Computer Vision

We tackle fundamental unsupervised learning approaches that we consider to be the key to true unconstrained learning, which best simulates how humans discover the surrounding world. We aim to combine the unsupervised visual perception with supervised cognitive video representation. We want to build systems that understand our world by only watching videos. And we go further, teaching the system to describe, in natural language, the discovered elements.

Elena Burceanu, Ema Haller, Andrei Nicolicioiu, coordinated by Marius Leordeanu

Unsupervised learning of objects from video sequences

We address an essential problem in computer vision, that of unsupervised foreground object segmentation in video, where a main object of interest in a video sequence should be automatically separated from its background. Video object segmentation and object discovery are strongly related tasks, but we tackle the problem from a fully unsupervised perspective, building object representations from raw video sequences. An efficient solution to this task would enable large-scale video interpretation at a high semantic level in the absence of the costly manual labeling.

We are focused on generating foreground object soft masks based on automatic selection and learning from highly probable positive features. We show that such features can be selected efficiently by taking into consideration the spatio-temporal appearance and motion consistency of the object in the video sequence. We also emphasize the role of the contrasting properties between the foreground object and its background. Our work is also focused on theoretically proving the properties of our unsupervised learning method, which under some mild constraints is guaranteed to learn the correct classifier even in the unsupervised case.

Graph methods for video processing

Given a video we want to have a good understanding of the scene and to be able to identify the key events in order to extract useful information. Our goal is to capture the complex interactions between multiple entities in a scene. We improve the clasical convolutional models by proposing a graph model, that has a strong, explicit bias towards modeling relationships and at the same time being able to model long range interactions. We also design a novel method, devised at the object-level, that can discover salient regions and we quantitatively show that they correlate with objects locations and help the relational procesing

Multi-task graph

Existing methods in multi-task graphs rely on expensive manual supervision. In contrast, our proposed solution, with consensus shift learning, relies only on pseudo-labels provided by expert models. In our graph, every node represents a task, and every edge learns to transform one input node into another. Once initialized, the graph learns by itself on virtually any novel target domain. An adaptive selection mechanism finds consensus among multiple paths reaching a given node and establishes the pseudo-ground truth at that node. Such pseudo-labels, given by ensemble pathways in the graph, are used during the next learning iteration when single edges distill this distributed knowledge.

Video Captioning

Video captioning is the task of describing a video in natural language. It lies at the intersection of computer vision, natural language processing and machine learning requiring both high level visual comprehension and the ability to produce meaningful sentences. Our goal is to detect objects and events in a video and be capable of understanding the interactions between them in spatial and temporal dimensions.

We investigated multiple methods to analyze a video and extract information from it. To overcome limitations of the task and the available data, we design multiple models, to explore different video encoding strategies, to explore intermediate video-language representation and to investigate the gains brought by additional tasks and features. We propose a method for video captioning by selecting from the results of multiple encoder-decoder models. Our selection method based on consensus among multiple sentences is more likely to produce results with the same meaning as the video. We designed a methods that surpassed the state-of-the-art results on the challenging MSR-VTT dataset.

Unsupervised Object Tracking

Object tracking is one of the first and most fundamental problems that has been addressed in computer vision. While it has attracted the interest of many researchers over several decades of computer vision, it is far from being solved.

The task is hard for many reasons. Difficulties could come from severe changes in object appearance, presence of background clutter and occlusions that might take place in the video.

The only ground-truth knowledge given to the tracker is the bounding box of the object in the first frame. Thus, without knowing in advance the properties of the object being tracked, the tracking algorithm must learn them on the fly. It must adapt correctly and make sure it does not jump toward other objects in the background. That is why the possibility of drifting to the background poses on of the main challenges in tracking.

Natural Language Processing

A large amount of today's data is stored in databases. Building AI tools that facilitate the access to knowledge requires processing of natural language and structured data. We focus on neural approaches for natural language interfaces to databases, in particular structure-aware and semi-supervised methods.

