Search

In this paper, we show that adversarial training time attacks by a few pixel modifications can cause undesirable overfitting in neural networks for both discriminative and generative models. We propose an evolutionary algorithm to search for an optimal pixel attack using a novel cost function inspired by domain adaptation literature to design our training time attack. The proposed cost function explicitly maximizes the generalization gap and domain divergence between clean and corrupted images. Empirical evaluations demonstrate that our adversarial training attack can achieve significantly low testing accuracy (with high training accuracy) on multiple datasets by just perturbing a single pixel in the training images. Even under the use of popular regularization techniques, we identify a significant performance drop compared to clean data training. Our attack is more successful than previous pixel-based training time attacks on state-of-the-art Convolutional Neural Networks (CNNs) architectures, as evidenced by significantly lower testing accuracy. Interestingly, we find that the choice of optimization plays an essential role in robustness against our attack. We empirically observe that Stochastic Gradient Descent (SGD) is resilient to the proposed adversarial training attack, different from adaptive optimization techniques such as the popular Adam optimizer. We identify that such vulnerabilities are caused due to over-reliance on the cross-entropy (CE) loss on highly predictive features. Therefore, we propose a robust loss function that maximizes the mutual information between latent features and input images, in addition to optimizing the CE loss. Finally, we show that the discriminator in Generative Adversarial Networks (GANs) can also be attacked by our proposed training time attack resulting in poor generative performance. Our paper is one of the first works to design attacks for generative models.

A real-time multi-bot system consists of a group of robots working together to accomplish a common task. Each robot in the environment is provided with its own set of sensors, communication modules, vision systems, and control systems. A robot in the multi-bot environment is given complete autonomy or partial autonomy or teleoperated based on the task. For a robot to navigate autonomously, they should be capable of knowing their own position and pose in the environment, I.e., localization [1]. Localization is achieved by the process of Data Fusion. Data fusion is the process of combining the data from sensors that sense motion and sensors that represent the environment. For a robot, the data collected from its sensors are available only to itself. In a multi-bot system, the data collected from a single robot should be made available to all the other robots in the environment. Each robot collects the sensor data of its own and shares this information among all the other robots in the environment. The connectivity and exchange of data between individual robots is a key issue in a multi-bot environment. The multi-bot system performance and safety can be increased by high speed, low latency exchange of data between individual robots. Moreover, keeping a robot informed about others in their surrounding increases the knowledge and makes each robot take a global decision in the environment rather than local decisions [2]. This experiment demonstrates the high-speed exchange of information between robots and quantifies the delay elapsed in it. This also proves ROS2 has good performance over ROS and best suited for the multi-bot system [3].

Today in this competitive world, people mostly tend to focus on students who are very focused with their career and have their ideologies and goals set but in that long run the students who have no aim in their lives and are without any ideologies or goals set but have a great potential to perform well in certain aspects are forgotten. Our paper is entirely for them. The research has been conducted in such a way so that a platform can be provided for students to interact with and set up their goals thereby exploring their pros and cons in such a way that they achieve their own goals in their respective fields and not feel left out. Here a decision tree model has been proposed to efficiently help in determining career goals of students. Later performance evaluation has been carried out to compute the effectiveness of classification.

Existing methods for distillation do not efficiently utilize the training data. This work presents a novel approach to perform distillation using only a subset of the training data, making it more data-efficient. For this purpose, the training of the teacher model is modified to include self-regulation wherein a sample in the training set is used for updating model parameters in the backward pass either if it is misclassified or the model is not confident enough in its prediction. This modification restricts the participation of samples, unlike the conventional training method. The number of times a sample participates in the self-regulated training process is a measure of its significance towards the model's knowledge. The significance values are used to weigh the losses incurred on the corresponding samples in the distillation process. This method is named significance-based distillation. Two other methods are proposed for comparison where the student model learns by distillation and incorporating self-regulation as the teacher model, either utilizing the significance information computed during the teacher's training or not. These methods are named hybrid and regulated distillations, respectively. Experiments on benchmark datasets show that the proposed methods achieve similar performance as other state-of-the-art methods for knowledge distillation while utilizing a significantly less number of samples., Comment: 8 pages, 5 figures, 6 tables, 27 references

This paper analyzes a deep learning based classification process for common East Asian dermatological conditions. We have chosen ten common categories based on prevalence. With more than 85% accuracy in our experiments, we have tried to investigate why current models are yet to reach accuracy benchmarks seen in object identification tasks. Our current attempt sheds light on how deep learning based dermoscopic identification and dataset creation could be improved.

Our paper introduces an efficient combination of established techniques to improve classifier performance, in terms of accuracy and training time. We achieve two-fold to ten-fold speedup in nearing state of the art accuracy, over different model architectures, by dynamically tuning the learning rate. We find it especially beneficial in the case of a small dataset, where reliability of machine reasoning is lower. We validate our approach by comparing our method versus vanilla training on CIFAR-10. We also demonstrate its practical viability by implementing on an unbalanced corpus of diagnostic images.

