We introduce the multi-prediction deep Boltzmann machine (MP-DBM). The MPDBM can be seen as a single probabilistic model trained to maximize a variational approximation to the generalized pseudolikelihood, or as a family of recurrent nets that share parameters and approximately solve different inference problems. Prior methods of training DBMs either do not perform well on classification tasks ...

Melanoma is the most lethal malignant tumour, and its prevalence increasing. Early detection diagnosis of skin cancer can alert patients to manage precautions dramatically improve lives people. Recently, deep learning has grown increasingly popular in extraction categorization features for effective prediction. A model learns co-adapts representations from training data point where it fails per...

We discuss general techniques, centered around the “Layerwise Separation Property” (LSP) of a planar graph problem, that allow to develop algorithms with running time c √ k|G|, given an instance G of a problem on planar graphs with parameter k. Problems having LSP include planar vertex cover, planar independent set, and planar dominating set. Extensions of our speed-up technique to basically al...

We propose a distributed approach to train deep neural networks (DNNs), which has guaranteed convergence theoretically and great scalability empirically: close to 6 times faster on instance of ImageNet data set when run with 6 machines. The proposed scheme is close to optimally scalable in terms of number of machines, and guaranteed to converge to the same optima as the undistributed setting. T...

We report on the fabrication and characterization of iron oxide nanoparticle thin film superlattices. The formation into different film morphologies is controlled by tuning the particle plus solvent-to-substrate interaction. It turns out that the wetting vs dewetting properties of the solvent before the self-assembly process during solvent evaporation plays a major role in determining the resul...

In this work, we design a neural network for recognizing emotions in speech, using the standard IEMOCAP dataset. Following the latest advances in audio analysis, we use an architecture involving both convolutional layers, for extracting highlevel features from raw spectrograms, and recurrent ones for aggregating long-term dependencies. Applying techniques of data augmentation, layerwise learnin...

This paper presents a multilayered/multidirector and shear-deformable finite-element formulation of shells for the analysis of composite laminates. The displacement field is assumed continuous across the finiteelement layers through the composite thickness. The rotation field is, however, layerwise continuous and is assumed discontinuous across these layers. This kinematic hypothesis results in...

Variable-angle-tow (VAT) composite laminates can eventually improve the mechanical performance of lightweight structures by taking advantage a larger design space compared to straight-fiber counterparts. Here, we provide scalable low- high-fidelity methodology retrieve tow angles that maximize buckling load and fundamental frequency VAT plates. A genetic algorithm is used solve optimization pro...