Spam can be defined as unsolicited bulk email. In an effort to evade text-based spam filters, spammers can embed their spam text in an image, which is referred to as image spam. In this research, we consider the probl...
详细信息
Spam can be defined as unsolicited bulk email. In an effort to evade text-based spam filters, spammers can embed their spam text in an image, which is referred to as image spam. In this research, we consider the problem of image spam detection, based on image analysis. We apply various machine learning and deep learning techniques to real-world image spam datasets, and to a challenge image spam-like dataset. We obtain results comparable to previous work for the real-world datasets, while our deep learning approach yields the best results to date for the challenge dataset.
Sketching has been used by humans to visualize and narrate the aesthetics of the world for a long time. With the onset of touch devices and augmented technologies, it has attracted more and more attention in recent ye...
详细信息
Sketching has been used by humans to visualize and narrate the aesthetics of the world for a long time. With the onset of touch devices and augmented technologies, it has attracted more and more attention in recent years. Recognition of free-hand sketches is an extremely cumbersome and challenging task due to its abstract qualities and lack of visual cues. Most of the previous work has been done to identify objects in real pictorial images using neural networks instead of a more abstract depiction of the same objects in sketch. This research aims at comparing the performance of different machine learning algorithms and their learned inner representations. This research studies some of the famous machine learning models in classifying sketch images. It also does a study of legacy and the new datasets to classify a new sketch through various classifiers like support vector machines and the use of deep neural networks. It achieved remarkable results but still lacking behind the accuracy in the classification of the sketch images.
Classical methods for psychometric function estimation either require excessive resources to perform, as in the method of constants, or produce only a low resolution approximation of the target psychometric function, ...
详细信息
Classical methods for psychometric function estimation either require excessive resources to perform, as in the method of constants, or produce only a low resolution approximation of the target psychometric function, as in adaptive staircase or up-down procedures. This thesis makes two primary contributions to the estimation of the audiogram, a clinically relevant psychometric function estimated by querying a patientâs for audibility of a collection of tones. First, it covers the implementation of a Gaussian process model for learning an audiogram using another audiogram as a prior belief to speed up the learning procedure. Second, it implements a use case of Bayesian active model selection to determine whether two audiograms differ. Both algorithms were tested using audiometric data from the National Institute for Occupational Safety and Health (NIOSH).
In a cloud computing environment, enterprises have the flexibility to request resources according to their application demands. This elastic feature of cloud computing makes it an attractive option for enterprises to ...
详细信息
In a cloud computing environment, enterprises have the flexibility to request resources according to their application demands. This elastic feature of cloud computing makes it an attractive option for enterprises to host their applications on the cloud. Cloud providers usually exploit this elasticity by auto-scaling the application resources for quality assurance. However, there is a setup-time delay that may take minutes between the demand for a new resource and it being prepared for utilization. This causes the static resource provisioning techniques, which request allocation of a new resource only when the application breaches a specific threshold, to be slow and inefficient for the resource allocation task. To overcome this limitation, it is important to foresee the upcoming resource demand for an application before it becomes overloaded and trigger resource allocation in advance to allow setup time for the newly allocated resource. Machine learning techniques like time-series forecasting can be leveraged to provide promising results for dynamic resource allocation. In this research project, I developed a predictive analysis model for dynamic resource provisioning for cloud infrastructure. The researched solution demonstrates that it can predict the upcoming workload for various cloud infrastructure metrics upto 4 hours in future to allow allocation of virtual machines in advance.
In our current political climate, state level legislators have become increasingly impor- tant. Due to cuts in funding and growing focus at the national level, public oversight for these legislators has drastically de...
详细信息
In our current political climate, state level legislators have become increasingly impor- tant. Due to cuts in funding and growing focus at the national level, public oversight for these legislators has drastically decreased. This makes it difficult for citizens and activists to understand the relationships and commonalities between legislators. This thesis provides three contributions to address this issue. First, we created a data set containing over 1200 features focused on a legislatorâs activity on bills. Second, we created embeddings that represented a legislatorâs level of activity and engagement for a given bill using a custom model called Democracy2Vec. Third, we provided a case study focused on the 2015-2016 California State Legislator and had our results verified by a political expert. Our results show that our embeddings can explain relationships between legislator and how they will likely act during the legislative process.
The chief purpose of this study is to detect and eliminate the sentiment bias in a search engine. Sentiment bias means a bias induced in the search results based on the sentiment of the userâs search query. As ...
详细信息
The chief purpose of this study is to detect and eliminate the sentiment bias in a search engine. Sentiment bias means a bias induced in the search results based on the sentiment of the userâs search query. As people increasing depend on search engines for information, it is important to understand the quality of results produced by the search engines. This study does not try to build a search engine but leverage the existing search engines to provide better results to the user. In this study, only the queries that have high sentiment polarity are analyzed and the machine learning models are used to predict the sentiment polarity of the input query, sentiment polarity of the documents produced by the search engine for the given query and also to change the sentiment polarity of the input query to its opposite sentiment. This project proposes an end-to-end system that eliminates the search engine bias by producing results that align with the query sentiment as well as the opposite sentiment. The system comprising of three models for document level sentiment analysis, aspect level sentiment analysis and sentiment style transfer. The document level sentiment analyzer is an LSTM based model that uses GloVe word embeddings to analyze the sentiment of the documents produced by the search engine. The aspect level sentiment analyzer uses deep memory network with attention and auxiliary memory to analyze the sentiment of each search query. In order to obtain the iv documents of the opposite polarity, the sentiment of the search query is reversed using the sentiment style transfer model that uses a bi-directional LSTM. The results are analyzed to determine the sentiment bias of the search engine based on the input query. In our experiments, we observed that positive sentiment queries yielded 67% documents with positive sentiment and negative sentiment queries yielded 70% documents with negative sentiment. The proposed system eliminates this bias by providing the users with
With the progression of the internet and social media, people are given multiple platforms to share their thoughts and opinions about various subject matters freely. However, this freedom of speech is misused to direc...
