The Barcelona Supercomputing Center - Centro Nacional de Supercomputación (BSC-CNS) is the leading supercomputing center in Spain. It houses Mare Nostrum, one of the most powerful supercomputers in Europe, was a founding and hosting member of the former European HPC infrastructure PRACE (Partnership for Advanced Computing in Europe), and is now hosting entity for Euro HPC JU, the Joint Undertaking that leads large-scale investments and HPC provision in Europe. The mission of BSC is to research, develop and manage information technologies in order to facilitate scientific progress. BSC combines HPC service provision and R&D into both computer and computational science (life, earth and engineering sciences) under one roof, and currently has over 1000 staff from 60 countries.

Look at the BSC experience:
BSC-CNS You Tube Channel
Let's stay connected with BSC Folks!

We are particularly interested for this role in the strengths and lived experiences of women and underrepresented groups to help us avoid perpetuating biases and oversights in science and IT research. In instances of equal merit, the incorporation of the under-represented sex will be favoured.

We promote Equity, Diversity and Inclusion, fostering an environment where each and every one of us is appreciated for who we are, regardless of our differences.

If you consider that you do not meet all the requirements, we encourage you to continue applying for the job offer. We value diversity of experiences and skills, and you could bring unique perspectives to our team.

The Department of Computer Applications for Science and Engineering (CASE) participates in a 5-year European project focused on advanced Language Technologies. To advance that initiative, we seek a young Researcher to embark in the development of algorithms to make learning machines ultimately capable of rapidly comparing the semantic contents of two texts in any language. This application belongs to the general area of Machine Learning-Aided Mathematical Optimization and Natural Language Processing (NLP). It focuses on how to reduce the computational load of a class of mathematical optimization problems, resorting to surrogate ML models.

Mathematical optimization, aka integer programming (IP or ILP), formulations are convex problems solved by discrete optimization, where the objective is the optimal allocation of given resources (represented by decision variables). We already formulated the pure binary algorithm that solves this type of joint set-packing and set-partitioning problem. That entails exploring the complete feasible solutions space, i.e. the space of all possible pairing combinations between the semantic constituents of A and B, a computationally costly and time consuming endeavor. Our challenge lies in diminishing that cost and time-to-solution. Rather than considering the complete feasible solutions space, we are interested in either reducing the size of the feasible solution space from the very outset of the ILP resolution, or altogether replacing the ILP resolution by an inference produced by a surrogate Machine Learning (ML) model.

Among the Natural Language Processing solutions proposed to compare the semantic contents of two texts are various modalities of similarity computations. They extend from the now superseeded, traditional n-gram based lexical comparison, to the comparison of context-aware, semantic representations, in the form of high dimensional vector embeddings, or semantic encodings, of the constituent syntagmata of texts A and B. Vector embeddings are obtained from learned transformer-based language models (LMs), fine-tuned for Natural Language Understanding (NLU).

The successful candidate will use pre-trained, NLU-optimized, distilled LMs to obtain high dimensional semantic encodings for the semantic constituents (SCs) of textual summaries. SCs are (subject, verb, object) triples, automatically extracted from arbitrary texts by custom pre-processing. In this manner comparing texts A and B becomes equivalent to comparing two sets of vector embeddings in high dimensional space. The optimized pairing of set A embeddings with set B embeddings ultimately yield a bipartite graph where each vertex degree is at most 1 and each edge has the similarity of the paired USCs as attribute.

The successful hiree will focus their research & development on:
(i) establishing a containerized HPC-capable Python workflow, applying a free and open-source ILP solver to our joint set-packing and set-partitioning problem, and following the already formulated ILP constraints, thus establishing a baseline for computation.
(ii) establishing a Reinforced Learning (RL) formulation and/or a Graph Neural Network (GNN) design to train as surrogate models, using the sophisticated Marenostrum 5 research infrastructure and its general purpose (CPU-based) and accelerated (GPU-based) partitions.
(iii) setting up a complete performance evaluation framework and obtaining suitable perrformance metrics.
(iv) developping an easy to deploy containerized Python module, exposing end-points via a REST API, capable of running on GPU, or on multi-CPU hosts.

Successful candidates will benefit from expert training and BSC-CNS staff benefits: international multidisciplinary scientific environment and advanced applied research training. We encourage applications from highly motivated candidates with demonstrated experience and interest in applied research, within the context of major European Projects.