Context-dependent deep neural network HMMs have been shown to achieve recognition accuracy superior to Gaussian mixture models in a number of recent works. Typically, neural networks are optimized with stochastic gradient descent. On large datasets, stochastic gradient descent improves quickly during the beginning of the optimization. But since it does not make use of second order information, ...

We propose a novel regularizer when training an auto-encoder for unsupervised feature extraction. We explicitly encourage the latent representation to contract the input space by regularizing the norm of the Jacobian (analytically) and the Hessian (stochastically) of the encoder’s output with respect to its input, at the training points. While the penalty on the Jacobian’s norm ensures robustne...

The relationship between the n-dimensional surfaces of smooth, strictly convex objects and the m-dimensional surfaces of their orthogonal projections, or shadows, is investigated. Our main results concern the relationships between the local properties of the surface at a point and those of its shadow. Specifically, the curvature Hessian of the projection at a boundary point is shown to be simpl...

Since Milnor’s discovery of exotic spheres [Mi], one of the most intriguing problems in Riemannian geometry has been whether there are exotic spheres with positive curvature. It is well known that there are exotic spheres that do not even admit metrics with positive scalar curvature [Hi] . On the other hand, there are many examples of exotic spheres with positive Ricci curvature (cf. [Ch1], [He...

We improve Chen-Ricci inequalities for a Lagrangian submanifold Mn of dimension n (n 2) in a 2n -dimensional complex space form M̃2n(4c) of constant holomorphic sectional curvature 4c with a semi-symmetric metric connection and a Legendrian submanifold Mn in a Sasakian space form M̃2n+1(c) of constant φ -sectional curvature c with a semi-symmetric metric connection, respectively.

Doubly stochastic matrix plays an essential role in several areas such as statistics and machine learning. In this paper we consider the optimal approximation of a square set doubly matrices. A structured BFGS method is proposed to solve dual primal problem. The resulting algorithm builds curvature information into diagonal components true Hessian, so that it takes only additional linear cost o...

We obtain a conceptually new differential geometric proof of P.F. Klembeck’s result (cf. [9]) that the holomorphic sectional curvature kg(z) of the Bergman metric of a strictly pseudoconvex domain Ω ⊂ C approaches −4/(n + 1) (the constant sectional curvature of the Bergman metric of the unit ball) as z → ∂Ω.

We analyze the limit of the p-form Laplacian under a collapse with bounded sectional curvature and bounded diameter to a singular limit space. As applications, we give results about upper and lower bounds on the j-th eigenvalue of the p-form Laplacian, in terms of sectional curvature and diameter.

In this paper we describe a new shape-from-shading method. We show how the parallel transport of surface normals can be used to impose curvature consistency and also to iteratively update surface normal directions so as to improve the brightness error. We commence by showing how to make local estimates of the Hessian matrix from surface normal information. With the local Hessian matrix to hand,...

Here s is the scalar curvature, D2s the Hessian of s, ∆s = trD2s the Laplacian, and ◦ R the action of the curvature tensor R on symmetric bilinear forms, c.f. [B, Ch.4H] for further details. The equation (0.3) is just the trace of (0.2). It is obvious from the trace equation (0.3) that there are no non-flat R2 critical metrics, i.e. solutions of (0.2)-(0.3), on compact manifolds N ; this follow...