BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//SAIL Imperial//Seminars//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-WR-CALNAME:SAIL seminars
X-WR-TIMEZONE:Europe/London
BEGIN:VEVENT
UID:2026-10-23-pawan-kumar@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20261023T120000Z
DTEND:20261023T130000Z
SUMMARY:SAIL seminar: Red teaming frontier AI at the UK AI Security Instit
 ute (Pawan Kumar)
LOCATION:311 Huxley Building
DESCRIPTION:Pawan Kumar (UK AI Security Institute)\n\nBio: M. Pawan Kumar 
 is a Research Director at AISI. Prior to that\, he was a Research Scienti
 st at Google DeepMind\, and a faculty member in the Department of Enginee
 ring Science at the University of Oxford. Alarmed by the enormous potenti
 al harms of AI\, he has recently shifted his research focus away from AI 
 capabilities towards AI security.\n\nThe UK’s AI Security Institute (AI
 SI)\, situated within the Cabinet Office\, works to equip government with
  an evidenced understanding of the risks posed by advanced AI. AISI's Red
  Team stress-tests frontier safety measures: finding ways misuse safeguar
 ds can fail\, investigating unintended and harmful model behaviours\, and
  assessing whether protocols intended to keep models under control actual
 ly work.\n\nIn this talk\, I will introduce AISI and take you inside the 
 Red Team’s research. I’ll explain how the team leverages expertise an
 d creativity in adversarial machine learning to uncover safety failures\,
  establish whether they reflect risks in real deployments\, and understan
 d what causes them. Red Team findings go directly to frontier AI develope
 rs to help improve their systems\, and to UK and allied governments to in
 form decisions about AI risk and mitigations.\n\nFor students and researc
 hers interested in frontier AI safety and security\, or tackling difficul
 t technical problems with direct public impact\, the talk will offer a wi
 ndow into the field’s open research questions\, what working on them lo
 oks like\, and opportunities to join AISI.
URL:https://sail.doc.ic.ac.uk/seminars/#2026-10-23-pawan-kumar
END:VEVENT
BEGIN:VEVENT
UID:2026-06-10-mykel-kochenderfer@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20260610T140000Z
DTEND:20260610T150000Z
SUMMARY:SAIL seminar: Decision making for safety critical systems (Mykel K
 ochenderfer)
LOCATION:Room 340\, Huxley Building
DESCRIPTION:Mykel Kochenderfer (Stanford University)\n\nBuilding robust de
 cision making systems for autonomous systems is challenging. Decisions mu
 st be made based on imperfect information about the environment and with 
 uncertainty about how the environment will evolve. In addition\, these sy
 stems must carefully balance safety with other considerations\, such as o
 perational efficiency. Typically\, the space of edge cases is vast\, plac
 ing a large burden on human designers to anticipate problem scenarios and
  develop ways to resolve them. This talk discusses major challenges assoc
 iated with ensuring computational tractability and establishing trust tha
 t our systems will behave correctly when deployed in the real world. We w
 ill outline some methodologies for addressing these challenges and point 
 to some research applications that can serve as inspiration for building 
 safer systems.\n\nBio: Mykel Kochenderfer is an Associate Professor of Ae
 ronautics and Astronautics at Stanford University. He is the director of 
 the Stanford Intelligent Systems Laboratory (SISL)\, conducting research 
 on advanced algorithms and analytical methods for the design of robust de
 cision making systems. His research contributed to what became the intern
 ational standard for aircraft collision avoidance. Prof. Kochenderfer is 
 a co-director of the Center for AI Safety and senior fellow of the Stanfo
 rd Institute for Human-Centered Artificial Intelligence (HAI). He is edit
 or-in-chief of the Journal of Artificial Intelligence Research. He is an 
 author of the textbooks Decision Making under Uncertainty: Theory and App
 lication (MIT Press\, 2015)\, Algorithms for Optimization (MIT Press\, 20
 19)\, Algorithms for Decision Making (MIT Press\, 2022)\, and Algorithms 
 for Validation (MIT Press\, 2026).