详细信息
With the progression of the internet and social media, people are given multiple platforms to share their thoughts and opinions about various subject matters freely. However, this freedom of speech is misused to direct hate towards individuals or group of people due to their race, religion, gender etc. The rise of hate speech has led to conflicts and cases of cyber bullying, causing many organizations to look for optimal solutions to solve this problem. Developments in the field of machine learning and deep learning have piqued the interest of researchers, leading them to research and implement solutions to solve the problem of hate speech. Currently, machine learning techniques are applied to textual data to detect hate speech. With the ample use of video sharing sites, there is a need to find a way to detect hate speech in videos. This project deals with classification of videos into normal or hateful categories based on the spoken content of the videos. The video dataset is built using a crawler to search and download videos based on offensive words that are specified as keywords. The audio is extracted from the videos and is converted into textual format using a speech-to-text converter to obtain a transcript of the videos. Experiments are conducted by training four models with three different feature sets extracted from the dataset. The models are evaluated by computing the specified evaluation metrics. The evaluated metrics indicate that random forest classifier model delivers the best results in classifying videos.
A graph is a very powerful abstract data type that can be used to model entities (nodes) and relationships (edges). Many real world networks like biological, computer and friendship networks can be represented as grap...
详细信息
A graph is a very powerful abstract data type that can be used to model entities (nodes) and relationships (edges). Many real world networks like biological, computer and friendship networks can be represented as graphs. Graphs can be mined to extract interesting patterns and interactions between the participating entities. Recently, various artificialintelligence (AI) and Machine Learning (ML) techniques are used for this purpose. In order to do that, the nodes of a graph have to be represented as low dimensional feature vectors. Node embedding is the process of generating a �-dimensional feature vector corresponding to each node of a graph, such that the structurally similar nodes remain close in the �-dimensional space. There are many state-of-the-art methods, like node2vec and DeepWalk to com- pute node embeddings. These techniques borrow methods like the Skip-Gram model, used in the domain of Natural Language Processing (NLP) to compute word embed- dings. This project explores the idea of porting the GloVe (Global Vectors for Word Representation) model, a popular technique for word embeddings, to a new method called GloVeNoR to compute node embeddings in a graph. We evaluate the modelâs quality by comparing it with node2vec and DeepWalk on the problem of community detection on five different data sets. We observe that GloVeNoR discovers similar or better communities than the other existing models on all the datasets.
The emergence of large datasets and major improvements in Deep Learning has lead to many real-world applications. These applications have been focused on automotive markets, mobile markets, stock markets, and the heal...
详细信息
The emergence of large datasets and major improvements in Deep Learning has lead to many real-world applications. These applications have been focused on automotive markets, mobile markets, stock markets, and the healthcare market. Although Deep Learning has strong foundations across many areas, the few applications in Sports, Fitness, or even Injury Rehabilitation could benefit greatly from it. For example, if you are performing a workout and you need to evaluate your form, but do not have access or resources for an instructor to evaluate your form, it would be great to have an artificial Intelligent agent provide real time feedback through your laptop or phone. Therefore our goal in this research study is to find a foundation for an exercise feedback application by comparing two computer vision models. The two approaches we will be comparing will be pose estimation and action recognition. The latter will be covered in more depth, as we will provide an end to end approach, while the former will be used as a benchmark to compare with. Action recognition will cover the collection, labeling, and organization of the data, training and integrating with real-time data to provide the user with feedback. The exercises we will focus on during our testing and analysis will be squats and push-ups. We were able to achieve an accuracy score of 79% with our best model, given a validation set of 391 squatting images from the PennAction dataset for squat exercise action recognition.
Image compression is a well-studied field of Computer Vision. Recently, many neural network based architectures have been proposed for image compression as well as enhancement. These networks are also put to use by fr...
详细信息
Image compression is a well-studied field of Computer Vision. Recently, many neural network based architectures have been proposed for image compression as well as enhancement. These networks are also put to use by frameworks such as end-to-end image compression. In this project, we have explored the improvements that can be made over this framework to achieve better benchmarks in compressing images. Generative Adversarial Networks are used to generate new fake images which are very similar to original images. Single Image Super-Resolution Generative Adversarial Networks (SI-SRGAN) can be employed to improve image quality. Our proposed architecture can be divided into four parts : image compression module, arithmetic encoder, arithmetic decoder, image reconstruction module. This ar- chitecture is evaluated based on compression rate and the closeness of the reconstructed image to the original image. Structural similarity metrics and peak signal to noise ratio are used to evaluate the image quality. We have also measured the net reduction in file size after compression and compared it with other lossy image compression techniques. We have achieved better results in terms of these metrics compared to legacy and newly proposed image compression algorithms. In particular, an average PSNR of 28.48 and SSIM value of 0.86 is achieved as compared to 28.45 PSNR and 0.81 SSIM value in end to end image compression framework [1]
暂无评论