Currently listed open PhD positions have the following timeframe:
Research Supervisor | Assistant Professor Maxim Panov () |
Research Group | Statistical Machine Learning | https://sites.skoltech.ru/statml/ |
Project Description | The objective of this project is to develop methods for estimating the uncertainty of predictions of machine learning algorithms and to develop experiment planning procedures and Bayesian optimization based on them.
In many areas of modern science and technology, in particular in industrial design, modeling of chemical compounds and materials, industrial production and economic process control systems, and many others, it is necessary to construct models of the systems under consideration which make it possible to draw conclusions about their properties and predict their behavior. Moreover, it is important to not only create the models but also to estimate uncertainty for them and to optimize these systems. In all the problems described above, the possibility to estimate the uncertainty of the model predictions plays a major role. For the prediction problem, an uncertainty estimate is necessary to detect those points at which the prediction of the model is unreliable. In the problem of detecting anomalies, the estimation of uncertainty is necessary in order to distinguish between the deviations associated with a high dispersion of the prediction and the anomalous deviations of the average value of the forecast. In the problem of optimizing functions, the calculation of which is expensive, the estimation of uncertainty allows balancing between the search for the optimum in the vicinity of the best-known point and a broader study of the search space. Finally, in the task of designing an experiment, points with high forecast uncertainty are good candidates for addition to the training set. Note that not all machine learning models have built-in capabilities for assessing the uncertainty of their forecast. The most popular model with a “built-in” uncertainty estimate is a regression model based on Gaussian processes, in which the posterior dispersion of the Gaussian process acts as an estimate of uncertainty. Currently, there are a large number of optimization procedures, adaptive design of an experiment, detection of anomalies, and others based on this model. However, a significant drawback of this model is its high computational complexity, which limits its applicability to problems with large amounts of data. Also, this model was initially limited by the need to specify a parametric family of covariance functions, the choice of which in practice is complicated since one needs to explicitly take into account the features of the problem being solved. Predictive models based on modern neural networks have significantly less computational complexity at the training stage, and also make it relatively easy to incorporate into the model various features of the problem being solved (for example, translational invariance in the image recognition problem). However, methods for estimating the uncertainty for neural networks have begun to be actively developed only in recent years and are still at a relatively early stage of development, without giving sufficiently accurate estimates of the prediction uncertainty in many applications. In this project, we plan to develop a methodology for estimating uncertainty for neural networks and apply it to the development of specific active learning and adaptive design of experiments procedures, which in turn will be used to solve a number of important applied problems. This project focuses on solving the following key tasks:
|
Eligibility criteria | Candidates should hold or expect an MSc (or an equivalent) degree in the following or closely related disciplines:
Other eligibility criteria:
|
Research Supervisor | Assistant Professor Dmitry Yudin () |
Research Group | Dmitry Yudin |
Project Description | Nowadays neuromorphic computing is considered one of the most promising approaches for resolving the critical problems which the conventional CMOS technology faces upon continual miniaturization and ever-increasing power consumption. Owing to the low-power performance and brain-inspired massively parallel computing principles, a large number of bio-inspired algorithms and devices have been attempted in complex pattern recognition, image processing, and data mining. Intensive research has been conducted towards developing learning-based artificial synapses and neurons, attempting to reproduce the behavior of these two fundamental building blocks in biological neural networks. Owing to its similarity of ion movement in the device, a memristor has been proposed as a suitable artificial synapse and neuron for this emerging neuromorphic computing applications. In the meantime, to precisely emulate the high dynamic ranges of the neural activity to train a neural network in memristor arrays, changes in weight values in the form of memristive conductance should be subtle and uniform.