Florin Brad in collaboration with Traian Rebedea, Ionel Hosu, Radu Iacob

Natural Language Interface to Databases

Natural Language Interface to Databases (NLIDB) bridges the gap between technical and non-technical users by allowing the latter to query large amounts of structured data through the use of instructions written in natural language.

Despite long-standing research efforts, progress has been slow and widespread adoption has failed to pick up. Data-driven approaches have been hindered by the lack of large parallel corpora to train the models on, but recent datasets alleviate this problem. We seek to improve existing SEQ2SEQ models by leveraging syntax to guide the generation process and by using semi-supervised techniques to overcome the low parallel data regime.

Anomaly detection in Text

Recent deep methods for detecting anomalies in images learn better features of normality in an end-to-end self-supervised setting. These methods train a model to discriminate between different transformations applied to visual data and then use the output to compute an anomaly score. We use this approach for Anomaly Detection in text, by introducing a novel pretext task on text sequences. We learn our model end-to-end, enforcing two independent and complementary self-supervision signals, one at the token-level and one at the sequence-level.

Domain Adaptation for Authorship Verification

The task of identifying the author of a text spans several decades and was tackled using linguistics, statistics, and, more recently, machine learning. We tackle it inspired by the impressive performance gains across a broad range of natural language processing tasks and by the recent availability of the PAN large-scale authorship dataset.

Visual Question Answering

The VQA task refers to answering questions about an image. Existing methods fuse image and text representations and are able to exploit superficial correlations and produce the correct answer, but often for the wrong reason. The recently introduced Visual Commonsense Reasoning dataset facilitates cognition-level understanding on top of recognition, by requiring not only the correct answer, but also a justification for it. We improve existing methods by leveraging structured representations of images.

Reinforcement Learning

Within the field of artificial intelligence reinforcement learning seems a natural setting for training agents that interact with the world we are living in. We engage in furthering the field by developing agents able to learn continuously in different environments. We are also investigating models sitting at the intersection of generative models and reinforcement learning.

Florin Gogianu, collaborating with Tudor Berariu

Learning Representations for Deep Reinforcement Learning

We are interested in furthering the research on learning good representations for Deep Reinforcement Learning towards the goals of more stable optimisation, improvede generalization and better sample complexity.

Continual Learning

Recent advances in machine learning are still limited to stationary tasks, but general purpose intelligence would require agents able to acquire knowledge in a continual manner dealing with interleaving tasks. Lifelong learning scenarios deal exactly with this problem: training agents on new tasks while preserving performance on old ones.

We are currently exploring memory based, optimization related and architectural methods to train neural models in lifelong learning scenarios.

    Malmo AI Challenge

    The Microsoft Malmo AI Challenge proposed a time-extended stag hunt scneario build on top of the well-known Minecraft platform. In order to maximize its payoff in such a game an agent needs to predict the level of commitment to a collaborative strategy of the other player, decide on a specific plan and navigate through the environment to execute it.

    Multi-agent setups pose additional optimization problems stemming from the non-stationarity of the training objective. Also, situated environments ask agents to learn dynamic strategies capable of dealing with sudden changes in the course of action (such as another playing abandoning the collaborative strategy).

    We approached the contest by training agents through deep reinforcement learning techniques using recurrent neural networks for policies and for value estimation. We also added auxiliary loss functions (such as next reward, or next frame prediction) in order to complement the sparse reward signal.

    Our submission ranked second for the AI Summer School placement prize and third for the Microsoft Azure for Research Grant prize.


    Lattice-based cryptography is a great promise for post-quantum cryptography. We build advanced primitives whose security relies on the hardness of lattice problems.

    Miruna Rosca, Radu Țițiu, Mădălina Bolboceanu

    Hardness of lattice problems

    Lattice-based cryptography is a great promise for post-quantum cryptography. It aims at harnessing the security of cryptographic primitives in the conjectured hardness of well-identified and well-studied algorithmic problems involving euclidean lattices. In order to build post-quantum cryptographic primitives based on lattices, we actually make use of some intermediate, more versatile problems, Learning With Errors (LWE) and Shortest Integer Solutions (SIS), which are provably at least as hard as classical lattice problems.

