disruptions."

Context-aware machine translation models are designed to leverage contextual information, but often fail to do so. As a result, they inaccurately disambiguate pronouns and polysemous words that require context for resolution. In this paper, we ask several questions: What contexts do human translators use to resolve ambiguous words? Are models paying large amounts of attention to the same context? What if we explicitly train them to do so? To answer these questions, we introduce SCAT (Supporting Context for Ambiguous Translations), a new English-French dataset comprising supporting context words for 14K translations that professional translators found useful for pronoun disambiguation. Using SCAT, we perform an in-depth analysis of the context used to disambiguate, examining positional and lexical characteristics of the supporting words. Furthermore, we measure the degree of alignment between the model's attention scores and the supporting context from SCAT, and apply a guided attention strategy to encourage agreement between the two.

Neural networks have revolutionized performance in multip le domains, such as computer vision and natural language processing, and have proven to be a highly effectiv e tool for solving many challenging problems. This impressive practical success of neural networks is not well understood from the theoretical point of view. In particular, despite extensive research in recent y ears, it is not clear which models are learnable by neural networks algorithms. Historically, there were many negative results for learnin g neural networks, and it is now known that under certain complexity assumptions, it is computational ly hard to learn the class of functions computed by a neural network, even if the architecture is very simple. Indeed, it has been shown that learning neural networks is hard already for networks of depth 2 [ Klivans and Sherstov, 2006, Daniely and Shalev-Shwartz, 2016 ]. These results hold already for improper learning, namely where the learning algorithm is allowed to return a hypothesis that does not belong to the considered hy pothesis class. In recent years, researchers have considered several ways t o circumvent the discrepancy between those hardness results and the empirical success of neural n etworks. Namely, to understand which models are still learnable by neural networks algorithms.

Current road-traffic optimisation practice around the world is a combination of hand tuned policies with a small degree of automatic adaption. Even state-ofthe-art research controllers need good models of the road traffic, which cannot be obtained directly from existing sensors. We use a policy-gradient reinforcement learning approach to directly optimise the traffic signals, mapping currently deployed sensor observations to control signals. Our trained controllers are (theoretically) compatible with the traffic system used in Sydney and many other cities around the world. We apply two policy-gradient methods: (1) the recent natural actor-critic algorithm, and (2) a vanilla policy-gradient algorithm for comparison. Along the way we extend natural-actor critic approaches to work for distributed and online infinite-horizon problems.

Current road-traffic optimisation practice around the world is a combination of hand tuned policies with a small degree of automatic adaption. Even state-ofthe-art research controllers need good models of the road traffic, which cannot be obtained directly from existing sensors. We use a policy-gradient reinforcement learning approach to directly optimise the traffic signals, mapping currently deployed sensor observations to control signals. Our trained controllers are (theoretically) compatible with the traffic system used in Sydney and many other cities around the world. We apply two policy-gradient methods: (1) the recent natural actor-critic algorithm, and (2) a vanilla policy-gradient algorithm for comparison. Along the way we extend natural-actor critic approaches to work for distributed and online infinite-horizon problems.

Current road-traffic optimisation practice around the world is a combination of hand tuned policies with a small degree of automatic adaption. Even state-ofthe-art researchcontrollers need good models of the road traffic, which cannot be obtained directly from existing sensors. We use a policy-gradient reinforcement learningapproach to directly optimise the traffic signals, mapping currently deployed sensor observations to control signals. Our trained controllers are (theoretically) compatiblewith the traffic system used in Sydney and many other cities around the world. We apply two policy-gradient methods: (1) the recent natural actor-critic algorithm, and (2) a vanilla policy-gradient algorithm for comparison. Along the way we extend natural-actor critic approaches to work for distributed and online infinite-horizon problems.