In this project, we aim to design memristors with alloying conducting channels for neuromorphic realizations of on-chip deep learning. With special attention paid to the silicon switching medium, we investigate how these memristors may be tailored to provide the element base for novel concept devices to secure stable and controllable device operation, thus enabling large-scale implementation of crossbar arrays. In a purposely designed alloying ratio, substantial improvement in spatial/temporal switching uniformity, stable data retention over the large-conductance range, and significantly enhanced program symmetry in analog conductance states can be achieved. With such an alloyed memristor one is capable of fabricating large-scale crossbar arrays featuring high device yield and accurate analog programming capability, thus paving the way to beyond von Neumann computing. |
Eligibility criteria | Candidates should hold or expect an MSc (or an equivalent) degree in the following or closely related disciplines:
Other eligibility criteria:
|
Research Supervisor | Assistant Professor Vladimir Palyulin () |
Research Group | Vladimir Palyulin |
Project Description | Random processes produce random trajectories. In order to understand the nature of stochastic phenomena behind a wide range of phenomena from motions of proteins in a living cell to human behavioral strategies, one has to infer a model from the trajectory information. Recent advances in machine learning offer an opportunity to create tools for trajectory characterization. These tools can be noticeably more efficient than the conventionally used mean-squared displacement approach.
The project consists of the development of new ML-based frameworks for trajectory characterization, segmentation, and inpainting. The latter problem appears when only the parts of the trajectory are known from the measurements. |
Eligibility criteria | Candidates should hold or expect an MSc (or an equivalent) degree in the following or closely related disciplines:
Other eligibility criteria:
|
Research Supervisor | Assistant Professor Natalia Strushkevich () |
Research Group | Natalia Strushkevich |
Project Description | Mycobacterium tuberculosis (Mtb) is among the top 10 causes of death worldwide; and the number of drug-resistant species is increasing. The major challenge for anti-tuberculosis drug discovery is the lack of knowledge about the function of many mycobacterial proteins [PMID: 9634230]. Among them are enzymes of the cytochrome P450 (CYP) family, 75% of which do not have an assigned function. Various experimental approaches are used to deorphanize this class of proteins [PMID: 20493973], however, most of them have limitations and cannot be applied to the enzymes that utilize undiscovered molecules. Significant amounts of information accumulated over years of study of human and pathogenic proteins in transcriptomic, metabolomics, structural biology, and other fields could be used for machine learning to predict the protein function.
The goal of the project is the application of machine learning to deduce consistent patterns that define the function of unknown CYPs. Predictive algorithms and/or models will be developed with further docking of predicted or de novo designed compounds into the active site of CYPs with known 3D structure. Experimental validation of results is planned within collaborations with different labs. |
Eligibility criteria | Candidates should hold or expect an MSc (or an equivalent) degree in the following or closely related disciplines:
Other eligibility criteria:
|
Research Supervisor | Visiting Professor Christoph Hermann Borchers () |
Research Group | Laboratory of Omics Technology and Big Data for Personalized Medicine and Health | Christoph Hermann Borchers |
Project Description | The Laboratory for Omics Technologies and Big Data for Personalized Medicine and Health is looking for PhD students who will be part of the recently awarded Megagrant to Dr. Borchers.The R&D of the laboratory is focused on absolute quantitation of the protein expression, modification status, and mutation rate in biological and clinical samples. One emphasis of the laboratory is on the development and implementation of the “Next Generation Proteomics”, making the protein analysis more sensitive, multiplex, and faster allowing for more applicable and comprehensive analysis of highly complex systems like the proteome of tumors and blood plasma proteome. The students will be educated and trained to use state-of-the-art laboratory equipment, analytical instruments like modern mass spectrometers, and computational software including deep learning and AI.
The students will work on technology development projects directly applied to answer biological and medical questions. The students have the opportunity to visit and further enhance their studies at the proteomics laboratory of Dr. Borchers at the McGill University in Montreal, Canada. |
Eligibility criteria | Candidates should hold or expect an MSc (or an equivalent) degree in the following or closely related disciplines:
Other eligibility criteria:
|
Research Supervisor | Assistant Professor Pavel Osinenko () |
Research Group | Artificial Intelligence in Dynamic Action | https://sites.skoltech.ru/aida |
Project Description | When machine learning comes into the world of dynamical environments, it faces a qualitatively new set of challenges. Think of a task of face generation like in StyleGAN, or photo enhancement – these tasks are performed in an inherently safe environment: there is nothing that threatens anyone during the learning phase.