URL:https://sail.doc.ic.ac.uk/seminars/#2026-06-10-mykel-kochenderfer
END:VEVENT
BEGIN:VEVENT
UID:2025-11-19-krishnamurthy-dj-dvijotham@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20251119T150000Z
DTEND:20251119T160000Z
SUMMARY:SAIL seminar: Malice in Agentland (Krishnamurthy (Dj) Dvijotham)
LOCATION:140 Huxley Building
DESCRIPTION:Krishnamurthy (Dj) Dvijotham\n\nBio: Krishnamurthy (Dj) Dvijot
 ham is a senior staff research scientist at Google DeepMind where he lead
 s efforts on the development of secure and trustworthy AI agents. He prev
 iously founded the AI security research team at ServiceNow Research and c
 o-founded the robust and verified AI team at DeepMind. His past research 
 has received best paper awards at many leading AI conferences\, including
  most recently at ICML and CVPR 2024. His research led to the framework u
 sed for AI security testing at ServiceNow and has been deployed in severa
 l Google products\, including the Android Play Store\, YouTube and Gemini
 .\n\nAbstract: Modern AI systems are trained on troves of data scraped fr
 om the internet\, often with light oversight and filtering. Recent works 
 have shown that this creates significant risks of embedding malicious con
 tent in the training data\, leading to the creation of poisoned models th
 at systematically bias in favor of specific commercial or political entit
 ies\, or contain hard to detect backdoors that can lie dormant until a sp
 ecific trigger is encountered causing a dangerous output. Furthermore\, r
 elative to prior work on defending against attacks\, these attacks are dy
 namic and adaptive\, responding to the evolving mitigations and defenses.
  We present three results in this context:\n1. A novel fine tuning attack
  that causes frontier AI models to break alignment without triggering any
  of the safeguards implemented in frontier fine tuning APIs.\n2. A novel 
 approach to certifying robustness of learning algorithms to adaptive and 
 dynamic data poisoning attacks.\n3. The evolution of poisoning risks in t
 he world of digital AI agents that can exploit several parts of the suppl
 y chain that exists in the creation of deployed AI agents.\n\nWe end with
  a discussion of the future outlook for the scope of risks and possible m
 itigations as we train increasingly capable and widely deployed AI agents
 .
URL:https://sail.doc.ic.ac.uk/seminars/#2025-11-19-krishnamurthy-dj-dvijot
 ham
END:VEVENT
BEGIN:VEVENT
UID:2025-11-03-dario-paccagnan@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20251103T120000Z
DTEND:20251103T130000Z
SUMMARY:SAIL seminar: Pick to Learn: State of the Art Safety Guarantees fo
 r Machine Learning and Control (Dario Paccagnan)
LOCATION:I-X Level 5 Conference Room
DESCRIPTION:Dario Paccagnan\n\nAI models are increasingly embedded in scie
 ntific research and industrial production\, where they inform predictions
  and drive decisions. In safety-critical settings\, such as autonomous dr
 iving or medical diagnostics\, deploying these models requires rigorous s
 afety and performance guarantees\, a need that has motivated extensive wo
 rk at the interface of statistical learning and verification. In this tal
 k\, Professor Dario Paccagnan will present a recent line of work from his
  group that overcomes these limitations by enabling the use of all availa
 ble data to jointly train models and equip them with tight safety and per
 formance guarantees.
URL:https://sail.doc.ic.ac.uk/seminars/#2025-11-03-dario-paccagnan
END:VEVENT
BEGIN:VEVENT
UID:2024-02-05-omri-isac@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20240205T130000Z
DTEND:20240205T140000Z
SUMMARY:SAIL seminar: Neural Network Verification with Proof Production (O
 mri Isac)
LOCATION:I-X Seminar Room
DESCRIPTION:Omri Isac\n\nDeep neural networks (DNNs) are increasingly bein
 g employed in safety-critical systems\, and there is an urgent need to gu
 arantee their correctness. Consequently\, the verification community has 
 devised multiple techniques and tools for verifying DNNs. When DNN verifi
 ers discover an input that triggers an error\, that is easy to confirm\; 
 but when they report that no error exists\, there is no way to ensure tha
 t the verification tool itself is not flawed. As multiple errors have alr
 eady been observed in DNN verification tools\, this calls the applicabili
 ty of DNN verification into question. In this talk\, I will present a nov
 el mechanism for enhancing Simplex-based DNN verifiers with proof product
 ion capabilities. We will discuss how to create such a mechanism based on
  a constructive adaptation of the well-known Farkas’ lemma\, combined w
 ith mechanisms for handling piecewise-linear functions. We will conclude 
 by discussing how proof production could be leveraged in order to improve
  the performance of DNN verifiers\, using conflict clauses. Omri is a PhD
  student advised by Guy Katz in the Hebrew University of Jerusalem. His r
 esearch mainly focuses on proof production for deep neural network (DNN) 
 verification\, and its theoretical computability and complexity aspects.
