ERC FUNDED PROJECTS

The explosion of information around us has democratized knowledge and transformed its availability for people around the world. Still, since information is mediated through automated systems, access is bounded by their ability to understand language. Consider an economist asking “What fraction of the top-5 growing countries last year raised their co2 emission?”. While the required information is available, answering such complex questions automatically is not possible. Current question answering systems can answer simple questions in broad domains, or complex questions in narrow domains. However, broad and complex questions are beyond the reach of state-of-the-art. This is because systems are unable to decompose questions into their parts, and find the relevant information in multiple sources. Further, as answering such questions is hard for people, collecting large datasets to train such models is prohibitive. In this proposal I ask: Can computers answer broad and complex questions that require reasoning over multiple modalities? I argue that by synthesizing the advantages of symbolic and distributed representations the answer will be “yes”. My thesis is that symbolic representations are suitable for meaning composition, as they provide interpretability, coverage, and modularity. Complementarily, distributed representations (learned by neural nets) excel at capturing the fuzziness of language. I propose a framework where complex questions are symbolically decomposed into sub-questions, each is answered with a neural network, and the final answer is computed from all gathered information. This research tackles foundational questions in language understanding. What is the right representation for reasoning in language? Can models learn to perform complex actions in the face of paucity of data? Moreover, my research, if successful, will transform how we interact with machines, and define a role for them as research assistants in science, education, and our daily life.

1 499 375 €

Start date: 2019-04-01, End date: 2024-03-31

Understanding how the Milky Way arrived at its present state requires a large volume of precision measurements of our Galaxy’s current makeup, as well as an empirically based understanding of the main processes involved in the Galaxy’s evolution. Such data are now about to arrive in the flood of quality information from Gaia and SDSS-V. The demography of the stars and of the compact stellar remnants in our Galaxy, in terms of phase-space location, mass, age, metallicity, and multiplicity are data products that will come directly from these surveys. I propose to integrate this information into a comprehensive picture of the Milky Way’s present state. In parallel, I will build a Galactic chemical evolution model, with input parameters that are as empirically based as possible, that will reproduce and explain the observations. To get those input parameters, I will measure the rates of supernovae (SNe) in nearby galaxies (using data from past and ongoing surveys) and in high-redshift proto-clusters (by conducting a SN search with JWST), to bring into sharp focus the element yields of SNe and the distribution of delay times (the DTD) between star formation and SN explosion. These empirically determined SN metal-production parameters will be used to find the observationally based reconstruction of the Galaxy’s stellar formation history and chemical evolution that reproduces the observed present-day Milky Way stellar population. The population census of stellar multiplicity with Gaia+SDSS-V, and particularly of short-orbit compact-object binaries, will hark back to the rates and the element yields of the various types of SNe, revealing the connections between various progenitor systems, their explosions, and their rates. The plan, while ambitious, is feasible, thanks to the data from these truly game-changing observational projects. My team will perform all steps of the analysis and will combine the results to obtain the clearest picture of how our Galaxy came to be.

1 859 375 €

Start date: 2019-10-01, End date: 2024-09-30

Supercomputers are strategically crucial for facilitating advances in science and technology: in climate change research, accelerated genome sequencing towards cancer treatments, cutting edge physics, devising engineering innovative solutions, and many other compute intensive problems. However, the future of super-computing depends on our ability to cope with the ever increasing rate of faults (bit flips and component failure), resulting from the steadily increasing machine size and decreasing operating voltage. Indeed, hardware trends predict at least two faults per minute for next generation (exascale) supercomputers. The challenge of ascertaining fault tolerance for high-performance computing is not new, and has been the focus of extensive research for over two decades. However, most solutions are either (i) general purpose, requiring little to no algorithmic effort, but severely degrading performance (e.g., checkpoint-restart), or (ii) tailored to specific applications and very efficient, but requiring high expertise and significantly increasing programmers' workload. We seek the best of both worlds: high performance and general purpose fault resilience. Efficient general purpose solutions (e.g., via error correcting codes) have revolutionized memory and communication devices over two decades ago, enabling programmers to effectively disregard the very likely memory and communication errors. The time has come for a similar paradigm shift in the computing regimen. I argue that exciting recent advances in error correcting codes, and in short probabilistically checkable proofs, make this goal feasible. Success along these lines will eliminate the bottleneck of required fault-tolerance expertise, and open exascale computing to all algorithm designers and programmers, for the benefit of the scientific, engineering, and industrial communities.

1 824 467 €

Start date: 2019-06-01, End date: 2024-05-31

Machine learning has rapidly evolved in the last decade, significantly improving accuracy on tasks such as image classification. Much of this success can be attributed to the re-emergence of neural nets. However, learning algorithms are still far from achieving the capabilities of human cognition. In particular, humans can rapidly organize an input stream (e.g., textual or visual) into a set of entities, and understand the complex relations between those. In this project I aim to create a general methodology for semantic interpretation of input streams. Such problems fall under the structured-prediction framework, to which I have made numerous contributions. The proposal identifies and addresses three key components required for a comprehensive and empirically effective approach to the problem. First, we consider the holistic nature of semantic interpretations, where a top-down process chooses a coherent interpretation among the vast number of options. We argue that deep-learning architectures are ideally suited for modeling such coherence scores, and propose to develop the corresponding theory and algorithms. Second, we address the complexity of the semantic representation, where a stream is mapped into a variable number of entities, each having multiple attributes and relations to other entities. We characterize the properties a model should satisfy in order to produce such interpretations, and propose novel models that achieve this. Third, we develop a theory for understanding when such models can be learned efficiently, and how well they can generalize. To achieve this, we address key questions of non-convex optimization, inductive bias and generalization. We expect these contributions to have a dramatic impact on AI systems, from machine reading of text to image analysis. More broadly, they will help bridge the gap between machine learning as an engineering field, and the study of human cognition.

1 932 500 €

Start date: 2019-02-01, End date: 2024-01-31