Reinforcement learning, which is considered by some as the vanguard of ML, is supposed to imitate the behavior of living beings by an adaptation to the environment via trial-and-error. Living beings, however, do not blindly try out actions — they reason, although, depending on species, maybe in a somewhat primitive way. In most reinforcement learning applications, we rather deal with more or less safe environments where we can first train the neural nets, e.g., actor, critic. model estimator, with sufficient data, and then run them on-field. The cost of a failure during the learning phase is relatively low in such scenarios. Consider AlphaZero, the famous chess bot, based on reinforcement learning. Evidently, a loss in a blitz game is just yet another entry in the data set. Open AI Gym, a popular Python tool of reinforcement learning, learns the agents by trial-and-error also. The challenge of this project is to develop a reinforcement learning approach free of the necessity of a long training phase and capable of more or less “hot start” like traditional control. Here is a question to think about: can we develop a hybrid of reinforcement learning and model-based, predictive engines to possess the intelligence of machine learning and strict guarantees like traditional control? For the time being, reinforcement learning is barely applicable in industry precisely due to the limitations described. If we want to roll out a really commercializable algorithm – we need to address the above challenge. We may take the general reinforcement learning setup and investigate how it can be augmented so as to become fully online-capable. There are several ways of tackling this. A particular one is to build reinforcement learning on top of some nominal controller that guarantees safety by default and launch the learning in a virtual environment where such a controller is ambient. Another way is to balance data-driven elements of reinforcement learning with predictive controls. |
Eligibility criteria | Candidates should hold or expect an MSc (or an equivalent) degree in the following or closely related disciplines:
Other eligibility criteria:
|
Research Supervisor | Assistant Professor Alexander Panchenko () |
Research Group | Natural Language Processing | https://sites.skoltech.ru/nlp |
Project Description | The aim of the research project is to create new methods for active learning (AL) that are applicable in the tasks of information extraction from natural language texts using deep neural network models.
Most of the research and development in natural language processing is based on machine learning methods, which require a large amount of annotated resources (training corpora). Both industry and scientific groups regularly encounter a shortage of such resources when conducting their research and development. Annotation of text corpora is a very time-consuming process, so only a few languages have a substantial amount of resources suitable for machine learning. For most other languages, including Russian, the lack of resources is a very acute problem. The situation is even more severe if it is necessary to analyze texts from specific subject areas, for example, from clinical medicine or biomedicine. In such cases, corpus annotation is very expensive, because instead of simple crowdsourcing, it is necessary to attract highly qualified specialists (physicians and scientists). In this case, annotation with the help of active training can significantly reduce costs and accelerate the development of text analysis tools.Active learning iteratively selects unlabeled objects and shows them to a human annotator. It shows only objects that are considered “informative”. This helps to reduce annotation redundancy and avoid unimportant objects, for example, outliers. One of the essential components of active learning methods is query strategies of unlabeled training examples that are shown to human annotators. Applying active learning commonly reduces the amount of annotation by several times but retains the high quality of the trained model. In the near future, this technology is expected to become widely used in both industrial and research projects related to machine learning, and, in particular, in projects related to natural language processing.The state-of-the-art methods for natural language processing are based on deep neural models with a large number of parameters that take advantage of transfer learning: ELMo, BERT, RoBERTa, ELECTRA, inter alia. The project aims is to develop methods of active learning that are in alignment with the abilities of these recently proposed models. The research can be conducted along with one of the two main directions:
|
Eligibility criteria | Candidates should hold or expect an MSc (or an equivalent) degree in the following or closely related disciplines:
Other eligibility criteria:
|
Research Supervisor | Assistant Professor Alexander Panchenko () |
Research Group | Natural Language Processing | https://sites.skoltech.ru/nlp |
Project Description | The output of a neural generation model, such as RNN or Transformer, is difficult to control. This happens because such models are trained on large volumes of text data which usually come from the web. Such texts exist in large amounts and are constantly renewed, using them allows training models on up-to-date topics and vocabulary. However, such an abundance of texts cannot be checked for quality. User-generated content can include obscene lexis, toxicity, and other undesirable elements that are difficult to detect automatically.As a result, a model trained on data from the web can potentially generate an undesirable text. Since the preemptive check of text appropriateness is impossible, using neural models in commercial applications is too risky. This often makes developers of generation models resort to less flexible, but more reliable and controllable models, such as rule-based dialogue systems based on scenarios.
