Papers by Luciana Morogan
Speech Emotion Recognition for Emergency Services
This paper aims to create a background for information processing support. We introduce formal cl... more This paper aims to create a background for information processing support. We introduce formal classes called object compounds. By object compounds, we refer to the formalism needed for the construction of (biological inspired equivalent) classes of molecules, compounds or complex-like objects.
Tracing One Neuron Activity from the Inside of Its Structure
International journal of new computer architectures and their applications, 2012
As new areas of neural computing are trying to make at least one step beyond the definition of di... more As new areas of neural computing are trying to make at least one step beyond the definition of digital computing, the neural networks field, was developed around the idea of creating models of real neural systems. The key point is based on learning rather than programming. We designed the model we present in this paper, based on the the needs introduced by new trends in neural computing. We created a model for information processing inside neurons. It is constructed, from a neural inside point of view, as a feedback system that controls the flow of information passing through the neuron. Processes of internal neural learning take place. We translated them as processes of learning at ”molecular” level. Our work is disseminated into the construction of an algorithm of learning, that we introduce in the end of this paper. Keywords-Neural networks; hybrid intelligent systems; nature inspired computing technique; evolutionary computing.
RoWordNet – A Python API for the Romanian WordNet
The paper presents (1) a new library (API) that provides easy access to the Romanian WordNet and ... more The paper presents (1) a new library (API) that provides easy access to the Romanian WordNet and (2) an application based on the API to perform embedded vector retrofitting. The new API (further referred to as RoWn)is built in Python and offers direct access to all the data provided in the Romanian WordNet. It implements all basic I/O and query operations. As RoWnis based on a directed graph structure powered by the versatile networkx library, users are provided with the groundwork to implement powerful graph algorithms. The second part of the paper presents an application built on RoWn: the process of vector retrofitting. “Retrofitting” refers to the semantic specialization of word vectors, a process of fine-tuning each vector under a specified constraint. We use the synonymy/antonymy relations extracted from RoWnto tune the word vectors under these semantic constraints, effectively specializing the vector space.
OpenNIG - Open Neural Image Generator
Generative models are statistical models that learn a true underlying data distribution from samp... more Generative models are statistical models that learn a true underlying data distribution from samples using unsupervised learning, aiming to generate new data points with some variation. In this paper, we introduce OpenNIG (Open Neural Image Generator), an open-source neural networks toolkit for image generation. It offers the possibility to easily train, validate and test state of the art models. The framework also contains a module that enables the user to directly download and process some of the most common databases used in deep learning. OpenNIG is freely available via GitHub.
Prion neural system: Modeling the binding affinities between neurons of a network
There is a need (that came from biological hypothesis about the functionality of certain subsyste... more There is a need (that came from biological hypothesis about the functionality of certain subsystems of the nervous system) of designing a more realistic model of a dynamic complex system. An idea of such a comlex system started with the prion neural system (already introduced in some previous papers). This article is focused on modeling synapses between neurons of a network by modeling the binding affinities between them. The original idea, of biological inspiration, is that in a network of neurons of a prion neural system (or any other network of neuron devices) any neuron should not be able to bind with any other neuron. In this article it is also designed the synaptic formation between different kinds of neurons in a network.
Generation of the path to counter-examples by backward state space traversal in symbolic model checking based on term rewriting
Model checking has proved to be a very useful formal verification technique in the design of comm... more Model checking has proved to be a very useful formal verification technique in the design of communication and security protocols. Researches have used it to validate the communications protocols from the point of view of their functional and security specifications. Over the years, model checking has evolved from explicit encoding of the state space to symbolic encoding, thus overcoming the state space explosion problem and being able to handle almost infinite state spaces. Research has indicated that another means of improving the symbolic encoding of the state space is to encode only the states of the system, but not the transition relations. But this raises the problem of how to compute the path to the counter-examples. This paper address the issue and proposes the use of a backward traversal of the state space. The idea is to start from the counter-example state and compute its predecessor(s), then take each of the predecessors and apply recursively the same inversing operation until the initial state is reached, thus obtaining the sought path. The new approach was formally defined and exemplified.
