Academic Year 2016-2017
The Institute for Genomics and Bioinformatics is pleased to announce its Distinguished Speaker Series for the academic year 2016-2017.
| Feb 23, 2017 Bren Hall 4011 4:00pm |
Bruce McNaughton Distinguished Professor, Center for the Neurobiology of Learning and Memoryr School of Biological Sciences, UC IrvineThe extraction of knowledge from memory: How our brains create a model of the worldAbstract: Damage to the hippocampal formation (HC) results in severe deficits in the acquisition of two types of memory, which are categorized as “episodic memory” and “semantic memory”. The former refers to internal records of autobiographical experience that can be verbally described in terms of specific events and their spatiotemporal contexts, whereas the latter refers to generalized knowledge; our internal model of the world and its statistical and categorical structure and temporal dynamics, from which we are able to make predictions about the properties of objects and situations, and the outcomes of events and of our own actions. Although the issue of whether episodic memory, once stored, survives hippocampal removal has lately come under some re-evaluation, there is general agreement that semantic memory is retained after such damage. Thus, semantic memory must reside elsewhere in the brain, particularly in the neocortex (NC). As an overall research program, I am concerned with the neural interactions between and within HC and NC that may underlie the creation of knowledge, and how a NC that is rich in knowledge differs in its representational statistics from a relatively naïve cortex. This is a broad topic with many important questions to be addressed. We currently use high density “neural ensemble” recording and optical imaging of cortical voltage activity (VSDI) to test two key hypotheses: the “interleaved memory reactivation” hypothesis and the “hierarchical clustering” hypothesis for neural patterns underlying cortical knowledge representation. |
| Feb 14, 2017 Bren Hall 4011 11:00am |
Casey S. Greene, Ph.D. Assistant Professor Dept. of Systems Pharmacology and Translational Therapeutics Perelman School of Medicine University of PennsylvaniaDiscovery-oriented analysis of public gene expression compendiaAbstract: There’s a lot of public data available. For example, anybody with an internet connection can download more than 2 million genome-wide assays of gene expression. Analyzing these data remains challenging. We’d traditionally use meta-analysis methods, but often public data lack the annotations that enable meta-analysis. If we could surmount these barriers, however, we’d have a multi-billion-dollar resource at our fingertips. Our lab develops algorithms to integrate these heterogeneous, noisy, and often poorly or incorrectly annotated data. We focus specifically on algorithms that are unsupervised and robust to noise, which allows us to tackle unannotated noisy data. We’ve shown that these algorithms can robustly reveal biological features in data from cancer biopsies to microbial systems. In addition to our research focus, we put a heavy emphasis on transparency and reproducibility. I’ll discuss how we’re using continuous integration systems to perform inherently reproducible bioinformatics and computational biology analyses. |
Academic Year 2014-2015
The Institute for Genomics and Bioinformatics is pleased to announce its Distinguished Speaker Series for the academic year 2014-2015.
| Feb 20, 2015 Bren Hall 6011 11:00am |
Pavel Pevzner Department of Computer Science and Engineering University of California at San DiegoLife After MOOCs: Online Science Education Needs a New RevolutionAbstract: Universities continue to pack hundreds of students into a single classroom, despite the fact that this “hoarding” approach has little pedagogical value. Hoarding is particularly objectionable in STEM courses, where learning a complex idea is comparable to navigating a labyrinth. In the large classroom, once a student takes a wrong turn, the student has limited opportunities to ask a question, resulting in an educational breakdown, or the inability to progress further without individualized guidance.A recent revolution in online education has largely focused on making low-cost equivalents of hoarding classes, as many MOOCs are mirror images of their offline counterparts. This is one of the reasons why prominent computer scientist Moshe Vardi published an editorial in Communications of the ACM expressing concerns about the pedagogical quality of MOOCs and including the sentiment, “If I had my wish, I would wave a wand and make MOOCs disappear”. I share the concerns about the quality of early primitive MOOCs, which have been hyped by many as a cure-all for education. At the same time, I feel that much of the criticism of MOOCs stems from the fact that truly disruptive online educational resources have not been developed yet! For this reason, if I had a wand, I would transform MOOCs into a more effective educational product called a Massive Adaptive Interactive Text (MAIT) that can outperform a professor in a classroom. I argue that computer science is a unique discipline where this transition is about to happen and describe our first steps towards transforming a MOOC into a MAIT that has already outperformed me. I further argue that the future MAIT revolution, in difference from the ongoing MOOC revolution, will profoundly affect the way we all teach (even at top universities that do not consider MOOCS as a threat) and will eventually eliminate hoarding classes. |
