Getting hints from random walks in Optimization and Deep Learning
Comparison to random behavior can help monitor and automatically control an optimization process, as demonstrated more than 20 years ago by the CMA-ES algorithm (Covariance Matrix Adaptation Evolution Strategy), today recognized as one of the best performing Black Box continuous optimizer.
After briefly surveying CMA-ES principles, the first part of the talk will provide evidence that the hyperparameters that control CMA-ES covariance matrix adaptation could in turn be tuned to the problem at hand, and present EPM-CMA-ES, a Per Instance Algorithm Configuration for CMA-ES: Based on features describing the objective function, and provided enough examples of the algorithm behavior are available, an Empirical Performance Model (EPM) is learned, and used to configure the hyperparameters for unknown objective functions. EPM-CMA-ES won the GECCO Black-Box Optimization Competition last July.
The second part of the talk will introduce S-ALERA, yet another optimizer for Deep Learning: Based on a similar principle of comparison with a random walk, ALERA (Agnostic LEarning Rate Adaptation) gracefully increases or decreases the learning rate of Stochastic Gradient Descent. However, whereas allowing the learning rate to also increase can indeed speed-up the optimization, it makes the algorithm more prone to catastrophic events, a well-known issue in Deep Learning. A statistical test using Page-Hinkley change point detection is hence added, in order to detect such events, and then cool down the optimization process. The resulting algorithm (SALERA — Safe ALERA) compares favorably to the state-of-the-art on standard benchmarks, while limiting the risks of dramatic explosions.
Marc Schoenauer graduated at Ecole Normale Supérieure in Paris. He is today Principled Senior Researcher (DR1), co-head of the TAO team with Michèle Sebag at INRIA Saclay Île-de-France since Sept. 2003. He has been working in the fields of Evolutionary Computation and Machine Learning since the early 90s, is author of more than 150 papers in journals and major conferences of the fields, and has been (co-)advisor of 33 PhD students.
Marc Schoenauer is Chair of SIGEVO, the ACM Special Interest Group for Evolutionary Computation. He was the founding president of Evolution Artificielle, the French Society for Evolutionary Computation, and has been president of AFIA, the French Association for Artificial Intelligence. He has been Editor in Chief of Evolutionary Computation Journal, is or has been Associate Editor of IEEE Transactions on Evolutionary Computation, TCS-C (Theory of Natural Computing), Genetic Programming and Evolvable Machines, and the Journal of Applied Soft Computing.
On Constructing Cyber-Analytics
Enterprise network defense is providing great opportunities for the development and deployment of statistical and machine learning methods. Such methods are intended to complement existing defenses, such as firewalls, virus scanners, and intrusion detection systems – which are predominantly signature-based. The role of data analysis methods is to provide enhanced situation awareness, by providing monitoring and alerting mechanisms to detect departures from “normal” behavior. In developing analytics in this context, a variety of challenging problems need to be addressed, including the volume and velocity of the data, high levels of heterogeneity, temporal variation, and more. We review aspects of the problem and characteristics of the various data sources. At present, the vision of jointly modelling various data sources at different levels of network abstraction, appears out of reach due to data volume and timeliness concerns. Instead, we describe a set of novel, and often simple, analytics that operate within different levels of the abstract hierarchy.
Niall Adams is Professor of Statistics, and head of the Statistics section, in the Mathematics Department at Imperial College London. His primary research focus is the development of statistical and machine learning methodology for enterprise cyber-security. From 2011 to 2016, he was seconded to the Heilbronn Institute for Mathematical Research at the University of Bristol, to lead the cyber-security data mining team. Streaming analytics are an area of particular methodological focus, and he has developed a variety of methods for streaming classification and change-point detection.
Dynamic EEG signatures underlying coma recovery after cardiac arrest: a machine learning approach
What is the major problem being addressed by this research?
A cardiac arrest is a medical emergency in which the heart stops beating. Lack of oxygen delivery to the brain during a cardiac arrest can cause significant damage to the brain. Because of that, many patients are unresponsive by the time they arrive to the hospital and severe brain injury is the leading cause of mortality in these patients. Most families want to know if the brain injury leading to their loved ones’ unresponsive state is reversible or not. The
electroencephalogram (EEG) is a test that measures brain’s electrical activity continuously using small electrodes attached to the patient’s head. Our research aims to develop a ‘brain wave’ monitoring technology that will measure brain function and predict which patients might be able to wake up and which patients had severe enough brain injury to make recovery unlikely.
