Events

Hide Filters Show Filters
Hide Filters Show Filters

Event Type

Impact Area

Research Area

Seminar Series

Reset Filters

May 25, 2022

No events scheduled

May 26, 2022

Gael Varoquaux: DirtyData: statistical learning on non-curated databases

Gael Varoquaux, Inria
11:00A
-12:00P

Location

32G-889 (Hewlett)
Add to Calendar 2022-05-26 11:00:00 2022-05-26 12:00:00 America/New_York Gael Varoquaux: DirtyData: statistical learning on non-curated databases Abstract: According to industry surveys, the number one hassle of data scientists is cleaning the data to analyze it. On two specific aspects of dirty data, I will show how machine-learning can readily injest data without curation. Specificaly I will cover prediction from missing values and non-normalized entries. The normalization problem can be tackled with character-level modeling to recover latent caterories. The missing-values problem will lead us to revisit classic statistical results in the setting of supervised learning but leads to models that do not require assumptions on the missing-mechanism without Missing At Random requirements.References:* Cerda, Patricio, Gaël Varoquaux, and Balázs Kégl. "Similarity encoding for learning with dirty categorical variables." Machine Learning 107.8 (2018): 1477-1494. https://link.springer.com/article/10.1007/s10994-018-5724-2* Cerda, Patricio, and Gaël Varoquaux. "Encoding high-cardinality string categorical variables." IEEE Transactions on Knowledge and Data Engineering (2020). https://ieeexplore.ieee.org/abstract/document/9086128* Cvetkov-Iliev, Alexis, Alexandre Allauzen, and Gaël Varoquaux. "Analytics on Non-Normalized Data Sources: more Learning, rather than more Cleaning." IEEE Access (2022). https://ieeexplore.ieee.org/abstract/document/9758752* Le Morvan, M., Josse, J., Moreau, T., Scornet, E., & Varoquaux, G. (2020). NeuMiss networks: differentiable programming for supervised learning with missing values. Advances in Neural Information Processing Systems, 33, 5980-5990. https://proceedings.neurips.cc/paper/2020/hash/42ae1544956fbe6e09242e6cd752444c-Abstract.html* Le Morvan, Marine, Julie Josse, Erwan Scornet, and Gaël Varoquaux. "What’sa good imputation to predict with missing values?." Advances in Neural Information Processing Systems 34 (2021) https://proceedings.neurips.cc/paper/2021/hash/5fe8fdc79ce292c39c5f209d734b7206-Abstract.htmlBio:Gaël Varoquaux is a research director working on data science and health at Inria (French Computer Science National research). His research focuses on statistical-learning tools for data science and scientific inference, with an eye on applications in health and social science. He develops tools to make machine learning easier, with statistical models suited for real-life, uncurated data, and software for data science. For example, since 2008, he has been exploring data-intensive approaches to understand brain function and mental health. He co-funded scikit-learn, one of the reference machine-learning toolboxes, and helped build various central tools for data analysis in Python. Varoquaux has a PhD in quantum physics and is a graduate from Ecole Normale Superieure, Paris. 32G-889 (Hewlett)

June 01, 2022

Memory Locality Optimisations for Graph Processing

Hans Vandierendonck
Queen's University Belfast
4:00P
-5:00P

Location

Virtual Event
https://mit.zoom.us/meeting/register/tJMtdOqgrTwsGNV_k0Nk6JjLMk-G6GFzTRyk
Add to Calendar 2022-06-01 16:00:00 2022-06-01 17:00:00 America/New_York Memory Locality Optimisations for Graph Processing ******************IMPORTANT NOTE ABOUT REGISTRATION******************- The registration link for Spring/Summer 2022 is the same as the link from Fall 2021.- Please save the Zoom link that you receive after you register. This link will stay the same for all subsequent Fast Code seminars.- Zoom does not recognize a second registration, and will not send out the link a second time. The organizers will not be notified of any second registration.- If you have any problems with registration, please contact lindalynch@csail.mit.edu by May 31, so that we can try to resolve it before the seminar begins.*********************************************************************Abstract:Graph analytics present a challenging workload for modern computer architectures due to the size of data sets, unpredictable and randomly-looking access patterns and limited opportunities for data reuse in on-chip caches. This talk discusses three application-specific methods for improving the memory locality of graph processing applications. While the three techniques are fundamentally different, they all share the common goal of aiming to reduce the range of data accessed through random access patterns. Hub temporal locality aims to control random access patterns associated to hub vertices, enabling the majority of random access stores to hit in the on-chip CPU caches. Narrow vertex properties enable the cache to store more vertex properties, thereby making the cache appear larger. Finally, LOTUS improves locality by partitioning the graph adjacency matrix and adjusting the storage format for distinct components in an application-aware manner. These results demonstrate the performance potential, and need for, bespoke memory locality optimisations.Bio:Hans Vandierendonck is Professor of High-Performance and Data-Intensive Computing in the school of Electronics, Electrical Engineering and Computer Science, at Queen's University Belfast. He is Director of the Centre for Data Science and Scalable Computing (DSSC) in the Institute for Electronics, Communications and Information Technology. His research interests are in compilers, runtime systems and architectures for parallel systems with special attention to the programmability of such systems. Hans also has a vested interested in computer architecture, and particularly in cache architecture, prediction and performance evaluation. He has co-authored over a 100 papers and has supervised PhD dissertations of 4 students, one of whom, Dr Jiawen Sun, is finalist in the EPSRC Connected Nation Pioneers competition. Hans received the IBM Belgium Prize for Computer Science in 2000 for his graduation thesis and in 2004 for his PhD dissertation. He is a Senior Member of IEEE and ACM. https://mit.zoom.us/meeting/register/tJMtdOqgrTwsGNV_k0Nk6JjLMk-G6GFzTRyk

