Menu
INFORMS 2018 Annual Meeting - Phoenix logo
INFORMS Annual Meeting, November 4-7, 2018
Search
  • Registration
    • Register Now!
    • Visa Support Information
  • Program
    • Schedule
    • Quick Reference Guide
    • Plenaries & Keynotes
    • Call for Papers
    • Abstract Submission
    • Poster Sessions
    • Pre-Meeting Events
    • TutORials
    • Technology Tutorials
    • Speaker Guidelines
  • Events
    • Get Involved
    • Receptions
    • New Member Breakfast
  • Career Fair
  • Exhibitors & Sponsors
    • Exhibitors & Sponsors
    • Exhibiting Opportunities
    • Sponsorship Opportunities
  • Hotel & Travel
  • Info
    • For Students
    • For Practitioners
    • Childcare
    • Conference Media Kits
    • Organizing Committee
    • Sponsored Cluster Chairs
    • Emerging Topics Chairs
Close Menu
November 5, 2018

Joint Session Integer Programming/O.R. Frontiers: Machine Learning and Discrete Optimization II

INFORMS e-News

By Zulqarnain Haider

The marriage between machine learning and mixed integer programming (MIP) has recently been the talk of the town. Researchers around the world are imagining new ways of combining and merging the two powerful fields of machine learning and MIP. The joint session on machine learning and discrete optimization was second in a three-part series spanning multiple days and many speakers. Soon after the session started, the room was filled to capacity. It started with a talk by Anirbit Mukherjee who, dressed in traditional Indian garb, described his recent research in trying to understand the underlying mathematics of neural networks. In a wide-ranging survey of the field, the speaker described his work formalizing the connections between neural networks and sparse coding.

The second speaker, Elias B. Khalil from Georgia Tech, presented his work on machine learning security using integer programming. He introduced binarized neural networks (BNN) and an integer programming-based approach to attack or perturb the BNN to make sure that it predicts incorrectly. By identifying the vulnerabilities of machine learning models through adversarial examples, one could make them robust to “attacks,” an idea distilled in the maxim “attack to protect.” The speaker concluded his talk by asserting that the presented approach, called iProp compares favorably with other state of the art attack algorithms like FGSM.

The third speaker, Giulia Zarpellon from Polytechnique Montréal, spoke about learning new approaches to branch and bound (B&B) more efficiently. She suggested a three-pronged reinforcement learning framework for better defining the state of the B&B tree, the branching policy, and the rewards. She suggested to take context and temporality into account in variable branching. She concluded by describing the computational experiments with 54 heterogenous instances.

Thiago Serra (pictured), from Carnegie Mellon University was the fourth speaker and the session chair. In a fast-paced talk, he presented his results about the number of linear regions that piecewise linear functions represented by neural networks can attain, both theoretically and empirically. This number could be used to compare different, but similar, neural network configurations. The talk mainly focused on coming up with better bounds on the number of linear regions including a method for exact enumeration of the linear regions by modeling the DNN as a mixed-integer program.

The final speaker of the session, Ellen Vitercik (Carnegie Mellon), talked about algorithm design as a learning problem. She took the case of clustering problem, a special case of combinatorial partitioning problem. The aim of the study was to find an algorithm that is best over a sample of algorithms in a given domain and doing so on firm theoretical foundations.

Plenary: Networks to Save the World: O.R. in Action MCDM Junior Researcher Best Paper Award Finalists

Related Posts

News

Aggregation of Preference Data

By Yeawon Yoo The session titled, “Aggregation of Preference Data” was held Tuesday afternoon – the last day of the INFORMS Annual Meeting. Even though the session was held at the last time slot, the room was full of many audience members and discussions. The session chair, Adolfo Escobedo from Arizona State University, opened the session […]

News

Energy Market Disruption and Transition Through Advanced Analytics

In his keynote presentation, “Energy Market Distribution and Transition Through Advanced Analytics,” Jorge Calzada of National Grid in Boston, MA, discussed applications of advanced analytics in the energy industry. The energy industry, one of the last industries to adopt advanced analytics, is currently in the midst of a great transition, after a very stable period […]

News

Job search strategies for academic positions

By Yeawon Yoo The INFORMS Minority Issues Forum provided a session titled, “Job search strategies for academic positions” chaired by Eduardo Perez from Texas State University. The following is the list of panelists: Mingzhou Jin (Professor of the Department of Industrial & Systems Engineering at the University of Tennessee-Knoxville), Clara Novoa (Associate professor at Texas […]

INFORMS® Online - Institute for Operations Research and the Management Sciences

The Institute for Operations Research and the Management Sciences

5521 Research Park Drive, Suite 200
Catonsville, MD 21228 USA

phone 1 443-757-3500

phone 2 800-4INFORMS (800-446-3676)

fax 443-757-3515

email

Get the Latest Updates

  • Discover INFORMS
  • Explore OR & Analytics
  • Get Involved
  • Impact
  • Join Us
  • Recognizing Excellence
  • Professional Development
  • Resource Center
  • Meetings & Conferences
  • Publications
  • About INFORMS
  • PubsOnLine
  • 2021 Analytics Conference
  • Certified Analytics Professional
  • Career Center
  • INFORMS Connect
Copyright INFORMS. All Rights Reserved
  • Terms of Use
  • |
  • Privacy Policy
  • |
  • Contact INFORMS
Follow INFORMS on: