Marc van Zee

Google Research, Brain team, Amsterdam

Research

Summary

I am currently working as a Software Engineer at Google Research, in the Brain team in Amsterdam, the Netherlands. My overall research interest lies in building systems that can do commonsense reasoning, by applying techniques/ideas from formal knowledge representation to machine learning architectures. I studied Industrial Design (BSc.) and Artificial Intelligence (MSC., cum laude), and did a PhD on high-level decision-making in large enterprises (graduated with highest distinction). I've been working at Google since 2017. First I worked on building a prototype for a next-generation Assistant. In 2019, I helped building the Compositional Freebase Questions dataset (see my Google Research Blog post). Since 2019, I'm also a developer for Flax, a new neural network library built on JAX.


↑ Back to Top

Publications

2020
Intention as Commitment toward Time (Marc van Zee, Dragan Doder, Leendert van der Torre, Mehdi Dastani, Thomas Icard, Eric Pacuit), In Artificial Intelligence, Elsevier BV, volume 283, 2020. [doi]
Measuring Compositional Generalization (Marc van Zeee), In Google AI Blog, 2020.
2019
Measuring Compositional Generalization: A Comprehensive Method on Realistic Data (Daniel Keysers, Nathanael Schärli, Nathan Scales, Hylke Buisman, Daniel Furrer, Sergii Kashubin, Nikola Momchev, Danila Sinopalnikov, Lukasz Stafiniak, Tibor Tihon, Dmitry Tsarkov, Xiao Wang, Marc van Zee, Olivier Bousquet), In International Conference on Learning Representations (ICLR), 2019.
2018
Beyond the Hype: Making Progress on Natural Language Systems (Marc van Zee), In Leon50 - Arguing about Logic: Debates in Individual and Collective Reasoning, 2018.
2017
Rational Enterprise Architecture (Leendert van der Torre, Marc Van Zee), In International Conference on Industrial, Engineering and Other Applications of Applied Intelligent Systems, 2017.
2016
RationalGRL: A Framework for Rationalizing Goal Models Using Argument Diagrams (Marc van Zee, Diana Marosin, Sepideh Ghanavati, Floris Bex), In Proceedings of the 35th International Conference on Conceptual Modeling (ER'2016), 2016.
The RationalGRL toolset for Goal Models and Argument Diagrams (Marc van Zee, Diana Marosin, Floris Bex, Sepideh Ghanavati), In Proceedings of the 6th International Conference on Computational Models of Argument (COMMA'16), Demo abstract, 2016.
AGM-Style Revision of Beliefs and Intentions (Marc van Zee, Dragan Doder), In Proceedings of the 22nd European Conference on Artificial Intelligence (ECAI'16), 2016.
Formalizing and Modeling Enterprise Architecture (EA) Principles with Goal-oriented Requirements Language (GRL) (Diana Marosin, Marc van Zee, Sepideh Ghanavati), In Proceedings of the 28th International Conference on Advanced Information System Engineering (CAiSE16), 2016.
AGM-Style Revision of Beliefs and Intentions from a Database Perspective (Preliminary Version) (Marc van Zee, Dragan Doder), In Proceedings of the 16th International Workshop on Non-Monotonic Reasoning (NMR'16), 2016.
2015
Intention Reconsideration as Metareasoning (Marc van Zee, Thomas Icard), In Bounded Optimality and Rational Metareasoning NIPS 2015 Workshop, 2015.
ARMED: ARgumentation Mining and reasoning about Enterprise architecture Decisions (Marc van Zee, Dirk van der Linden), In Proceedings of the 27th Benelux Conference on Artificial Intelligence (BNAIC2015), 2015.
Insights from a Study on Decision Making in Enterprise Architecture (Dirk van der Linden, Marc van Zee), In Proceedings of the 8th IFIP WG 8.1 Working Conference on the Practice of Enterprise Modeling (PoEM), 2015.
Rationalization of Goal Models in GRL using Formal Argumentation (Marc van Zee, Floris Bex, Sepideh Ghanavati), In Proceedings of the Requirements Engineering Conference 2015 (RE'15), RE: Next! track, 2015.
Rational Architecture = Architecture from a Recommender Perspective (Marc van Zee), In Proceedings of the International Joint Conference on Artificial Intelligence, 2015.
AGM Revision of Beliefs about Action and Time (Marc van Zee, Mehdi Dastani, Dragan Doder, Leendert van der Torre), In Proceedings of the International Joint Conference on Artificial Intelligence, 2015.
Consistency Conditions for Beliefs and Intentions (Marc van Zee, Mehdi Dastani, Dragan Doder, Leendert van der Torre), In Twelfth International Symposium on Logical Formalizations of Commonsense Reasoning, 2015.
2014
Bridging Social Network Analysis and Judgment Aggregation (Silvano Colombo Tosatto, Marc van Zee), In Proceedings of the 6th International Conference on Social Informatics., 2014.
Capturing Evidence and Rationales with Requirements Engineering and Argumentation-Based Techniques (Marc van Zee, Sepideh Ghanavati), In Proceedings of the 26th Benelux Conference on Artificial Intelligence (BNAIC2014), 2014.
Collective Intention Revision from a Database Perspective (Marc van Zee, Mehdi Dastani, Yoav Shoham, Leendert van der Torre), In Collective Intentionality Conference, 2014.
Encoding Definitional Fragments of Temporal Action Logic into Logic Programming (Marc van Zee, Patrick Doherty, John-Jules Meyer), In International Workshop on Defeasible and Ampliative Reasoning (DARe), 2014.
On the Semantic Feature Structure of Modeling Concepts: an Empirical Study (Dirk van der Linden, Marc van Zee), In 16h IEEE Conference on Business Informatics (CBI), 2014.
Formalizing Enterprise Architecture Decision Models using Integrity Constraints (Marc van Zee, Georgios Plataniotis, Diana Marosin, Dirk van der Linden), In 16h IEEE Conference on Business Informatics (CBI), 2014.
Reasoning on Robot Knowledge from Discrete and Asynchronous Observations (Pouyan Ziafati, Yehia Elrakaiby, Marc van Zee, Leendert van der Torre, Holger Voos, Mehdi Dastani, John-Jules Meyer), In Knowledge Representation and Reasoning in Robotics, 2014.
Social Network Analysis for Judgment Aggregation (Silvano Colombo Tosatto, Marc van Zee), In 13th International Conference on Autonomous Agents and Multiagent Systems (AAMAS2014), 2014.
2013
Argument Revival in Annotated Argumentation Networks (Diego Agustin Ambrossio, Alessio Antonini, Yehia Elrakaiby, Dov Gabbay, Marc van Zee), In Second workshop on Argumentation in Artificial Intelligence and Philosophy: computational and philosophical perspectives (ARGAIP-13), 2013.
Multi-Cycle Query Caching in Agent Programming (Natasha Alechina, Tristan Behrens, Mehdi Dastani, Koen Hindriks, Koen Hubner, Fred Jomi, Brian Logan, Hai H. Nguyen, Marc van Zee), In Twenty-Seventh AAAI Conference on Artificial Intelligence (AAAI-13), 2013.
Belief Caching in 2APL (Mehdi Dastani, Marc van Zee), In The workshop on Engineering Multi-Agent Systems (EMAS), 2013.
Implementing Temporal Action Logics using Logic Programming and SMT Solving (Marc van Zee), Master's thesis, Utrecht University, 2013.
↑ Back to Top