URL:https://sail.doc.ic.ac.uk/seminars/#2024-02-05-omri-isac
END:VEVENT
BEGIN:VEVENT
UID:2024-01-29-lorenzo-cazzaro@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20240129T130000Z
DTEND:20240129T140000Z
SUMMARY:SAIL seminar: Efficient and Principled Verification of Tree-Based 
 Classifiers (Lorenzo Cazzaro)
LOCATION:Virtual
DESCRIPTION:Lorenzo Cazzaro\n\nThe increasing success of Machine Learning 
 (ML) has highlighted the importance of defining and verifying the trustwo
 rthy behavior of ML models. Existing research followed two directions for
  addressing these challenges: (i) establishing properties like robustness
  and fairness to characterize the trustworthy behavior of ML models and (
 ii) designing verification algorithms based on formal methods to provide 
 formal guarantees regarding compliance with the desired properties. Unfor
 tunately\, current proposals suffer from several limitations. Complete ve
 rification methods on ML models present limited scalability and do not al
 ways terminate within reasonable time and memory bounds. Additionally\, r
 obustness and fairness properties are not sufficiently expressive since t
 hey are local or data-dependent in that they are predicated solely on spe
 cific test instances rather than arbitrary inputs.\nIn this talk\, I will
  introduce three contributions that attempt to address these shortcomings
 \, focusing on a specific class of ML classifiers\, tree-based classifier
 s. In the first part of the talk\, I will shortly introduce two attempts 
 to go beyond the verification of local robustness and fairness properties
 . These contributions leverage a data-independent stability analysis of t
 he tree-based classifier to verify a generalization of robustness\, resil
 ience\, and a global fairness property. Then\, in the second part of the 
 talk\, I will deepen a novel approach to the security verification of tre
 e-based classifiers\, called verifiable learning\, i.e.\, identifying and
  training ML models that enable (formally proven) efficient security veri
 fication. In particular\, we identify and train a specific class of tree-
 based models called large-spread ensembles that admit efficient complete 
 robustness verification\, moving away from the NP-hardness of the robustn
 ess verification problem. Lorenzo Cazzaro is a third-year Ph.D. student i
 n Computer Science and a teaching assistant at Università Ca' Foscari Ve
 nezia. He graduated cum laude in Computer Science in 2021 and he was a re
 search fellow in the field of adversarial machine learning until 2021. Hi
 s main research interests include formal verification of properties of ma
 chine learning models\, adversarial machine learning\, and applications o
 f artificial intelligence in cybersecurity.
URL:https://sail.doc.ic.ac.uk/seminars/#2024-01-29-lorenzo-cazzaro
END:VEVENT
BEGIN:VEVENT
UID:2024-01-22-linyi-li@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20240122T170000Z
DTEND:20240122T180000Z
SUMMARY:SAIL seminar: Certifying Critical Trustworthiness Properties of La
 rge-Scale DL Systems (Linyi Li)
LOCATION:Virtual
DESCRIPTION:Linyi Li\n\nAlong with the wide deployment of deep learning (D
 L) systems\, their lack of trustworthiness (robustness\, fairness\, numer
 ical reliability\, etc) is raising serious social concerns\, especially i
 n safety-critical scenarios such as autonomous driving\, aircraft navigat
 ion\, and facial recognition. Hence\, a rigorous and accurate evaluation 
 of the trustworthiness of DL systems is critical before their large-scale
  deployment.\n\nIn this talk\, I will introduce my research on certifying
  critical trustworthiness properties of large-scale DL systems. Inspired 
 by techniques in optimization\, cybersecurity\, and software engineering\
 , my work computes rigorous worst-case bounds to characterize the degree 
 of trustworthiness for a given DL system and further improve such bounds 
 via strategic training. Specifically\, I will introduce two representativ
 e frameworks: (1) DSRS: a framework providing theoretically tightest cert
 ification for DL systems. DSRS along with our training method DRT and ope
 n-source tools (VeriGauge and alpha-beta-CROWN) is the state-of-the-art a
 nd award-winning solution for achieving DL robustness against constrained
  perturbations. (2) TSS: a framework providing most scalable certificatio
 n for DL systems against semantic transformations. TSS opens a series of 
 subsequent research on guaranteeing semantic robustness for various downs
 tream DL and AI applications. Besides\, I will briefly discuss some key t
 echniques in alpha-beta-crown\, the winning verification tool in recent V
 NN-COMP competitions. I will conclude this talk with a roadmap that outli
 nes several core research questions and future directions on trustworthy 
 machine learning.
