Richard Y. Zhang

Split-Brain Autoencoders: Unsupervised Learning by Cross-Channel Prediction

We propose split-brain autoencoders, a straightforward modification of the traditional autoencoder architecture, for unsupervised representation learning. The method adds a split to the network, resulting in two disjoint sub-networks. Each sub-network is trained to perform a difficult task – predicting one subset of the data channels from another. Together, the sub-networks extract features from the entire input signal. By forcing the network to solve cross-channel prediction tasks, we induce a representation within the network which transfers well to other, unseen tasks. This method achieves state-of-the-art performance on several large-scale transfer learning benchmarks.

Sensor fusion for semantic segmentation of urban scenes

Semantic understanding of environments is an important problem in robotics in general and intelligent autonomous systems in particular. In this paper, we propose a semantic segmentation algorithm which effectively fuses information from images and 3D point clouds. The proposed method incorporates information from multiple scales in an intuitive and effective manner. A late-fusion architecture is proposed to maximally leverage the training data in each modality. Finally, a pairwise Conditional Random Field (CRF) is used as a post-processing step to enforce spatial consistency in the structured prediction. The proposed algorithm is evaluated on the publicly available KITTI dataset [1] [2], augmented with additional pixel and point-wise semantic labels for building, sky, road, vegetation, sidewalk, car, pedestrian, cyclist, sign/pole, and fence regions. A per-pixel accuracy of 89.3% and average class accuracy of 65.4% is achieved, well above current state-of-the-art [3].

Simplified design method for litz wire

A simplified approach to choosing number and diameter of strands in litz wire is presented. Compared to previous analyses, the method is easier to use. The parameters needed are only the skin depth at the frequency of operation, the number of turns, the breadth of the core window, and a constant from a table provided in the paper. In addition, guidance is provided on litz wire construction - how many strands or sub-bundles to combine at each twisting operation. The maximum number of strands to combine in the first twisting operation is given by a simple formula requiring only the skin depth and strand diameter. Different constructions are compared experimentally.

Four-legged converter 3-D SVM scheme over-modulation study

A novel three-dimensional space vector modulation (3-D SVM) along with two different overmodulation correction schemes is proposed for three-phase four-legged converters. Although the overmodulation is desired in motor drive applications for better bus voltage utilization, it can cause voltage distortion, and thus is undesirable in utility interface applications. The overmodulation case may occur during transient conditions. When it happens, the duty cycle will saturate, and the output voltage waveform quality deteriorates. In this paper, the overmodulation surface for a 3D SVM scheme is defined, and the correction methods that keep the system from entering into overmodulation are evaluated with computer simulation.

Fast Simulation of Complicated 3-D Structures Above Lossy Magnetic Media

A fast numerical method is presented for the simulation of complicated 3-D structures, such as inductors constructed from Litz or stranded wires, above or sandwiched between the planar lossy magnetic media. Making use of its smoothness, the quasi-static multilayer Greens function is numerically computed using finite differences, and its source height dependence is computed using an optimal Toeplitz-plus-Hankel decomposition. We show that a modified precorrected fast Fourier transform method can be applied to reduce the dense linear algebra problem to near-linear time, and that frequency-dependent setups can be avoided to result in a considerable speed-up. Experimental verifications are made for a 16-strand Litz wire coil realistically modeled down to each individual strand. Results are obtained in 2-3 h, showing an excellent agreement to measurements, and can be used to study the impact of transposition patterns in Litz wire construction.

Realistic litz wire characterization using fast numerical simulations

The losses of realistic litz wires are characterized while explicitly accounting for their construction, using a procedure that computes the current-driven and magnetic-field-driven copper losses using fast numerical simulations. We present a case study that examines loss variation in one- and two-level litz wires as a function of twisting pitch, over a wide range of values and in small increments. Experimental confirmation is presented for predictions made by numerical simulations. Results confirm the capability and efficiency of numerical methods to provide valuable insights into the realistic construction of litz wire.

Automatic identification of window regions on indoor point clouds using LiDAR and cameras

In this paper, we propose an algorithm to automatically identify window regions on exterior facing facades of buildings using interior 3D point cloud resulting from an ambulatory backpack sensor system, outfitted with multiple LiDAR sensors and cameras. We develop a set of discriminative features for the task, namely visual brightness, infrared opaqueness, and an occlusion indicator, within a Markov Random Field (MRF) framework to provide structured prediction for window or glass regions. A preprocessing classifier is trained on the features to produce node potentials, and large margin parameter training is used to boost performance. Our algorithm has been trained on data taken at the 3rd floor of Cory Hall at UC Berkeley, with a total façade area of 269.1 m2, and has been tested on walls taken on the 2nd floor of Cory Hall, a Walgreens, and an office building in San Francisco, with a total exterior façade area of 454.6 m2. Window regions are successfully identified with 85.5% F1-score and 94.2% accuracy.

Design of resonance damping via control synthesis

Dampers are widely used in power electronics to damp resonances, in order to reduce device stress, power loss, and electromagnetic interference. In this paper we formulate the damping problem so as to expose the fundamental tradeoff between damping amplitude peaks and minimizing power dissipation, and then use a constrained optimization approach to compute optimal Pareto frontier as a function of damper order. We use the procedure to demonstrate the diminishing returns of increasing damper order using a simple filter example, and then we demonstrate the power of the method for multiport converters. In particular, we show that using the constrained optimization procedure dramatically outperforms the standard port-by-port method.

GMRES-Accelerated ADMM for Quadratic Objectives

We consider the sequence acceleration problem for the alternating direction method-of-multipliers (ADMM) applied to a class of equality-constrained problems with strongly convex quadratic objectives, which frequently arise as the Newton subproblem of interior-point methods. Within this context, the ADMM update equations are linear, the iterates are confined within a Krylov subspace, and the General Minimum RESidual (GMRES) algorithm is optimal in its ability to accelerate convergence. The basic ADMM method solves a

Inertial and frequency response of microgrids with induction motors

\kappa