Teaching

2016
  • Assistant Selected Topics in A.I. by Emil Weydert, University of Luxembourg
  • Assistant Intelligent Agents 1 by Leendert van der Torre, University of Luxembourg
2015
  • Assistant Selected Topics in A.I. by Emil Weydert, University of Luxembourg
  • Assistant Intelligent Agents 1 by Leendert van der Torre, University of Luxembourg
2014
  • Assistant Selected Topics in A.I. by Emil Weydert, University of Luxembourg
  • Assistant Intelligent Agents 1 by Leendert van der Torre, University of Luxembourg
2013
  • Assistant Selected Topics in A.I. by Emil Weydert, University of Luxembourg
2012
  • Assistant Introductie Adaptieve Systemen (Introduction to Adaptive System), by Gerard Vreeswijk, Utrecht University
  • Assistant Introductie Taalkunde (Introduction to Linguistics), by Y. Winter and A. Chernilovskaya, Utrecht University
↑ Back to Top

Program Committee (PC) Member

↑ Back to Top

Other Academic Activities

↑ Back to Top

Open-source projects

Note This list is probably outdated. View all projects with descriptions on Github.


  • Flax: a neural network library

    Flax is a high-performance neural network library for JAX that is designed for flexibility: Try new forms of training by forking an example and by modifying the training loop, not by adding features to a framework. I've been working on various things in this project, including writing the first seq2seq LSTM example, and setting up the HOWTO system.