URL:https://sail.doc.ic.ac.uk/seminars/#2024-01-22-linyi-li
END:VEVENT
BEGIN:VEVENT
UID:2023-11-20-mark-muller@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20231120T130000Z
DTEND:20231120T140000Z
SUMMARY:SAIL seminar: Training and Verification of Robust Neural Networks 
 (Mark Muller)
LOCATION:Room 311\, Huxley Building
DESCRIPTION:Mark Muller\n\nFollowing the discovery of adversarial examples
 \, provable robustness guarantees for neural networks have received incre
 asing attention from the research community. While relatively small or he
 avily regularized models with limited standard accuracy can now be effici
 ently analyzed\, obtaining guarantees for more accurate models remains an
  open problem. Recently\, a new verification paradigm has emerged that ta
 ckles this challenge by combining a Branch-and-Bound approach with precis
 e multi-neuron constraints. The resulting\, more precise verifiers have i
 n turn enabled novel certified training methods which reduce (over-)regul
 arization to obtain more precise yet certifiable networks. In this talk\,
  we discuss these certification and training methods.\nMark N. Müller is
  a senior Ph.D. student at the Secure\, Reliable\, and Intelligent System
 s Lab of ETH Zürich\, advised by Prof. Martin Vechev. Mark’s research 
 focuses on provable guarantees for machine learning models. This includes
  both deterministic and probabilistic certification methods for a diverse
  range of architectures and certified training methods. He has led the re
 search on these topics at SRI Lab for the last two years\, including two 
 industry collaborations\, co-organized the Verification of Neural Network
 s Competition 2022 (VNN-Comp’22)\, and was the lead organizer for the 2
 nd Workshop on Formal Verification of Machine Learning (WFVML'23) at ICML
 ’23 this year.
URL:https://sail.doc.ic.ac.uk/seminars/#2023-11-20-mark-muller
END:VEVENT
BEGIN:VEVENT
UID:2022-11-25-mykel-kochenderfer@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20221125T110000Z
DTEND:20221125T120000Z
SUMMARY:SAIL seminar: Automated Decision Making for Safety Critical Applic
 ations (Mykel Kochenderfer)
LOCATION:Room 144\, Huxley Building
DESCRIPTION:Mykel Kochenderfer\n\nBuilding robust decision making systems 
 for autonomous systems is challenging. Decisions must be made based on im
 perfect information about the environment and with uncertainty about how 
 the environment will evolve. In addition\, these systems must carefully b
 alance safety with other considerations\, such as operational efficiency.
  Typically\, the space of edge cases is vast\, placing a large burden on 
 human designers to anticipate problem scenarios and develop ways to resol
 ve them. This talk discusses major challenges associated with ensuring co
 mputational tractability and establishing trust that our systems will beh
 ave correctly when deployed in the real world. We will outline some metho
 dologies for addressing these challenges and point to some research appli
 cations that can serve as inspiration for building safer systems.\nMykel 
 Kochenderfer is an Associate Professor of Aeronautics and Astronautics at
  Stanford University. He is the director of the Stanford Intelligent Syst
 ems Laboratory (SISL)\, conducting research on advanced algorithms and an
 alytical methods for the design of robust decision making systems. Prior 
 to joining the faculty in 2013\, he was at MIT Lincoln Laboratory where h
 e worked on aircraft collision avoidance\, leading to the creation of the
  ACAS X international standard for manned and unmanned aircraft. He recei
 ved his Ph.D. from the University of Edinburgh in 2006 and B.S. and M.S. 
 degrees in computer science from Stanford University in 2003. Prof. Koche
 nderfer is a co-director of the Center for AI Safety. He is an associate 
 editor of the Journal of Artificial Intelligence Research and the Journal
  of Aerospace Information Systems. He is an author of the textbooks Decis
 ion Making under Uncertainty - Theory and Application (MIT Press\, 2015)\
 , Algorithms for Optimization (MIT Press\, 2019)\, and Algorithms for Dec
 ision Making (MIT Press\, 2022).
