Search

Conditional Generative Adversarial Networks (cGANs) have enabled controllable image synthesis for many vision and graphics applications. However, recent cGANs are 1-2 orders of magnitude more compute-intensive than modern recognition CNNs. For example, GauGAN consumes 281G MACs per image, compared to 0.44G MACs for MobileNet-v3, making it difficult for interactive deployment. In this work, we propose a general-purpose compression framework for reducing the inference time and model size of the generator in cGANs. Directly applying existing compression methods yields poor performance due to the difficulty of GAN training and the differences in generator architectures. We address these challenges in two ways. First, to stabilize GAN training, we transfer knowledge of multiple intermediate representations of the original model to its compressed model and unify unpaired and paired learning. Second, instead of reusing existing CNN designs, our method finds efficient architectures via neural architecture search. To accelerate the search process, we decouple the model training and search via weight sharing. Experiments demonstrate the effectiveness of our method across different supervision settings, network architectures, and learning methods. Without losing image quality, we reduce the computation of CycleGAN by 21×, Pix2pix by 12×, MUNIT by 29×, and GauGAN by 9×, paving the way for interactive image synthesis. [ABSTRACT FROM AUTHOR]

3D neural networks are widely used in real-world applications (e.g., AR/VR headsets, self-driving cars). They are required to be fast and accurate; however, limited hardware resources on edge devices make these requirements rather challenging. Previous work processes 3D data using either voxel-based or point-based neural networks, but both types of 3D models are not hardware-efficient due to the large memory footprint and random memory access. In this paper, we study 3D deep learning from the efficiency perspective. We first systematically analyze the bottlenecks of previous 3D methods. We then combine the best from point-based and voxel-based models together and propose a novel hardware-efficient 3D primitive, Point-Voxel Convolution (PVConv). We further enhance this primitive with the sparse convolution to make it more effective in processing large (outdoor) scenes. Based on our designed 3D primitive, we introduce 3D Neural Architecture Search (3D-NAS) to explore the best 3D network architecture given a resource constraint. We evaluate our proposed method on six representative benchmark datasets, achieving state-of-the-art performance with 1.8-23.7× measured speedup. Furthermore, our method has been deployed to the autonomous racing vehicle of MIT Driverless, achieving larger detection range, higher accuracy and lower latency. [ABSTRACT FROM AUTHOR]

Inductive wear debris sensor has been proved to be an effective device for lubricant oil condition monitoring and fault diagnosis. However, the narrow internal diameter of current sensing coil, designed for high sensitivity, limits the oil throughput. To improve the throughput without sacrificing the sensitivity, a novel multichannel wear debris detection method based on time division multiplexing is presented in this work. Only one excitation signal is used to excite multiple sensing coils. Then, the signals from multiple sensing coils are combined with a serials of square waves separately, which are self-designed specially. As a result, the peak waveforms are lifted in different timeslots by the high-level voltages in squares waves. After that, the peaks of the signals are cut out and combined into one output signal. Then, we used synchronized sampling method to record the peak values of the output signal. The signals for all sensing channels are finally extracted from the recorded peak values. To validate the feasibility of the proposed method, we designed a ten-channel sensor system using time division multiplexing. Through the pseudo-dynamic and dynamic test, it is proved that the sensor could detect wear debris in different channels simultaneously and independently without sacrificing the sensitivity. In addition, the proposed method has the potential to integrate more channels into one system, which would contribute to high throughput real time lubricant oil detection. [ABSTRACT FROM AUTHOR]

Efficient deep learning inference requires algorithm and hardware codesign to enable specialization: we usually need to change the algorithm to reduce memory footprint and improve energy efficiency. However, the extra degree of freedom from the neural architecture design makes the design space much larger: it is not only about designing the hardware architecture but also codesigning the neural architecture to fit the hardware architecture. It is difficult for human engineers to exhaust the design space by heuristics. We propose design automation techniques for architecting efficient neural networks given a target hardware platform. We investigate automatically designing specialized and fast models, auto channel pruning, and auto mixed-precision quantization. We demonstrate that such learning-based, automated design achieves superior performance and efficiency than the rule-based human design. Moreover, we shorten the design cycle by 200× than previous work, so that we can afford to design specialized neural network models for different hardware platforms. [ABSTRACT FROM AUTHOR]

In this paper, the LP11 mode evolution in a few mode-polarization maintaining fiber is experimentally investigated. Selective conversion from the LP11a/LP11b mode to the LP11b/LP11a , OAM+1, and OAM−1 modes are demonstrated with varying wavelengths and fiber lengths. This property may have nice applications in spatial division multiplexing (SDM) systems, and one of them, i.e., the all-fiber compact and robust scalar orbital angular momentum (OAM) generator, which is cascaded by a mode-selective coupler and a few mode-polarization maintaining fiber, is also demonstrated with expected performance for conversion between LP01 mode and OAM+1/OAM−1. [ABSTRACT FROM PUBLISHER]

Recently, mode division multiplexing (MDM) has been introduced into optical transmission systems and networks for capacity enhancement. In this letter, we propose a bidirectional passive optical network (PON) architecture based on MDM. The bidirectional optical distribution network of the MDM-PON consists of few-mode fiber, passive mode multiplexer/demultiplexer, and few-mode circulators, all of which are characterized by low modal-crosstalk. Signals in different modes are able to be transmitted and received independently for both downstream and upstream transmission. We experimentally demonstrate the bidirectional MDM-PON transmission over 10-km two-mode optical fiber. The 10-Gb/s ON–OFF-keying signals can be independently transmitted and directly detected without coherent receiver and multiple-input multiple-output processing. Moreover, nonlinear transmission impairments are experimentally investigated. [ABSTRACT FROM PUBLISHER]