The COVID-19 pandemic has put a strain on health facilities worldwide. However, remote screening ... more The COVID-19 pandemic has put a strain on health facilities worldwide. However, remote screening can lessen the burden on medical resources. If done manually, screening does not scale to the large number of people that require it. We are constructing an automated remote screening system for Romania. The system needs to support simultaneous use by many persons (presumably a significant part of the population). Considering the urgency, we propose a lightweight web-based platform that can run on low-end server infrastructure. One auxiliary goal of the platform is for it to use machine learning (ML) to be able to adapt its screening results (recommendations) to the evolution of pandemic. We considered using cloud services, however, due to privacy considerations regarding medical data and legal issues we designed the platform to be able to run, but to not require cloud services. Some of the decisions that we made were to offload computation (ML model inference) to the browser, minimize the number of requests to the server, as well as the code size, and construct a server architecture that uses buffers for writing to the database and which can scale on demand.
The cloud computing providers need to offer security warranties. As we all know, one of the criti... more The cloud computing providers need to offer security warranties. As we all know, one of the critical points is the confidentiality and access to customer data which, these days, is migrated and managed in cloud environments. In this sense, one solution is based on encrypting data before its upload in cloud. But this approach sets a limit regarding data processing. In this article we present a practical application of the homomorphic encryption schemes, namely the problem of finding maximum/minimum from a collection of encrypted integers. First, we present our algorithm that can be run directly in cloud without the need for an intermediate data exchange with the client. Second, our experimental results show the time resources necessary to evaluate the proposed algorithm.
Efficient and robust perceptual hashing using log-polar image representation
ABSTRACT Robust image hashing seeks to transform a given input image into a shorter hashed versio... more ABSTRACT Robust image hashing seeks to transform a given input image into a shorter hashed version using a key-dependent non-invertible transform. These hashes find extensive applications in content authentication, image indexing for database search and watermarking. Modern robust hashing algorithms consist of feature extraction, a randomization stage to introduce non-invertibility, followed by quantization and binary encoding to produce a binary hash. This paper describes a novel algorithm for generating an image hash based on Log-Polar transform features. The Log-Polar transform is a part of the Fourier-Mellin transformation, often used in image recognition and registration techniques due to its invariant properties to geometric operations. First, we show that the proposed perceptual hash is resistant to content-preserving operations like compression, noise addition, moderate geometric and filtering. Second, we illustrate the discriminative capability of our hash in order to rapidly distinguish between two perceptually different images. Third, we study the security of our method for image authentication purposes. Finally, we show that the proposed hashing method can provide both excellent security and robustness.
Fast Searching in Image Databases Using Multi-index Robust Fingerprinting
Springer eBooks, 2015
Robust image fingerprinting seeks to transform a given input image into a compact binary hash usi... more Robust image fingerprinting seeks to transform a given input image into a compact binary hash using a non-invertible transform. These binary hashes exhibit robustness against common image processing and find their extensive application in multimedia databases where near neighbor index search is often employed. Unfortunately, robust fingerprinting length is usually longer than 32 bits which makes impossible to use them as direct indices in multimedia databases.
This paper aims to create a background for information processing support. We introduce formal cl... more This paper aims to create a background for information processing support. We introduce formal classes called object compounds. By object compounds, we refer to the formalism needed for the construction of (biological inspired equivalent) classes of molecules, compounds or complex- like objects.
RoWordNet – A Python API for the Romanian WordNet
2018 10th International Conference on Electronics, Computers and Artificial Intelligence (ECAI), 2018
The paper presents (1) a new library (API) that provides easy access to the Romanian WordNet and ... more The paper presents (1) a new library (API) that provides easy access to the Romanian WordNet and (2) an application based on the API to perform embedded vector retrofitting. The new API (further referred to as RoWn)is built in Python and offers direct access to all the data provided in the Romanian WordNet. It implements all basic I/O and query operations. As RoWnis based on a directed graph structure powered by the versatile networkx library, users are provided with the groundwork to implement powerful graph algorithms. The second part of the paper presents an application built on RoWn: the process of vector retrofitting. “Retrofitting” refers to the semantic specialization of word vectors, a process of fine-tuning each vector under a specified constraint. We use the synonymy/antonymy relations extracted from RoWnto tune the word vectors under these semantic constraints, effectively specializing the vector space.
Researchers considered the DNA and the cell membranes as devices for the storage of genetic mater... more Researchers considered the DNA and the cell membranes as devices for the storage of genetic material. Recently, they implemented the storage of huge quantities of information in DNA and more recently, in some abstract devices called membranes systems, known as well as P systems – called after their inventor, prof. dr. Gh. Păun. This paper presents the necessary terminology of understanding the new concepts in section 2, called “general remarks”. Those two different categories: of writing the information into DNA and of coding the information into the membrane structure using multisets of atomic objects, are presented in sections 3 and 4, respectively. In section 3 are presented the solid and the liquid faze of writing the information in the DNA sequences along with the constraints necessary for such processes. A totally different aproach is illustrated in the next chapter, because in the base model we don`t have to deal with sequences, as in DNA computing, but with multisets of atom...