What specific questions are you asking and how will you attempt to answer them?
We are specifically studying how the evolution of EEG signal trends in comatose cardiac arrest patients might help us identify which patients have better chances of recovering enough to carry on regular daily activities independently. Previous studies suggest that specific EEG features are surrogate markers of thalamocortical networks function. Integrity of these large-scale cerebral networks have been linked with emergence from coma.
In this study, we have recorded continuous EEG signals in more than 500 comatose patients for few days after a cardiac arrest, and we have also acquired several quantitative EEG features from the original raw EEG signal. As recovery from brain injury might take weeks to months, we have evaluated their neurological function six-months after hospital discharge to evaluate the patient’s recovery. We expect that patients who recover to be independent have distinct QEEG signatures trends from those that do not. Machine-learning methods may help us identify which features at specific time-windows have the best performance predicting functional recovery.
Overall, what is the potential impact of this work to patients, families, and society?
Families and health-care providers caring for a loved one with brain injury after cardiac arrest frequently are unsure if continuing aggressive medical treatments will help the patient improve or if that will prolong the inevitable process of dying. By developing a diagnostic and monitoring tool that can early and accurately predict which patients have potential to recover will facilitate patient selection to clinical trials focusing on treatments that enhance neurological recovery. The ability to monitor the brain’s response to these new treatments may offer new insights into the mechanisms underlying coma recovery and also support the development of individualized interventions based on specific patient’s responses. Taken altogether, this brain-monitoring platform will open several opportunities to improve the health of patients with brain injury caused by cardiac arrest and potentially to other neurological diseases such as traumatic brain injury, stroke, and disorders of consciousness.
Dr. Edilberto Amorim is a neurologist who specializes in critical care and clinical neurophysiology. Dr. Amorim completed medical training at the Bahiana School of Medicine in Brazil and neurology residency at the University of Pittsburgh. He subsequently pursued fellowship training in Neurocritical care at the Massachusetts General Hospital and Brigham and Women’s Hospital followed by a Clinical Neurophysiology fellowship at the Massachusetts General Hospital. Dr. Amorim is board certified in Neurology and devotes clinical time to the MGH Neurosciences ICU. He is currently a Research Fellow at the Massachusetts General Hospital and Visiting Scientist at the Massachusetts Institute of Technology. His clinical research is focused on coma and disorders of consciousness, hypoxic-ischemic brain injury, and noninvasive neuromonitoring in critical ill patients.
Evolving Neural Programs for Continuous Learning
Artificial neural networks (ANNs) have recently made large advances in the field of continuous control tasks. Both on-policy and off-policy reinforcement learning (RL) algorithms which train ANNs have shown impressive results in tasks such as classic RL control tasks, robotic control, and video game playing with pixel input. However, these training methods are limited by their inability to generalize to different tasks after learning a specific task, termed catastrophic forgetting, and their need for a large set of training examples. These key features of continuous learning, the ability to learn new skills while retaining previous knowledge and the ability to learn on a small set of examples, are found in biologic neural networks and contemporary neuroscience has greatly advanced understanding of some of their underlying mechanisms. In this talk, I will examine existing artificial neural models, ranging from deep learning to evolutionary and developmental methods, as they relate to continuous learning. I will then discuss an evolutionary model, currently under development, which explores existing neural models and discovers new models for competition in an open continuous learning environment that assesses the catastrophic forgetting and learning rate of each model.
Dennis G Wilson '14 is currently a PhD candidate at the Institut de Recherche en Informatique de Toulouse studying artificial neural development. During their time at MIT, they worked for three years as a UROP in the Anyscale Learning For All group in CSAIL, applying evolutionary strategies and developmental models to the complex problem of wind farm layout optimization. Their current work and more can be found at d9w.xyz.
Machine Learning and Evolutionary Computation in Cryptography
In this talk, we address several applications of metaheuristics and machine learning in cryptography. First, we start with a brief introduction on machine learning and metaheuristic techniques and following that, we address two practical scenarios. More specifically, as the first example we consider a combinatorial optimization problem in the form of designing S-boxes with good cryptographic and implementation properties (with a special emphasis on area and power perspective). Moreover, we discuss how to use population based metaheuristic techniques like those belonging to the evolutionary computation area in the design process.
