Jorge Mendez

I am a Ph.D. student at the Lifelong Machine Learning Group at the University of Pennsylvania, where I am advised by Prof. Eric Eaton. I previously obtained a Master's degree in Robotics from the GRASP Lab at Penn, and got my Bachelor's degree in Electronics Engineering from Universidad Simon Bolivar in Venezuela.

I spent the summer of 2021 as a research intern at FAIR NYC working with Arthur Szlam and Ludovic Denoyer on mixture of experts models. Previously, I spent the summer of 2020 as a remote research intern at MSR Montreal working with Harm van Seijen on lifelong learning of compositional reinforcement learning problems. Before that, I spent the summer of 2019 as a research intern at Facebook with the Conversational AI team, working under Alborz Geramifard and co-advised by Mohammad Ghavamzadeh. I also spent a year at Politecnico di Milano in Italy, where I took graduate courses in AI and Robotics, and spent five months at the Advanced Systems Technology group in STMicroelectronics.

I am on the academic job market this Fall 2021. Please find here my research, teaching, and diversity statements.

Email  /  CV  /  Google Scholar  /  LinkedIn  /  Twitter

Research

I'm interested in the creation of versatile artificially intelligent systems that learn to accumulate knowledge over their lifetimes. I focus on the question of how agents can decompose the complex knowledge required to model a lifelong data stream into simpler units that can be adapted and reused in the future. My work applies these methods to computer vision, robotics, and natural language.

Publications
Conference Papers

Modular Lifelong Reinforcement Learning via Neural Composition
PDF
Jorge A. Mendez, Harm van Seijen, Eric Eaton
To appear in International Conference on Learning Representations (ICLR), 2022

We explore the problem of lifelong RL of functionally compositional knowledge, and develop an algorithm that demonstrates zero-shot and forward transfer, avoidance of forgetting, and backward transfer in discrete 2-D and robotic manipulation domains.

Lifelong Learning of Compositional Structures
PDF  /  Code  /  Talk
Jorge A. Mendez, Eric Eaton
In International Conference on Learning Representations (ICLR), 2021

We study the question of how to learn compositional parameterized structures from an empirical standpoint, and propose a general-purpose framework that can learn with various forms of knowledge representations and base algorithms.

Lifelong Policy Gradient Learning of Factored Policies for Faster Training Without Forgetting
PDF  /  Code
Jorge A. Mendez, Boyu Wang, Eric Eaton
In Neural Information Processing Systems (NeurIPS), 2020

We introduced an algorithm for directly optimizing factored policies via policy gradients in a lifelong learning setting, and showed theoretically and empirically that our approach avoids catastrophic forgetting.

Transfer Learning via Minimizing the Performance Gap Between Domains
PDF  /  Code  /  Poster
Boyu Wang, Jorge A. Mendez, Mingbo Cai, Eric Eaton
In Neural Information Processing Systems (NeurIPS), 2019

We introduced the notion of performance gap as a label-dependent notion of domain discrepancy, and developed an boosting-based algorithm, gapBoost, that exploits the insights from gap minimization.

Lifelong Inverse Reinforcement Learning
PDF  /  Code  /  Video  /  Poster
Jorge A. Mendez, Shashank Shivkumar, Eric Eaton
In Neural Information Processing Systems (NeurIPS), 2018

We introduced the problem of lifelong learning from demonstrations, and created an efficient lifelong inverse reinforcement learning (ELIRL) algorithm.

Wokrshop Papers

Lifelong Learning of Factored Policies via Policy Gradients
PDF  /  Code  /  Video  /  Talk
Jorge A. Mendez, Eric Eaton
In 4th Lifelong Learning Workshop at the International Conference on Machine Learning (LML-ICML), 2019

Best paper award. This work was superseded by the NeurIPS version with theoretical results. We introduced an algorithm for directly optimizing factored policies via policy gradients in a lifelong learning setting, and showed empirically that our approach avoids catastrophic forgetting.

A General Framework for Continual Learning of Compositional Structures
PDF  /  Code  /  Poster
Jorge A. Mendez, Eric Eaton
In Continual Learning Workshop at the International Conference on Machine Learning (CL-ICML), 2020

This work was superseded by the pre-print version with more empirical evidence. We study the question of how to learn compositional parameterized structures from an empirical standpoint, and propose a general-purpose framework that can learn with various forms of knowledge representations and base algorithms.

Reinforcement Learning of Multi-Domain Dialog Policies Via Action Embeddings
PDF  /  Talk
Jorge A. Mendez, Alborz Geramifard, Mohammad Ghavamzadeh, Bing Liu
In 3rd Conversational AI Workshop at Neural Information Processing Systems (ConvAI-NeurIPS), 2019

We developed an architecture for learning multi-domain task oriented dialog policies, based on the notion of action embeddings, which capture domain agnostic representations of how to respond to user's queries.

Teaching
teaching

University of Pennsylvania
CIS 192 Python Programming
Fall '19, Spring '20
Instructor

CIS 419/519 Introduction to Machine Learning
Fall '17
Head Teaching Assistant

Universidad Simon Bolivar
CI 2125 Programming I
Winter '16, Spring '14, Winter '14, Fall '14, Spring '13, Winter '13, Fall '13
Teaching Assistant

EC 2272 Electric Circuit Analysis II
Spring '11
Teaching Assistant

MA 1112 Calculus II
Winter '11
Teaching Assistant

MA 1111 Calculus I
Fall '10
Teaching Assistant


Website template