    To obtain more efficient primitives, different structured variants of LWE and SIS have been introduced. We are interested in studying the hardness of all these problems and giving reductions between them.


    Elena Burceanu

    My interest is in understanding videos in an unsupervised manner, currently working on object tracking for my PhD at University of Bucharest and Institute of Mathematics of the Romanian Academy. I have a strong background in Mathematics and Physics and I have finished my BSc in Computer Science and the MSc in Distributed Systems at University Politehnica Bucharest.


    Emanuela Haller

    I am a second year PhD student, co-supervised by Marius Leordeanu (Institute of Mathematics of the Romanian Academy) and Adina Magda Florea (University Politehnica of Bucharest). I have a BSc in Computer Science and a MSc in Artificial Intelligence, both from University Politehnica of Bucharest. My main focus is the problem of unsupervised learning and I am currently working on the task of unsupervised learning of objects from video sequences.


    Andrei Nicolicioiu

    I am interested in deep learning methods with a focus on relational representations and graph neural networds. I studied at University Politehnica of Bucharest, where I obtained a Bachelor's degree and a Master's degree in Artificial Inteligence. I researched topics in multilabel classification, video captioning, occluded regions segmentation, few-shot learning, relational processing of visual data.


    Florin Brad

    I am interested in neural generative models for natural language processing, especially for code generation. In particular, I am interested in leveraging discrete structure (syntax trees) to improve the expresiveness of the latent space and to guide the generation process.


    Andrei Manolache

    I got my bachelor's degree in Mathematics and Computer Sciences from the University of Bucharest and I've currently enrolled in its Artificial Intelligence Master's degree program. I'm interested in unsupervised and self-supervised representation learning, especially using deep neural language models. At the moment I'm tackling the problem of out-of-distribution sample detection in natural language and more general unstructured data.


    Marius Drăgoi

    I studied at Politehnica University of Bucharest, where I obtained a Bachelor's degree in Computer Science and a Master's degree in Artificial Intelligence. I am currently researching the problem of out-of-distribution sample detection and distributional shift in log data.


    Florin Gogianu

    Currently pursuing a PhD in Reinforcement Learning unde the supervision of Prof. Lucian Bușoniu following an MSc in Artificial Intelligence from University Politehnica of Bucharest and a BSc in Philosophy. I have a broad interest in Reinforcement Learning topics and I am currently focusing on questions regarding sample efficiency in the context of model-free value-based methods with neural network estimators.


    Miruna Roșca

    I am interested in post-quantum cryptography, with a focus on lattice-based solutions. I have a strong background in mathematics and I obtained my Phd in cryptography from École Normale Supérieure de Lyon, advised by Damien Stehlé. During my Phd, I worked on algebraic variants of Learning With Errors. More info about me on the website link below.


    Mădălina Bolboceanu

    I obtained both my undergraduate and Master degrees in Mathematics from the University of Bucharest. My goal is to use my strong mathematical skills and experience in mathematical contests and olympiads to solve cryptographic challenges. I am interested in applications of lattices in cryptography, including lattice based homomorphic encryption schemes.


    Radu Țițiu

    I am interested in lattice-based cryptography, which proposes promising cryptographic schemes in the eventuality that quantum computers are built. Moreover, lattice-based cryptography enables the construction of some advanced cryptographic primitives that allow computation on encrypted data, including Functional Encryption or Homomorphic Encryption. I have obtained my PhD in Cryptography from École Normale Supérieure de Lyon, under the supervision of Benoît Libert. In my undergraduate and my Master programs I studied mathematics at the University of Bucharest..


    Marius Leordeanu

    I am an Associate Professor (Senior Lecturer) at the University Politehnica of Bucharest and senior researcher at the Institute of Mathematics of the Romanian Academy. I am interested in the nature of intelligence, life and consciousness. In particular, my research focuses on computer vision, machine learning and robotics. At the university I teach the graduate level computer vision and robotics classes. I have received a Ph.D. in Robotics from Carnegie Mellon University in 2009 and Bachelor degrees in Mathematics and Computer Science from the City University of New York, 2003.