URL:https://sail.doc.ic.ac.uk/seminars/#2022-11-25-mykel-kochenderfer
END:VEVENT
BEGIN:VEVENT
UID:2020-03-02-michael-akintunde@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20200302T120000Z
DTEND:20200302T130000Z
SUMMARY:SAIL seminar: Formal Verification of Neural Agents in Non-determin
 istic Environments (Michael Akintunde)
LOCATION:Room 217\, Huxley Building
DESCRIPTION:Michael Akintunde\n\nWe introduce a model for agent-environmen
 t systems where the agents are implemented via feed-forward ReLU neural n
 etworks and the environment is non-deterministic. We study the verificati
 on problem of such systems against CTL properties. We introduce a bounded
  fragment of CTL and show its usefulness in identifying shallow bugs in t
 he system. We present a novel parallel algorithm for MILP-based verificat
 ion of agent-environment systems\, present an implementation\, and report
  the experimental results obtained against a variant of the VerticalCAS u
 se-case.
URL:https://sail.doc.ic.ac.uk/seminars/#2020-03-02-michael-akintunde
END:VEVENT
BEGIN:VEVENT
UID:2020-02-17-elena-botoeva@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20200217T120000Z
DTEND:20200217T130000Z
SUMMARY:SAIL seminar: Verification of Neural Networks: Specifying Global R
 obustness using Generative Models (Dr Elena Botoeva)
LOCATION:Room 218\, Huxley Building
DESCRIPTION:Dr Elena Botoeva\n\nThe success of neural networks across most
  machine learning tasks and the persistence of adversarial examples have 
 made the verification of such models an important quest. Several techniqu
 es have been successfully developed to verify robustness\, and are now ab
 le to evaluate neural networks with thousands of nodes. The main weakness
  of this approach is in the specification: robustness is asserted on a va
 lidation set consisting of a finite set of examples\, i.e. locally. We pr
 opose a notion of global robustness based on generative models\, which as
 serts the robustness on a very large and representative set of examples. 
 We show how this can be used for verifying neural networks. In this paper
  we experimentally explore the merits of this approach\, and show how it 
 can be used to construct realistic adversarial examples.
URL:https://sail.doc.ic.ac.uk/seminars/#2020-02-17-elena-botoeva
END:VEVENT
BEGIN:VEVENT
UID:2020-02-03-henrik-aslund@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20200203T120000Z
DTEND:20200203T130000Z
SUMMARY:SAIL seminar: Verifiably Safe Off-Model Reinforcement Learning (He
 nrik Åslund)
LOCATION:Room 218\, Huxley Building
DESCRIPTION:Henrik Åslund\n\nThe desire to use reinforcement learning in 
 safety-critical settings has inspired a recent interest in formal methods
  for learning algorithms. Existing formal methods for learning and optimi
 zation primarily consider the problem of constrained learning or constrai
 ned optimization. Given a single correct model and associated safety cons
 traint\, these approaches guarantee efficient learning while provably avo
 iding behaviors outside the safety constraint. Acting well given an accur
 ate environmental model is an important pre-requisite for safe learning\,
  but is ultimately insufficient for systems that operate in complex heter
 ogeneous environments. This paper introduces verification-preserving mode
 l updates\, the first approach toward obtaining formal safety guarantees 
 for reinforcement learning in settings where multiple possible environmen
 tal models must be taken into account. Through a combination of inductive
  data and deductive proving with design-time model updates and runtime mo
 del falsification\, we provide a first approach toward obtaining formal s
 afety proofs for autonomous systems acting in heterogeneous environments.