The goal of this work is to increase the reliability of neural dialogue models by creating an external automatic censor that will evaluate a generated utterance and decide if it is acceptable for a dialogue. The censor will consist of two components: a model that evaluates the quality of generated text and a model that re-formulates the unacceptable text. The Quality evaluation model will define if a text contains inappropriate elements, namely obscene phrases, offense, too informal phrases, undesirable topics. It will also need to model the dialogue flow because a neutral phrase can become offensive in some contexts. The re-formulation model should re-write an unacceptable utterance in such a way that it keeps its content but the tone becomes acceptable.There currently exist no systems that could identify and correct unacceptable utterances on the fly. The existing automatic censors for chatbots usually rely on lists of potentially dangerous keywords. This approach can overlook offenses which do not contain obscene words. We would like to create a model that considers features of a higher level. Moreover, existing models do not suggest inoffensive alternatives for the discovered offensive messages. While text paraphrasing and style transfer of texts is an active topic of research, the task of transforming toxic text to neutral has only been considered in a few research works. Thus, the work will include collecting new datasets and devising new methods that take into account the specificity of the task. |
Eligibility criteria | Candidates should hold or expect an MSc (or an equivalent) degree in the following or closely related disciplines:
Other eligibility criteria:
|
Research Supervisor | Associate Professor Sergey Rykovanov () |
Research Group | High-Performance Computing Research and Education (Hipercore) | http://sites.skoltech.ru/hipercore/ |
Project Description | High-Performance Computing became a necessity in modern science due to the highly nonlinear nature of the most interesting phenomena in chemistry, engineering, materials science, physics, and astrophysics. Nobel Prize in Physics in 2018 was given for the development of extremely powerful and short laser pulses. Such pulses ionize materials almost instantly and cause extreme, nonlinear, relativistic phenomena similar to that in Space. There are several important areas of extreme laser physics:
Of course, all these areas are impossible to study without the usage of powerful supercomputers. There are several PhD positions within this project depending on the interest and skills of the applicant.
Students, of course, will have the opportunity to interact with the research group and tackle parts of different projects. Projects will be conducted on one of the most powerful supercomputers in Russia “Zhores” boasting 1 petaFLOPS theoretical performance in the GPU partition. Successful students have the opportunity for academic mobility and collaboration with our national and international partners in the UK, Germany, China, and the USA. |
Eligibility criteria | Candidates should hold or expect an MSc (or an equivalent) degree in the following or closely related disciplines:
Other eligibility criteria:
|
Research Supervisor | Associate Professor Sergey Rykovanov () |
Research Group | High-Performance Computing Research and Education (Hipercore) | http://sites.skoltech.ru/hipercore/ |
Project Description | Scientific Simulations of complex systems (plasmas, atmosphere, pedestrian flows, molecules, etc) are often extremely expensive in terms of computational time prohibiting large parameter scans. On the other hand, the training of Deep Learning networks typically requires a large dataset to succeed.The goal of this project is to apply and improve the recently proposed DENSE method that allows using deep neural networks even with small datasets. The successful candidate will apply this method to various simulations of complex systems such as (but not limited to) complex fluids, plasmas, nonlinear nanophotonics systems, fluid flows in micro- and porous structures.This is an extremely interesting, challenging, and promising project uniting traditional mathematical modeling and AI methods. |
Eligibility criteria | Candidates should hold or expect an MSc (or an equivalent) degree in the following or closely related disciplines:
Other eligibility criteria:
|
For more details, please get acquainted with the CDSE PhD program details and PhD admission procedure.
Skoltech Center CDISE aims at the conduction of cross-cutting interdisciplinary research driven by modern applications in the fields of computational and data sciences. CDISE has accumulated 20+ research groups all with unique expertise in Russia and competitive at the world level in their prospective areas:
All PhD studentships in Skoltech are by default covered with a generous stipend of 75,000₽ in addition to health coverage for the duration of graduate studies, provided that academic standards are met, and progress towards a PhD degree is maintained. Check for more details in Skoltech policy on student stipends and other benefits.