| Jan 26, 2015 Bren Hall 4011 1:00pm |
Dr. Richard Caruana MicrosoftDo Deep Nets Really Need To Be Deep?Abstract: Currently, deep neural networks are the state of the art on problems such as speech recognition and computer vision. By using a method called model compression, we show that shallow feed-forward nets can learn the complex functions previously learned by deep nets and achieve accuracies previously only achievable with deep models while using the same number of parameters as the original deep models. On the TIMIT phoneme recognition and CIFAR-10 image recognition tasks, shallow nets can be trained that perform similarly to complex, well-engineered, deeper convolutional architectures. The same model compression trick can also be used to compress impractically large deep models and ensembles of large deep models down to “medium-size” deep models that run more efficiently on servers, and down to “small” models that can run on mobile devices. In machine learning and statistics we used to believe that one of the keys to preventing overfitting was to keep models simple and the n umber of parameters small to force generalization. We no longer believe this — learning appears to generalize best when training models with excess capacity, but the learned functions can often be represented with far fewer parameters. We do not yet know if this is true just of current learning algorithms, or if it is a fundamental property of learning in general.Bio Rich Caruana is a Senior Researcher at Microsoft Research in Redmond (Seattle), Washington. Before joining Microsoft, Rich was on the faculty in the CS Department at Cornell University, at UCLA’s Medical School, and at CMU’s Center for Learning and Discovery (CALD). Rich’s Ph.D. is from Carnegie Mellon University where he worked with Tom Mitchell and Herb Simon. His thesis on Multi-Task Learning helped create a new subfield of machine learning called Transfer Learning. Rich received an NSF CAREER Award in 2004 (for Meta Clustering), best paper awards in 2005 (with Alex Niculescu-Mizil), 2007 (with Daria Sorokina), and in 2014 (with Todd Kulesza, Saleema Amershi, Danyel Fisher and Denis Charles), co-chaired KDD in 2007 (with Xindong Wu), and serves as area chair for the NIPS, ICML, and KDD conferences. His current research focus is on learning for medical decision making, deep learning, adaptive clustering, and computational ecology. |
| Oct 9, 2014 Bren Hall 6011 10:00am |
Geoffrey Hinton Professor University of Toronto and GoogleDark KnowledgeAbstract: A simple way to improve classification performance is to average the predictions of a large ensemble of different classifiers. This is great for winning competitions but requires too much computation at test time for practical applications such as speech recognition. In a widely ignored paper in 2006, Caruana and his collaborators showed that the knowledge in the ensemble could be transferred to a single, efficient model by training the single model to mimic the log probabilities of the ensemble average. This technique works because most of the knowledge in the learned ensemble is in the relative probabilities of extremely improbable wrong answers. For example, the ensemble may give an image of a BMW a probability of one in a billion of being a garbage truck but this is still far greater (in the log domain) than its probability of being a carrot. This “dark knowledge”, which is practically invisible in the class probabilities, defines a similarity metric over the classes that makes it much easier to learn a good classifier.I will describe a new variation of this technique called “distillation” and will show some surprising examples in which good classifiers over all of the classes can be learned from data in which some of the classes are entirely absent, provided the target probabilities come from an ensemble that has been trained on all of the classes. I will also show how this technique can be used to improve a state-of-the-art acoustic model and will discuss its application to learning large sets of specialist models without overfitting. This is joint work with Oriol Vinyals and Jeff Dean. |
Academic Year 2013-2014
The Institute for Genomics and Bioinformatics is pleased to announce its Distinguished Speaker Series for the academic year 2013-2014.
| Feb 25, 2014 Bren Hall 6011 11:00am |
Silvio Micali Ford Professor of Engineering Department of Electrical Engineering and Computer Science Massachusetts Institute of Technology Proof, Secrets, and Computation |