A human does not have to see all elephants to recognize an animal as an elephant. On contrast, current state-of-the-art deep learning approaches heavily depend on the variety of training samples and the capacity of the network. In practice, the size of network is always limited and it is impossible to access all the data samples. Under this circumstance, deep learning models are extremely fragile to human-imperceivable adversarial examples, which impose threats to all safety critical systems. Inspired by the association and attention mechanisms of the human brain, we propose reverse adversarial examples method that can greatly improve models' robustness on unseen data. Experiments show that our reverse adversarial method can improve accuracy on average 19.02% on ResNet18, MobileNet, and VGG16 on unseen data transformation. Besides, the proposed method is also applicable to compressed models and shows potential to compensate the robustness drop brought by model quantization - an absolute 30.78% accuracy improvement.

For a linear isometry T: C 0 (X) C 0 (Y) of finite corank, there is a cofinite subset Y 1 of Y such that Tf| Y 1= h f is a weighted composition operator and X is homeomorphic to a quotient space of Y 1 modulo a finite subset. When X= Y, such a T is called an isometric quasi-n-shift on C 0 (X). In this case, the action of T can be implemented as a shift on a tree-like structure, called a T-tree, in M (X) with exactly n joints. The T-tree is total in M (X) when T is a shift. With these tools, we can analyze the structure of T.

We present a quantum field theoretic description on the t−J model on a square lattice with dilute holes (i.e. near half-filling), based on the compact mutual Chern-Simons gauge theory. We show that, due to the presence of non-perturbative monopole plasma configuration from the antiferromagnetic background, holons (carrying electric charge) are linearly confined and strongly localized even without extrinsic disorder taken into account. Accordingly, the translation symmetry is spontaneously broken at ground state. Such an exotic localization is distinct from Anderson localization and essentially rooted in intrinsic Mott physics of the t−J model. Finally, a finite-temperature phase diagram is proposed. The metal-insulator transition observed in in-plane resistivity measurement is identified to a confinement-deconfinement transition from the perspective of gauge theory. The transition is characterized by the order parameter "Polyakov-line".

We propose a new nonmonotone filter method to promote global and fast local convergence for sequential quadratic programming algorithms. Our method uses two filters: a standard, global g-filter for global convergence, and a local nonmonotone l-filter that allows us to establish fast local convergence. We show how to switch between the two filters efficiently, and we prove global and superlinear local convergence. A special feature of the proposed method is that it does not require second-order correction steps. We present preliminary numerical results comparing our implementation with a classical filter SQP method.

We study knots in 3d Chern-Simons theory with complex gauge group SL(N,C), in the context of its relation with 3d N=2 theory (the so-called 3d-3d correspondence). The defect has either co-dimension 2 or co-dimension 4 inside the 6d (2,0) theory, which is compactified on a 3-manifold M^. We identify such defects in various corners of the 3d-3d correspondence, namely in 3d SL(N,C) Chern-Simons theory, in 3d N=2 theory, in 5d N=2 super Yang-Mills theory, and in the M-theory holographic dual. We can make quantitative checks of the 3d-3d correspondence by computing partition functions at each of these theories. This Letter is a companion to a longer paper, which contains more details and more results.