Cervical Cancer is the fourth leading cause of death due to cancer among women worldwide. Pap Smear Test is the commonly used method for Cervical Cancer screening. But Pap Smear pathology screening is very time consuming process. Therefore, an automatic detection method of nucleus of cervical cell is proposed in this paper which mainly focuses on time consumption which is an important parameter when it comes the automatic segmentation. The pre-processing is achieved using edge map with double threshold for de-noising of edges, and then segmentation of the nucleus of cervical cancer cell is achieved using Gradient Force Model and Balloon force Model. The two parametric deformable models are used to check the trade-off between the number of iterations and accuracy. Further, geometrical features like perimeter, area, eccentricity, mean intensity etc. are calculated followed by segmentation using both methods to detect whether cell is cancerous or normal. The calculated features are contrasted with each method. The experimental results shows time consumption is reduced using gradient force model in terms of number of iterations used for segmentation with the accuracy of 0.92 which is significant for clinical interpretation.

X-ray luminescent computed tomography (XLCT) is a promising new functional imaging modality based on computed tomography (CT). This imaging technique uses X-ray excitable nanophosphors to illuminate objects of interest in the visible spectrum. Though there are several validations of the underlying technology, none of them have addressed the issues of performance optimality for a given design of the imaging system. This study addresses the issue of obtaining best image quality through optimizing collimator width to balance the signal to noise ratio (SNR) and resolution. The results can be generalized as to any XLCT system employing a selective excitation scheme.

Overlapping of cervical cancerous cells and presence of debris, mucus and blood play a major issue in accurate segmentation of cells. Manual screening of overlapped cells in Pap smear slides is prone to error due to the complexity, high variation in shape and size and poor contrast of images. The automated system must be able to detect the nucleus and cytoplasm of clumped cells accurately as merging of cells is a characteristic of high stages of cervical cancer. In this paper, we propose a novel method to accurately segment the overlapping cells by dividing the whole image into many small non-overlapping pixel blocks, then extracting the texture features from Gray level co-occurrence matrix GLCM. The overlapped parts have a noticeable change in certain features which help us in selecting the area of interest which is marked explicitly and further the contours are marked using Independent level set method, accurately segmenting the cell nucleus and cytoplasm.

It is significant to analyze the dental images in order to improve and quantify medical images for correct diagnosis. Caries or cavity is one of the most prevalent diseases of the teeth. Dentists are putting the best effort to identify the problem at an earlier stage. The proposed method used in this paper is focused on the challenges faced during the root canal edge extraction from dental radiographic images, which is a major problem besides cavity detection and extraction. The image processing techniques helps to identify the caries that provide dentists with the precise results of the area affected by the caries. The proposed methodology consists of preprocessing of bitewing radiographic images using top hat bottom hat transformation followed by the sharpening filter for edge enhancement. This combinational approach provides qualitative and quantitative assessment to dentists on the presence of cavity. The caries are extracted by some morphological tools to grade the severity on the basis of some metric values. Preparatory experiments show the significance of the proposed method to extract cavity and grade its effect on the tooth.

Automatic identification of bird species based on the chirping sounds of birds was experimented using feature extraction method and classification based on support vector machines (SVMs). The proposed technique followed the extraction of cepstral features on mel scale of each audio recording from the collected standard database. Extracted mel frequency cepstral coefficients (MFCCs) formed a feature matrix. This feature matrix was then trained and tested for efficient recognition of audio events from audio test signals. 70% of the whole database was used for training purpose while the reamaining 30% for testing of samples. The classifier achieved upto 89.4% accuracy on a data set containing four species, commonly found in India.

This paper proposes a novel method for design of optimized inset-fed patch antennas using genetic algorithm. A patch antenna is thereby designed and the results were simulated to verify the resonating frequency against desired frequency, and the radiation pattern. It is a demonstration of how genetic algorithm can be used to optimize a patch antenna in accordance with various requirements and scenarios, with lower computational complexity resource utilization. The results obtained during simulation are quite promising to initiate antenna development using genetic algorithm to provide heuristic selection of designs, which best fit the purpose of antenna usage.

Achieving high gain resonating antenna at a prerequisite frequency is a real time problem. A Microstrip patch antenna is highly dependent on its designing. Although theoretically it is possible to find out the parameters for a particular frequency, the outcome is not always accurate. Since this process of combining sets of parameters is time consuming, using optimization algorithms seems a plausible option. Differential Evolution uses real parameters and uses the difference in the vectors to scan the contours of the fitness function. The function used to find the minima is the impedance function as a resonance is obtained when the impedance of antenna matches the input impedance. Unlike DE, Particle swarm optimization defines a set of locations and velocities of the members of vector array, using parallel search techniques starting with many vectors from initial position and communicating amongst themselves to find the global optima. On the other hand, Bacterial Foraging Optimization algorithm, despite being a swarm based algorithm, uses a different approach imitating the behavior of a bacteria at different situations. Although each algorithm is fed with the exact same parameters, the outcomes are strikingly different. Each algorithm optimizes the problem of resonating patch in its own way and time

Training on micro- and nano- computed tomography (CT) scanners has been traditionally conducted via extensive practice on the instrument. This entails presence of an instructor to guide through the training procedure, until reasonable experience is attained. Modern tomographic instruments being expensive to maintain, the operational costs escalates with increasing number of training conducted. In a pioneering approach, the technical know-how to operate such equipment has been partly imparted via virtual reality environment running on the Second Life grid. The experimentation has indicated a reduction of the total training time. The authors hope that in the long run, such techniques will aid in significant reduction of instruction time and costs associated with training.