Differences

This shows you the differences between two versions of the page.

--- education [2019/05/24 17:03]
fablpd
+++ education [2019/05/27 12:08]
fablpd
@@ Line 31: / Line 31: @@
   * **Distributed computing using RDMA and/or NVRAM**: contact [[https://people.epfl.ch/igor.zablotchi|Igor Zablotchi]] for more information.
+  * **[[Distributed ML|Distributed Machine Learning]]**: contact [[http://people.epfl.ch/georgios.damaskinos|Georgios Damaskinos]] for more information.
-  * **Robust Distributed Machine Learning**: The goal of such project is to work on the design and implementation of algorithms and systems to improve the robustness of distributed ML schemes that would tolerate poisoned data, software bugs as well as hardware failures. The practical work will be done on TensorFlow or Pytorch. Contact Arsany Guirguis or Sébastien Rouault for more information.
+  * **Robust Distributed Machine Learning**: With the proliferation of big datasets and models, Machine Learning is becoming distributed. Following the standard parameter server model, the learning phase is taken by two categories of machines: parameter servers and workers. Any of these machines could behave arbitrarily (i.e., said Byzantine) affecting the model convergence in the learning phase. Our goal in this project is to build a system that is robust against Byzantine behavior of both parameter server and workers. Our first prototype, AggregaThor(https://www.sysml.cc/doc/2019/54.pdf), describes the first scalable robust Machine Learning framework. It fixed a severe vulnerability in TensorFlow and it showed how to make TensorFlow even faster, while robust. Contact [[https://people.epfl.ch/arsany.guirguis|Arsany Guirguis]] or [[https://people.epfl.ch/sebastien.rouault|Sébastien Rouault]] for more information.
+  * **Stochastic gradient: (artificial) reduction of the ratio variance/norm for adversarial distributed SGD**:
+One computationally-efficient and non-intrusive line of defense for adversarial distributed SGD (e.g. 1 parameter server distributing the gradient estimation to several, possibly adversarial workers) relies on the honest workers to send back gradient estimations with sufficiently low variance; assumption which is sometimes hard to satisfy in practice.
+One solution could be to (drastically) increase the batch-size at the workers, but doing so may as well defeat the very purpose of distributing the computation.
+In this project, we propose two approaches that you can choose to explore (also you may propose a different approach) to (artificially) reduce the ratio variance/norm of the stochastic gradients, while keeping the benefits of the distribution.
+The first proposed approach, speculative, boils down to "intelligent" coordinate selection.
+The second makes use of some kind of "momentum" at the workers.
+[1] "Machine Learning with Adversaries: Byzantine Tolerant Gradient Descent" (https://papers.nips.cc/paper/6617-machine-learning-with-adversaries-byzantine-tolerant-gradient-descent)
+[2] "Federated Learning: Strategies for Improving Communication Efficiency" (https://arxiv.org/abs/1610.05492)
   * **Consistency in global-scale storage systems**: We offer several projects in the context of storage systems, ranging from implementation of social applications (similar to [[http://retwis.redis.io/|Retwis]], or [[https://github.com/share/sharejs|ShareJS]]) to recommender systems, static content storage services (à la [[https://www.usenix.org/legacy/event/osdi10/tech/full_papers/Beaver.pdf|Facebook's Haystack]]), or experimenting with well-known cloud serving benchmarks (such as [[https://github.com/brianfrankcooper/YCSB|YCSB]]); please contact [[http://people.epfl.ch/dragos-adrian.seredinschi|Adrian Seredinschi]] or [[https://people.epfl.ch/karolos.antoniadis|Karolos Antoniadis]]  for further information.