URL:https://sail.doc.ic.ac.uk/seminars/#2020-02-03-henrik-aslund
END:VEVENT
BEGIN:VEVENT
UID:2020-01-27-yi-ling-liu@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20200127T120000Z
DTEND:20200127T130000Z
SUMMARY:SAIL seminar: Adversarial Training for Image Recognition (Yi-Ling 
 Liu)
LOCATION:Room 218\, Huxley Building
DESCRIPTION:Yi-Ling Liu\n\nWe present a novel black-box adversarial traini
 ng algorithm to defend against state-of-the-art attack methods in machine
  learning. In order to search for an adversarial attack\, the algorithm a
 nalyses small regions around the input that are likely to make significan
 t contributions for the generation of adversarial samples. Unlike some of
  the literature in the area\, the proposed method does not require access
  to the internal layers of the model and is therefore applicable to appli
 cations such as security. We report the experimental results obtained on 
 models of different sizes built for the MNIST and CIFAR10 datasets. The r
 esults suggest that known attacks on the resulting models are less transf
 erable than those models trained by state-of-the art attack algorithms.
URL:https://sail.doc.ic.ac.uk/seminars/#2020-01-27-yi-ling-liu
END:VEVENT
BEGIN:VEVENT
UID:2020-01-20-francesco-leofante@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20200120T120000Z
DTEND:20200120T130000Z
SUMMARY:SAIL seminar: AI Planning meets Production Logistics (Francesco Le
 ofante)
LOCATION:Room 217\, Huxley Building
DESCRIPTION:Francesco Leofante\n\nPlanning for logistics requires algorith
 ms and tools that can handle the challenges such scenarios pose. On the o
 ne hand\, expressive languages are needed to build faithful models\; on t
 he other hand\, efficient solving techniques that can support these langu
 ages need to be devised. The existence of this trade-off\, however\, does
  not mean progress is not possible! Using the RoboCup Logistics League as
  a testbed\, I will focus on production logistics and show how recent adv
 ances in satisfiability checking can be used to solve some interesting pr
 oblems in the field.
URL:https://sail.doc.ic.ac.uk/seminars/#2020-01-20-francesco-leofante
END:VEVENT
BEGIN:VEVENT
UID:2020-01-13-patrick-henriksen@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20200113T120000Z
DTEND:20200113T130000Z
SUMMARY:SAIL seminar: Efficient Neural Network Verification via Adaptive R
 efinement and Adversarial Search (Patrick Henriksen)
LOCATION:Room 217\, Huxley Building
DESCRIPTION:Patrick Henriksen\n\nWe propose a novel verification method fo
 r high-dimensional feed-forward neural networks governed by ReLU\, Sigmoi
 d and Tanh activation functions. We show that the method is complete for 
 ReLU networks and sound for other networks. The technique extends symboli
 c interval propagation by using gradient descent to locate counter-exampl
 es from spurious solutions generated by the associated LP problems. The a
 pproach includes a novel adaptive splitting strategy intended to refine t
 he nodes with the greatest impact on the output of the network. The resul
 ting implementation\, called VERINET\, generally achieved speed-ups of an
  order of magnitude for safe cases and a speed-up of 3 orders of magnitud
 e for unsafe cases on MNIST models. We also show that VERINET can verify 
 networks of over 50\,000 ReLU nodes trained on the CIFAR-10 data-set\, a 
 network significantly larger than networks previously verified via comple
 te methods.
URL:https://sail.doc.ic.ac.uk/seminars/#2020-01-13-patrick-henriksen
END:VEVENT
BEGIN:VEVENT
UID:2019-12-19-dragos-margineantu-boeing@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20191219T100000Z
DTEND:20191219T110000Z
SUMMARY:SAIL seminar: Robust Machine Learning and Scalable Interactive Dec
 ision Systems (Dr Dragos Margineantu (Boeing))
LOCATION:Room 144\, Huxley Building
DESCRIPTION:Dr Dragos Margineantu (Boeing)\n\nDeployable decision systems 
 require much more than end-to-end learned models. This talk will focus on
  two of the questions that open up for the machine learning scientist in 
 the process of architecting and implementing modern decision systems: [A]
  how to design robust systems and [B] how to build interactive decision s
 ystems.\nResearch questions on machine learning robustness fall into two 
 broad categories: [A] model correctness analysis\, verification & validat
 ion and [B] bounding the risk of handling (predictions & actions) observa
 tions on which the learner is not qualified to handle. The first part of 
 the talk will focus on the latter set of questions: how do we learn model
 s that 'know when they don't know'\, what are the formalisms that we need
 \, what is practically doable - both\, for supervised learning and reinfo
 rcement learning. First we will explore methods for learning self-compete
 nce models in addition to the predictions. Next\, we will discuss self-co
 mpetence estimation approaches for reinforcement learning and decision ma
 king. Finally\, we will explore multi-faceted learning and related techni
 ques\, that can be applied in decision systems with the goal of increased
  robustness.\nThe second part of the talk will focus on architecting scal
 able interactive systems. Here\, we will explore how Inverse Reinforcemen
 t Learning methods can be applied in a user-in-the-loop setting for the d
 etection of anomalous actions and for intent recognition.
