ERC FUNDED PROJECTS

"Secure two-party and multiparty computation has long stood at the center of the foundations of theoretical cryptography. However, in the last five years there has been blistering progress on the question of efficient secure computation. We are close to the stage that secure computation can be applied to real-world privacy and security problems. There is thus considerable interest in secure computation solutions from governments, military and security organisations, and industry. However, in order to answer the needs of secure computation in practice, there is still a need to make secure computation protocols much faster. Until now, research in efficient cryptographic protocols has typically been in two different directions. The first direction, and the major one, is to construct more efficient protocols and prove them secure, where efficiency is measured by the amount of communication sent, the number of heavy cryptographic operations carried out (e.g., exponentiations), and so on. The second direction is to take the state-of-the-art protocols and implement them while optimising the implementation based on systems concerns. This latter direction has proven to improve the efficiency of existing protocols significantly, but is limited since it remains within the constraints of existing cryptographic approaches. We propose a synergetic approach towards achieving high-performance secure computation. We will design new protocols while combining research from cryptography, algorithms and systems. In this way, issues like load balancing, memory management, cache-awareness, bandwidth bottlenecks, utilisation of parallel computing resources, and more, will be built into the cryptographic protocol and not considered merely as an afterthought. If successful, HIPS will enable the application of the beautiful theory of secure computation to the problems of privacy in the digital era, cloud security and more."

1 999 175 €

Start date: 2014-10-01, End date: 2019-09-30

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

How do we make reliable decisions given sensory information that is often weak or ambiguous? Current theories center on a brain mechanism whereby sensory evidence is integrated over time into a “decision variable” which triggers the appropriate action upon reaching a criterion. Neural signals fitting this role have been identified in monkey electrophysiology but efforts to study the neural dynamics underpinning human decision making have been hampered by technical challenges associated with non-invasive recording. This proposal builds on a recent paradigm breakthrough made by the applicant that enables parallel tracking of discrete neural signals that can be unambiguously linked to the three key information processing stages necessary for simple perceptual decisions: sensory encoding, decision formation and motor preparation. Chief among these is a freely-evolving decision variable signal which builds at an evidence-dependent rate up to an action-triggering threshold and precisely determines the timing and accuracy of perceptual reports at the single-trial level. This provides an unprecedented neurophysiological window onto the distinct parameters of the human decision process such that the underlying mechanisms of several major behavioral phenomena can finally be investigated. This proposal seeks to develop a systems-level understanding of perceptual decision making in the human brain by tackling three core questions: 1) what are the neural adaptations that allow us to deal with speed pressure and variations in the reliability of the physically presented evidence? 2) What neural mechanism determines our subjective confidence in a decision? and 3) How does aging impact on the distinct neural components underpinning perceptual decision making? Each of the experiments described in this proposal will definitively test key predictions from prominent theoretical models using a combination of temporally precise neurophysiological measurement and psychophysical modelling.

1 382 643 €

Start date: 2015-05-01, End date: 2020-04-30

Staggering amounts of information are stored in natural language documents, rendering them unavailable to data-science techniques. Information Extraction (IE), a subfield of Natural Language Processing (NLP), aims to automate the extraction of structured information from text, yielding datasets that can be queried, analyzed and combined to provide new insights and drive research forward. Despite tremendous progress in NLP, IE systems remain mostly inaccessible to non-NLP-experts who can greatly benefit from them. This stems from the current methods for creating IE systems: the dominant machine-learning (ML) approach requires technical expertise and large amounts of annotated data, and does not provide the user control over the extraction process. The previously dominant rule-based approach unrealistically requires the user to anticipate and deal with the nuances of natural language. I aim to remedy this situation by revisiting rule-based IE in light of advances in NLP and ML. The key idea is to cast IE as a collaborative human-computer effort, in which the user provides domain-specific knowledge, and the system is in charge of solving various domain-independent linguistic complexities, ultimately allowing the user to query unstructured texts via easily structured forms. More specifically, I aim develop: (a) a novel structured representation that abstracts much of the complexity of natural language; (b) algorithms that derive these representations from texts; (c) an accessible rule language to query this representation; (d) AI components that infer the user extraction intents, and based on them promote relevant examples and highlight extraction cases that require special attention. The ultimate goal of this project is to democratize NLP and bring advanced IE capabilities directly to the hands of domain-experts: doctors, lawyers, researchers and scientists, empowering them to process large volumes of data and advance their profession.

1 499 354 €

Start date: 2019-05-01, End date: 2024-04-30