ELODI: Ensemble Logit Difference Inhibition for Positive-Congruent Training.

Negative flips are errors introduced in a classification system when a legacy model is updated. Existing methods to reduce the negative flip rate (NFR) either do so at the expense of overall accuracy by forcing a new model to imitate the old models, or use ensembles, which multiply inference cost prohibitively. We analyze the role of ensembles in reducing NFR and observe that they remove negative flips that are typically not close to the decision boundary, but often exhibit large deviations in the distance among their logits. Based on the observation, we present a method, called Ensemble Logit Difference Inhibition (ELODI), to train a classification system that achieves paragon performance in both error rate and NFR, at the inference cost of a single model. The method distills a homogeneous ensemble to a single student model which is used to update the classification system. ELODI also introduces a generalized distillation objective, Logit Difference Inhibition (LDI), which only penalizes the logit difference of a subset of classes with the highest logit values. On multiple image classification benchmarks, model updates with ELODI demonstrate superior accuracy retention and NFR reduction.

Person Re-Identification With Deep Kronecker-Product Matching and Group-Shuffling Random Walk.

Person re-identification (re-ID) aims to robustly measure visual affinities between person images. It has wide applications in intelligent surveillance by associating same persons' images across multiple cameras. It is generally treated as an image retrieval problem: given a probe person image, the affinities between the probe image and gallery images (P2G affinities) are used to rank the retrieved gallery images. There exist two main challenges for effectively solving this problem. 1) Person images usually show significant variations because of different person poses and viewing angles. The spatial layouts and correspondences between person images are therefore vital information for tackling this problem. State-of-the-art methods either ignore such spatial variation or utilize extra pose information for handling the challenge. 2) Most existing person re-ID methods rank gallery images considering only P2G affinities but ignore the affinities between the gallery images (G2G affinity). Such affinities could provide important clues for accurate gallery image ranking but were only utilized in post-processing stages by current methods. In this article, we propose a unified end-to-end deep learning framework to tackle the two challenges. For handling viewpoint and pose variations between compared person images, we propose a novel Kronecker Product Matching operation to match and warp feature maps of different persons. Comparing warped feature maps results in more accurate P2G affinities. To fully utilize all available P2G and G2G affinities for accurately ranking gallery person images, a novel group-shuffling random walk operation is proposed. Both Kronecker Product Matching and Group-shuffling Random Walk operations are end-to-end trainable and are shown to improve the learned visual features if integrated in the deep learning framework. The proposed approach outperforms state-of-the-art methods on Market-1501, CUHK03 and DukeMTMC datasets, which demonstrates the effectiveness and generalization ability of our proposed approach. Code is available at https://github.com/YantaoShen/kpm_rw_person_reid.

Non-linearity of Skin Properties in Electrotactile Applications: Identification and Mitigation.

Electrotactile displays can open a new sensory substitution channel to be utilized in a vast array of applications. Our Finger-Eye research used this approach to build a system for the blind to easily read any text not written in Braille. But there are still challenges in different aspects of such systems. One of the most crucial concerns, is the effects of receptor fatigue. Our tests show that during prolonged exposure of receptors to the electrical signals, their sensitivity to the signal level changes gradually and adjustments in the signal's power are required to keep the receptors is the stimulated state. This was confirmed by monitoring the electrical current passing through the skin and calculating the corresponding impedance. More interestingly, the rates of the impedance changes are different for each part of the skin, indicating inconsistent rates of receptor fatigue for each region of the skin. These electrical properties of the skin were addressed in this research for the purpose of rendering consistent sensations for the users regardless of the person or skin conditions. To solve these challenges, two methods are employed: a voltage control system based on pulse-width modulation is used to adjust the signal power; and Kalman filtering is used to predict impedance changes in advance and supply the skin with the proper signal. The result is a self-contained automated system capable of managing the signal power for any user at any given time or skin condition.

Light at night disrupts nocturnal rest and elevates glucocorticoids at cool color temperatures.

Nighttime light pollution is quickly becoming a pervasive, global concern. Since the invention and proliferation of light-emitting diodes (LED), it has become common for consumers to select from a range of color temperatures of light with varying spectra. Yet, the biological impacts of these different spectra on organisms remain unclear. We tested if nighttime illumination of LEDs, at two commercially available color temperatures (3000 and 5000 K) and at ecologically relevant illumination levels affected body condition, food intake, locomotor activity, and glucocorticoid levels in zebra finches (Taeniopygia guttata). We found that individuals exposed to 5000 K light had higher rates of nighttime activity (peaking after 1 week of treatment) compared to 3000 K light and controls (no nighttime light). Birds in the 5000 K treatment group also had increased corticosterone levels from pretreatment levels compared to 3000 K and control groups but no changes in body condition or food intake. Individuals that were active during the night did not consequently decrease daytime activity. This study adds to the growing evidence that the spectrum of artificial light at night is important, and we advocate the use of nighttime lighting with warmer color temperatures of 3000 K instead of 5000 K to decrease energetic costs for avian taxa.