URL:https://sail.doc.ic.ac.uk/seminars/#2019-12-19-dragos-margineantu-boei
 ng
END:VEVENT
BEGIN:VEVENT
UID:2019-12-02-elena-botoeva@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20191202T120000Z
DTEND:20191202T130000Z
SUMMARY:SAIL seminar: Strong Mixed-Integer Programming Formulations for Tr
 ained Neural Networks (Dr Elena Botoeva)
LOCATION:Room 217\, Huxley Building
DESCRIPTION:Dr Elena Botoeva\n\nWe present an ideal mixed-integer programm
 ing (MIP) formulation for a rectified linear unit (ReLU) appearing in a t
 rained neural network. Our formulation requires a single binary variable 
 and no additional continuous variables beyond the input and output variab
 les of the ReLU. We contrast it with an ideal "extended" formulation with
  a linear number of additional continuous variables\, derived through sta
 ndard techniques. An apparent drawback of our formulation is that it requ
 ires an exponential number of inequality constraints\, but we provide a r
 outine to separate the inequalities in linear time. We also prove that th
 ese exponentially-many constraints are facet-defining under mild conditio
 ns. Finally\, we study network verification problems and observe that dyn
 amically separating from the exponential inequalities 1) is much more com
 putationally efficient and scalable than the extended formulation\, 2) de
 creases the solve time of a state-of-the-art MIP solver by a factor of 7 
 on smaller instances\, and 3) nearly matches the dual bounds of a state-o
 f-the-art MIP solver on harder instances\, after just a few rounds of sep
 aration and in orders of magnitude less time.
URL:https://sail.doc.ic.ac.uk/seminars/#2019-12-02-elena-botoeva
END:VEVENT
BEGIN:VEVENT
UID:2019-11-18-alexander-manta@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20191118T130000Z
DTEND:20191118T140000Z
SUMMARY:SAIL seminar: Semantic Labelling and Learning for Parity Game Solv
 ing in LTL Synthesis (Alexander Manta)
LOCATION:Room 217\, Huxley Building
DESCRIPTION:Alexander Manta\n\nThe paper proposes "semantic labelling" as 
 a novel ingredient for solving games in the context of LTL synthesis. It 
 exploits recent advances in the automata-based approach\, yielding more i
 nformation for each state of the generated parity game than the game grap
 h can capture. Utilizing this extra information even in simple heuristics
  can help improve the standard strategy iteration approach to yield a win
 ning strategy directly without any computation and also makes strategy it
 eration yield smaller solutions. Furthermore\, equipping Q-learning with 
 semantic information makes it competitive to strategy iteration. The expe
 rimental results show that already the simplest heuristics can achieve si
 gnificant improvements demonstrating the utility of semantic labelling fo
 r more advanced learning approaches both for initialization and improveme
 nt of strategies.