June 07, 2022

THESIS DEFENSE: Format Abstractions for Compilation of Sparse Tensor Algebra

Stephen Chou
MIT-CSAIL
1:00P
-2:00P

Location

Virtual Event
Add to Calendar 2022-06-07 13:00:00 2022-06-07 14:00:00 America/New_York THESIS DEFENSE: Format Abstractions for Compilation of Sparse Tensor Algebra ABSTRACT: Multidimensional arrays (tensors) are commonly used to represent data in many domains, including data analytics, machine learning, engineering, and the physical sciences. Many highly-optimized libraries and compilers have been developed for efficiently computing on dense tensors. However, existing libraries and compilers are limited in their ability to support many real-world applications that work with sparse tensors, which contain mostly zeros. In particular, there exist countless specialized representations (formats) for storing sparse tensors in memory, each suited to specific types of applications and data. Since different formats can use very different data structures to store nonzeros though, computing with sparse tensors that are stored in different formats can require vastly dissimilar code that are difficult to implement by hand and also non-trivial to automatically generate. Existing libraries and compilers thus only support computing limited sets of operations on sparse tensors stored in limited sets of formats, sacrificing usability and performance as a result. In this thesis, I show how to build a compiler that supports efficiently computing with sparse tensors stored in disparate formats. I first show how a wide range of commonly-used sparse tensor formats, from array-based formats such as CSR, COO, and DIA to pointer-based formats such as linked lists, BSTs, and C-trees, can all be expressed as compositions of per-dimension formats. I further demonstrate how users can define such per-dimension formats by implementing abstractions we developed that precisely capture how different data structures store nonzeros in memory. I then show how, with these implementations at hand, a compiler can generate code to efficiently compute with tensors that are stored in any of the aforementioned---and countless other---formats. We have implemented our technique in the TACO sparse tensor algebra compiler, which is the first compiler to generate code that computes any basic tensor algebra expression with sparse tensors stored in arbitrary formats. Our technique generates code that has performance competitive with equivalent implementations in hand-optimized libraries and frameworks.THESIS COMMITTEE: Saman Amarasinghe, Julian Shun, Alan Edelman https://mit.zoom.us/j/94590727297?pwd=akxmbGhaR3pCcGVnOXcvZ3RuZ0R0Zz09 Contact Mary McDavitt mmcdavit@csail.mit.edu for password

June 20, 2022

Alphy: Social Media Tnat Can Serve a Community

Ted Selker
AlphyCo (formerly MIT Media Lab, CMU West)
3:00P
-5:00P

Location

Seminar Room G449 (Patil/Kiva)
Add to Calendar 2022-06-20 15:00:00 2022-06-20 17:00:00 America/New_York Alphy: Social Media Tnat Can Serve a Community Social media sites are places for people to meet and be. The news and literature are full of stories based on how users of such platforms find themselves self promoting, and taking extreme positions. From the beginning of online communication, people have noted the tendency for people to “flame” (go extreme) on line more than in person. In the physical world we meet people around a context. The context might be a french lesson, a ballroom dancing class, a hike, or sailboat ride. External contexts provide topics of mutual interest defusing the personal differences. How do we do that without furthering divisions in people's social and political experiences?Alphy creates a social media platform for supporting advancing women. Inspiring articles about people's successes, informative text, audio video & paced educational materials, and communication that encourages people to consider content and alphy traits attempt to make a safe and purpose built social media platform. A centerpeice of the approach is Alphysmarts. As well as incenting users to find interesting content, Alphysmarts gives feedback and encouragement for reflecting on things you are writing to others. The work expands on much of the sentiment analysis work done in my context aware and considerate computing labs at the MIT Media Lab and CMU too.---Ted Selker is CTO of Alphyco creating a community based considerate social media for advancing women. Ted is an entrepreneur inventor who also mentors innovation. Ted spent 5 years as director of Considerate Systems research at Carnegie Mellon University Silicon Valley and in developing the campus’s research mission. Prior to that, Ted spent 10 years as an associate Professor at the MIT Media Laboratory where he created the Context Aware Computing group, co-directed the Caltech/MIT Voting Technology Project, and directed the Industrial Design Intelligence future for product design project. Prior to that, his successes at targeted product creation and enhancement at IBM earned him the role of IBM Fellow. As an adjunct professor at Stanford at that time Ted also helped create the People Computers and Design project in their Computer Science Department.Ted's work has birthed successful products ranging from notebook computers to operating systems. It has accumulated numerous awards, patents, and papers and has often been featured in the press. Ted was co-recipient of the Computer Science Policy Leader Award for Scientific American 50, the American Association for People with Disabilities Thomas Paine Award for his work on voting technology and the Telluride Tech fest award. Seminar Room G449 (Patil/Kiva)