PSD microscopy: a new technique for adaptive local scanning of microscale objects.

A position-sensitive detector/device (PSD) is a sensor that is capable of tracking the location of a laser beam on its surface. PSDs are used in many scientific instruments and technical applications including but not limited to atomic force microscopy, human eye movement monitoring, mirrors or machine tool alignment, vibration analysis, beam position control and so on. This work intends to propose a new application using the PSD. That is a new microscopy system called scanning PSD microscopy. The working mechanism is about putting an object on the surface of the PSD and fast scanning its area with a laser beam. To achieve a high degree of accuracy and precision, a reliable framework was designed using the PSD. In this work, we first tried to improve the PSD reading and its measurement performance. This was done by minimizing the effects of noise, distortion and other disturbing parameters. After achieving a high degree of confidence, the microscopy system can be implemented based on the improved PSD measurement performance. Later to improve the scanning efficiency, we developed an adaptive local scanning system to scan the whole area of the PSD in a short matter of time. It was validated that our comprehensive and adaptive local scanning method can shorten the scanning time in order of hundreds of times in comparison with the traditional raster scanning without losing any important information about the scanned 2D objects. Methods are also introduced to scan very complicated objects with bifurcations and crossings. By incorporating all these methods, the new microscopy system is capable of scanning very complicated objects in the matter of a few seconds with a resolution that is in order of a few micrometers.

Improving low-cost inertial-measurement-unit (IMU)-based motion tracking accuracy for a biomorphic hyper-redundant snake robot.

This paper develops and experimentally validates a 3D-printed snake robot prototype. Its structure is designed to allocate limited room for each functional module (including an external power module, battery power module, the wireless control and transmission module and some detective sensors), so as to ensure the snake robot works in different environments. In order to control and track the snake robot, a low-cost MEMS-IMU (micro-electro-mechanical systems inertial measurement unit)-based snake robot motion tracking system is developed. Three algorithms (low-pass filter, baseline calibration, and Kalman filter) are used to eliminate noise from IMU's acceleration data, thus minimizing the noise influence to tracking accuracy. Through signal processing, the IMU acceleration data can be effectively used for motion tracking. The result from the video tracking software is employed as a reference for comparison, so as to evaluate the motion tracking algorithm efficiency. The comparison results demonstrate high efficiency of the proposed IMU-based motion tracking algorithm.

Dynamically characterizing bioimpedance of fingertip skin through a developed CVD based electrotactile rendering system.

In this paper, by combining a newly developed constant-voltage-driver (CVD) based electrotactile rendering system with an on-line identification method, parameters of the resistor-capacitor (R-C) load bioimpedance model of fingertip skin are dynamically characterized and analyzed. The CVD rendering system is capable of producing varying stimulation voltage/current waveforms to fingertip. The proposed on-line identification method is a discrete-time extended least squares (ELS) iterative approach with forgetting factor (FF), which serves as an adaptive law to estimate parameters of the bioimpedance model of fingertip skin during performing the stimulation current tracking control in z-domain. Experimental results demonstrate dynamic characteristics of the identified fingertip skin bioimpedance for the sampled population.

Electro-tactile preference identification using fuzzy logic.

Electro-tactile based rehabilitation systems must be capable of self-tuning to suit the tactile preference of different users. However, tactile preference is difficult to assess in practice. We propose a Takagi-Sugeno-Kang (TSK) fuzzy logic modeling and control approach for the on-line assessment of tactile preference. The method relies on real-time measurements of voltage and power absorbed by the fingertip. Our results show that the fuzzy logic approach successfully models user tactile preference. We are currently developing an electro-tactile based Braille display (E-Braille) for assisting the Blind and Visually Impaired (BVI) that exploits our fuzzy model.

Quantitative mechanical evaluation and analysis of Drosophila embryos through the stages of embryogenesis.

The fruit fly Drosophila embryo is one of the most important model organisms in genetics and developmental biology research. To better understand the biomechanical properties involved in Drosophila embryo research, this work presents a mechanical characterization of living Drosophila embryos through the stages of embryogenesis. Measurements of the mechanical forces of Drosophila embryos are implemented using a novel, in situ, and minimally invasive force sensing tool with a resolution in the range of microN. The measurements offer an essential understanding of penetration force profiles during the microinjection of Drosophila embryos. Sequentially quantitative evaluation and analysis of the mechanical properties, such as Young's modulus, stiffness, and mechanical impedance of living Drosophila embryos are performed by extracting the force measurements throughout the stages of embryogenesis. Experimental results illustrate the changing mechanical properties of Drosophila embryos during development, and thus mathematical models are proposed. The evaluation provides a critical step toward better understanding of the biomechanical properties of Drosophila embryos during embryogenesis, and could contribute to more efficient and significant genetic and embryonic development research on Drosophila.