URL:https://sail.doc.ic.ac.uk/seminars/#2019-11-18-alexander-manta
END:VEVENT
BEGIN:VEVENT
UID:2019-10-21-panagiotis-kouvaros@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20191021T120000Z
DTEND:20191021T130000Z
SUMMARY:SAIL seminar: Verification of ReLU-based Neural Networks (Dr Panag
 iotis Kouvaros)
LOCATION:Room 361\, Huxley Building
DESCRIPTION:Dr Panagiotis Kouvaros
URL:https://sail.doc.ic.ac.uk/seminars/#2019-10-21-panagiotis-kouvaros
END:VEVENT
BEGIN:VEVENT
UID:2018-12-10-markus-n-rabe@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20181210T100000Z
DTEND:20181210T110000Z
SUMMARY:SAIL seminar: Deep Learning for Automated Reasoning (Markus N. Rab
 e)
LOCATION:Room 217\, Huxley Building
DESCRIPTION:Markus N. Rabe\n\nUntil today\, automated reasoning and theore
 m proving mostly relied on combinatorial search algorithms and handwritte
 n heuristics. But the success of deep learning in various application are
 as prompts us to revisit those paradigms. In particular AlphaGo\, DeepMin
 d's system mastering board games such as Go and Chess\, has demonstrated 
 that deep reinforcement learning can be fused with combinatorial search a
 lgorithms with great success. This talk will discuss two projects on leve
 raging deep learning for automated reasoning and theorem proving. The fir
 st part concerns learning better heuristics for a solver for quantified B
 oolean formulas\, and the second part will give an outlook on the ongoing
  effort in Google on learning how to prove theorems in higher-order logic
 . I will focus on the challenges and tradeoffs that arise from the combin
 ation of the two worlds.
URL:https://sail.doc.ic.ac.uk/seminars/#2018-12-10-markus-n-rabe
END:VEVENT
BEGIN:VEVENT
UID:2018-07-24-prof-mykel-kochenderfer@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20180724T100000Z
DTEND:20180724T110000Z
SUMMARY:SAIL seminar: Building Trust in AI for Safety-Critical Systems (Pr
 of. Mykel Kochenderfer)
LOCATION:Room 217\, Huxley Building
DESCRIPTION:Prof. Mykel Kochenderfer\n\nStarting in the 1970s\, decades of
  effort went into building human-designed rules for providing automatic m
 aneuver guidance to pilots to avoid mid-air collisions. The resulting sys
 tem was later mandated worldwide on all large aircraft and significantly 
 improved the safety of the airspace. Recent work has investigated the fea
 sibility of using computational techniques to help derive optimized decis
 ion logic that better handles various sources of uncertainty and balances
  competing system objectives. This approach has resulted in a system call
 ed Airborne Collision Avoidance System (ACAS) X that significantly reduce
 s the risk of mid-air collision while also reducing the alert rate\, and 
 it is in the process of becoming the next international standard. Using A
 CAS X as a case study\, this talk will discuss lessons learned about buil
 ding trust in advanced decision support systems. This talk will also outl
 ine research challenges in facilitating greater levels of automation into
  safety critical systems.
URL:https://sail.doc.ic.ac.uk/seminars/#2018-07-24-prof-mykel-kochenderfer
END:VEVENT
BEGIN:VEVENT
UID:2018-07-09-xiaowei-huang@sail.doc.ic.ac.uk
DTSTAMP:20261118T150713Z
DTSTART:20180709T100000Z
DTEND:20180709T110000Z
SUMMARY:SAIL seminar: Approximate Verification of Deep Neural Networks wit
 h Provable Guarantees (Xiaowei Huang)
LOCATION:Room 218\, Huxley Building
DESCRIPTION:Xiaowei Huang\n\nDeep neural networks (DNNs) have achieved imp
 ressive experimental results in image classification\, but can surprising
 ly be unstable with respect to adversarial perturbations\, that is\, mini
 mal changes to the input image that cause the network to misclassify it. 
 With potential applications including perception modules and end-to-end c
 ontrollers for self-driving cars\, this raises concerns about their safet
 y. In this talk\, I will report our recent works towards the verification
  of feed-forward multi-layer DNNs. We enable exhaustive search of the reg
 ion by employing discretisation\, reduce the verification problem to the 
 finding of an optimal solution in a two-player game\, and propagate the a
 nalysis layer by layer. By utilising the fact that DNNs are Lipschitz con
 tinuous\, we show that our approach can achieve provable guarantees. We i
 mplemented our approach into a software tool and conducted experiments on
  several DNNs trained on datasets such as MNIST\, CIFAR10\, and imageNet\
 , and a DNN for Nexar traffic light challenge.
URL:https://sail.doc.ic.ac.uk/seminars/#2018-07-09-xiaowei-huang
END:VEVENT
END:VCALENDAR
