• Accueil
  • Accueil
  • Accueil
  • Accueil

CNRS

Rechercher




Accueil >

[PDS/HPDA Seminar] 14/1/2022 from 10:00 to 11:00 at 4A312 - Ali Mammadov (reading group) and Huiyuan Li (reading group)

[PDS/HPDA Seminar] 14/1/2022 from 10:00 to 11:00 at 4A312 - Ali Mammadov (reading group) and Huiyuan Li (reading group)

During the PDS/HPDA Seminar of 14/1/2022 from 10:00 to 11:00, Ali Mammadov will present a reading group talk and Huiyuan Li will present a reading group talk.

Visio : https://webconf.imt.fr/frontend/fra-vcg-byn-fxd

Location : 4A312

# Reading group : Lost in Pruning : The Effects of Pruning Neural Networks beyond Test Accuracy (MLSys’21)\n\nPresented by Ali Mammadov on 14/1/2022 at 10:00. Attending this presentation is mandatory for the master students.

Paper : https://proceedings.mlsys.org/paper/2021/file/2a79ea27c279e471f4d180b08d62b00a-Paper.pdf

Full post : https://www.inf.telecom-sudparis.eu/pds/seminars_cpt/555/

## Abstract
Neural network pruning is a popular technique used to reduce the inference costs of modern, potentially overparameterized, networks. Starting from a pre-trained network, the process is as follows : remove redundant parameters, retrain, and repeat while maintaining the same test accuracy. The result is a model that is a fraction of the size of the original with comparable predictive performance (test accuracy). Here, we reassess and evaluate whether the use of test accuracy alone in the terminating condition is sufficient to ensure that the resulting model performs well across a wide spectrum of "harder" metrics such as generalization to out-of-distribution data and resilience to noise. Across evaluations on varying architectures and data sets, we find that pruned networks effectively approximate the unpruned model, however, the prune ratio at which pruned networks achieve commensurate performance varies significantly across tasks. These results call into question the extent of genuine overparameterization in deep learning and raise concerns about the practicability of deploying pruned networks, specifically in the context of safety-critical systems, unless they are widely evaluated beyond test accuracy to reliably predict their performance. Our code is available at https://github.com/lucaslie/torchprune.

# Reading group : AUSOM : Autonomic Service-Oriented Middleware for IoT-Based Systems (IEEE World Congress Service)\n\nPresented by Huiyuan Li on 14/1/2022 at 10:30. Attending this presentation is mandatory for the master students.

Paper : https://www.computer.org/csdl/proceedings-article/services/2017/2002a102/12OmNwlqhRI

Full post : https://www.inf.telecom-sudparis.eu/pds/seminars_cpt/ausom-autonomic-service-oriented-middleware-for-iot-based-systems/

## Abstract
Service-oriented Architecture (SOA) has been recognized as a key technology for operating IoT-based systems, by abstracting sensing and actuation capabilities of IoT resources via services or microservices. However, such systems being dynamic in nature, must be proactive in responding to changing circumstances, hence they need to be autonomic in nature. Therefore this requires the design and deployment of an autonomic serviceoriented middleware that can mediate interactions, and control sensing & actuation, within the IoT-based system. To that end, in this paper, we present our vision of an autonomic service-oriented middleware AUSOM (pronounced "awesome") for IoT based systems. The key features of AUSOM are : incorporation of the well-known MAPE-K loop (Monitor, Analyze, Plan, Act, using stored Knowledge) from autonomic computing for proactive adaptation, incorporation of a multilayered context model, and using contextual information to facilitate adaptation at the IoT device (sensor and actuator) layer. We present the architecture of AUSOM and also illustrate how it would function via a simple yet realistic example.

See you soon,
The PDS/HPDA Seminar organizing committee

===========================
If you don’t want to receive the PDS/HPDA Seminar posts anymore, please contact francois.trahay@telecom-sudparis.eu