The cloud computing providers need to offer security warranties. As we all know, one of the criti... more The cloud computing providers need to offer security warranties. As we all know, one of the critical points is the confidentiality and access to customer data which, these days, is migrated and managed in cloud environments. In this sense, one solution is based on encrypting data before its upload in cloud. But this approach sets a limit regarding data processing. In this article we present a practical application of the homomorphic encryption schemes, namely the problem of finding maximum/minimum from a collection of encrypted integers. First, we present our algorithm that can be run directly in cloud without the need for an intermediate data exchange with the client. Second, our experimental results show the time resources necessary to evaluate the proposed algorithm.
A scalable COVID-19 screening platform
2020 IEEE Globecom Workshops (GC Wkshps, 2020
The COVID-19 pandemic has put a strain on health facilities world-wide. However, remote screening... more The COVID-19 pandemic has put a strain on health facilities world-wide. However, remote screening can lessen the burden on medical resources. If done manually, screening does not scale to the large number of people that require it.We are constructing an automated remote screening system for Romania. The system needs to support simultaneous use by many persons (presumably a significant part of the population). Considering the urgency, we propose a lightweight web-based platform that can run on low-end server infrastructure. One auxiliary goal of the platform is for it to use machine learning (ML) to be able to adapt its screening results (recommendations) to the evolution of pandemic.We considered using cloud services, however, due to privacy considerations regarding medical data and legal issues we designed the platform to be able to run, but to not require cloud services. Some of the decisions that we made were to offload computation (ML model inference) to the browser, minimize th...
OpenNIG - Open Neural Image Generator
2020 13th International Conference on Communications (COMM), 2020
Generative models are statistical models that learn a true underlying data distribution from samp... more Generative models are statistical models that learn a true underlying data distribution from samples using unsupervised learning, aiming to generate new data points with some variation. In this paper, we introduce OpenNIG (Open Neural Image Generator), an open-source neural networks toolkit for image generation. It offers the possibility to easily train, validate and test state of the art models. The framework also contains a module that enables the user to directly download and process some of the most common databases used in deep learning. OpenNIG is freely available via GitHub.
In our days, there is a need of designing a more realistic model of a dynamic complex neural syst... more In our days, there is a need of designing a more realistic model of a dynamic complex neural system. The inspiration arise from the biological hypothesis about the functionality of certain subsystems of the nervous system. An idea, that began in some previous papers, was materialized in the construction of the so called prion neural system. The present article focuses on the way the synapses between neurons of a network can be modeled by designing the binding affinities between them. Of biological inspiration is also the original idea that governs this article: in a network of neurons any neuron should not be able to bind with any other neuron. This is the reason we present a model of synaptic formation between different kinds of neurons in a network. A detailed case of study is presented in the end of
Enhanced models in deep image steganography
Seeking to take advantage of the innovations brought by machine learning, a field in a continuous... more Seeking to take advantage of the innovations brought by machine learning, a field in a continuous movement and development, the present paper aims to enhance a practice that has been used since ancient times: steganography. Thereby, we targeted the implementation of a system aimed to hide image-type messages with the aid of deep neural networks. We followed a baseline model designed according to the recommendations stated into the state-of-the-art section. Then, we progressively developed three new models, each adding a new improvement on top of the previous one.
Recent advances in NLP have been sustained by the availability of large amounts of 1 data and sta... more Recent advances in NLP have been sustained by the availability of large amounts of 1 data and standardized benchmarks, which are not available for many languages. As 2 a small step towards addressing this, we propose LiRo, a platform for benchmarking 3 models on the Romanian language on nine standard tasks: text classification, 4 named entity recognition, machine translation, sentiment analysis, POS tagging, 5 dependency parsing, language modelling, question-answering, and semantic textual 6 similarity. We also include a less standard task of Romanian embeddings debiasing, 7 to address the growing concerns related to gender bias in language models. The 8 platform exposes per-task leaderboards populated with baseline results for each 9 task. In addition, we create three new datasets: one from Romanian Wikipedia 10 and two by translating the Semantic Textual Similarity (STS) benchmark and 11 the Cross-lingual Question Answering Dataset (XQuAD) into Romanian. We 12 believe LiRo will no...
Uploads
Papers by Luciana Morogan