Following that, we investigate how to use machine learning in side-channel attacks. There, we elaborate on the life-cycle of a machine learning process with an example application taken from the DPA contest web site, more specifically, attacking masked AES implementation. Finally, we give a brief conclusion and address several possible future research avenues.
Stjepan Picek is a postdoctoral researcher in the Computer Security and Industrial Cryptography (COSIC) group at KU Leuven, Belgium. His research interests are applied cryptography, machine learning, and evolutionary computation. He finished his PhD in 2015 as a double doctorate under the supervision of Lejla Batina (Radboud University Nijmegen, The Netherlands) and Domagoj Jakobovic (Faculty of Electrical Engineering and Computing, Croatia). Prior to that, Stjepan worked in industry and government. He regularly publishes papers in both evolutionary computation and cryptography conferences. Stjepan is also a member of the organization committee for International Summer School in Cryptography and the vice-president of the Croatian IEEE CIS Chapter, as well as the member of several professional societies (ACM, IEEE, IACR).
Cognitive Learning using Evolutionary Computation
Artificial Cognitive Systems encompasses machine intelligence systems, such as robots, that interact with their environment. This talk will highlight research that enables such systems to learn and adapt to problems in their domain and in related domains. The symbolic evolutionary computation technique of Learning Classifier Systems (LCSs) was conceived 40 years ago as an artificial cognitive system. The work presented shows how LCSs can utilise building blocks of knowledge in heuristics ('if-then' rules) to transfer learnt knowledge from small to large scale problems in the same domain. Furthermore, the use of these
rules enables functionality learned in sub-problems to be transferred to related problems. Results show that provided the human experimenter can set a rough curricula for learning concepts, the underlying patterns/models in a problem domain can be learnt in an interpretable
Will Browne received a BEng Mechanical Engineering, Honours degree from the University of Bath, UK in 1993, MSc in Energy (1994) and EngD (Engineering Doctorate scheme, 1999) University of Wales, Cardiff. After eight years lecturing in the Department of Cybernetics, University Reading, UK, he was appointed to School of Engineering and Computer Science, Victoria University of Wellington, NZ in 2008. Associate Professor Browne's main area of research is Applied Cognitive Systems. This includes Learning Classifier Systems, Cognitive Robotics, and Modern Heuristics for industrial application. Blue skies research includes analogues of emotions, abstraction, memories, dissonance and machine consciousness. He is an Associate Editor for Neural Computing and Applications, and Applied Soft Computing. He has published over 100 academic papers, including in IEEE Transactions on Evolutionary Computation on scalable learning and two best paper awards in Genetic and Evolutionary Computation Conference.
The dynamics underlying global spread of emerging infectious diseases
Large-scale computational models parameterized with worldwide air network (WAN) and populations is the mainstream tool for studying global spread of emerging infectious diseases (EIDs). In addition to advanced global epidemic simulators such as GLEAM (http://www.gleamviz.org/), recent analytical studies have partially revealed how epidemic arrival time (EAT) for different populations in the WAN depend on local epidemic growth rates and the network features of the WAN. In this work, we developed a novel probabilistic model to describe global spread of EIDs by (i) using nonhomogeneous Poisson process to model air travel of infected cases in the WAN and (ii) accounting for the effect of continuous importation of infected cases on EAT in each population. We show that the resulting analytical formulas for EATs obtained from this model provide an accurate and explicit characterization of how EAT for different populations in the WAN depend on local epidemic growth rates and the network features of WAN.
Joseph Wu leads the infectious disease modeling research in the School of Public Health at The University of Hong Kong (HKU). He is an affiliated faculty member of the Center for Communicable Diseases Dynamics (CCDD) at the Harvard School of Public Health. He received his PhD in Operations Research from MIT in 2003 and BS in Chemical Engineering from MIT in 1999. His primary research is in influenza epidemiology and control, particularly focusing on pandemic preparedness and response. His work primarily entails developing mathematical models to assess the potential benefits and resource requirement of mitigation and surveillance strategies for influenza epidemics. Besides influenza, he is currently working on the health economic evaluation of EV71 vaccination in China and HPV vaccination in Hong Kong. In 2014, he led the production of HKU's first Massive Open Online Courseware (MOOC) called Epidemics which had more than 10,000 enrollees on the edX platform.