  • Go

    SAT Solver

    A simple implementation of a SAT solver that can use either a recursive or an iterative algorithm. It is based on Knuth's SAT0W watchlist-based backtracking algorithm. The iterative code follows Knuth's version much closer, but is a bit more complicated. The recursive version is more untuitive and easier to understand.

    Developed for fun.

  • Javascript

    Using a Neuroheadset as a Robot-Controlling Device

    The Emotiv Epoc is a neuroheadset that promises to be able to read one's mind using EEG measurements. When you wear it, it can separate 16 different thoughts, all based on a direction or a movement (for example: left, rotate forward, push, etc.). I developed a robot using digital production techniques and connecting the EPOC to it. By connecting two actions (push and pull) to two directions of the robot (forward and backward) using Javascript, I am able to control with my mind after a training period of about a month.

    Developed during an internship at Fablab, Amsterdam. Video of me controlling the robot

  • 2APL

    Starcraft AI for teamwork

    The main goal of this project to design (an aspect of) the AI in the game Starcraft, Broodwar and in this way enhance the gameplay for the human user. We have chosen to research cooperation between multiple AI players within an Real-time Strategy game. We have used the Multi-Agent System (MAS) paradigm to steer (small groups of) units in the game, such that they were able to outperform the default A.I. of the game in situations where teamwork was crucial.

    Developed together with Frank van Meeuwen and Roemer Vlasveld.

  • Java

    Random Project

    We started this project with a question: What is the essence of an object? By breaking down the essence, you create boundaries, or rules that can be used in an algorithm that generates objects. This idea is tested on a chair. We define a chair by pointing out ten points and their relations with each other. We add mass to the chair, and have developed algortihms that translate these chairs to STL format, so that models can be sent to a 3D printer. Jesse continued with this project and built actual chairs based on them.

    With Jesse Kirschner from Kirschner3D

  • Java

    Pursuer Evader Tracking game using a Self-Organising Map

    In the Pursuer Evader Tracking (PET) game there are two players: a pursuer and an evader. The evader tries to get away from the pursuer, while the pursuer has the task to catch it. The game is played on an infinitely large, two-dimensional field, where the agents have a constant startposition, but a random starting direction. The pursuer is moving fast, while the evader is able to turn faster. In this simulation I developed an A.I. that controls the behavior of the evader using a Self-Organizing Map (SOM).

    I developed this final assignment as Teaching Assistant for the course Introduction Adaptive Systems by Dr. Gerard Vreeswijk

  • Java

    Handwritten Digit Recognition using a Neural Network

    The MNIST database (Mixed National Institute of Standards and Technology database) is a large database of handwritten digits. I develop a neural network that is able to recognize these handwritten digits after a period of training. It is a three-layers feedforward neural network with backpropagation. The 256 input nodes correspond to the pixels of an image, and the 10 output layers correspond to the digits.

    Developed as part of Teaching Assistant task: student assignment for the course Introduction Adaptive Systems by Dr. Gerard Vreeswijk

  • Java

    Learning Optimal Plan Revision Strategies

    In order to specify situated agents - artificial systems that display effective, rational behavior in dynamic and often unpredictable environments - one of the key issues is to ensure that the agent responds to changes both appropriately and timely. There must be a rational balance between reasoning and acting. We develop a testbed consisting of an agent situated on a Markov Decision Process, By observing features of the environment, the agent is able to learn appropriate plan revision strategies through reinforcement learning.

    Intention Reconsideration as Metareasoning (Marc van Zee, Thomas Icard), In Bounded Optimality and Rational Metareasoning NIPS 2015 Workshop (BORM2015), 2015.

  • Netlogo

    Flow Control in a Decentralized Network using a Cellular Automaton

    This is a simulation of a cellular automata that models adaptive transport. White particles are transported over a green road, which is protected by brown walls. When the particles collide, they will die and a new road is created, perpendicular to the original one. This results in a cross-section. When a different particle reaches this crossing, it will become the flow controller of that crossing. Flow controllers have a red colour and decide whether particles are allowed to go through.

    For the course Introduction Adaptive Systems by Dr. Gerard Vreeswijk

  • Netlogo

    Optimizing Moonlanding using Genetic Programming

    This is a simulation that tries to land a moonlander as good as possible using genetic programming. The optimalisation goal is to let the moonlander land as quick as possible with as much remaining fuel as possible. I have divided this goal into two sub-goals: (1) Landing the moonlander as quick as possible; (2) Landing the moonlander with as much remaining fuel as possible. I have first tried to optimize these subgoals as much as possible, and then combined the solutions to find an optimal combination.

    For the course Introduction Adaptive Systems by Dr. Gerard Vreeswijk