Key point. Elementera's founders have published peer-reviewed research at ICLR, ICML, AAAI, EACL, NeurIPS and in journals including IEEE Access. The papers cover how learning systems generate audio, adapt to new tasks, decide, and pick up shortcuts from their data. Every paper below links to the published record.
An AI answer engine is a learning system that decides which sources to read and which companies to name. How such systems learn, decide and can be measured carefully are the questions our founders published on for years before Elementera was founded in 2024. This page lists that work in one place, with the co-authors exactly as published, so anyone can check it.
Each entry has one plain-language line. Where we have written a longer explainer, it links to the matching case study.
Which papers has Soroush Mehri published?
Soroush Mehri, Co-Founder and Chief AI Officer, did his research at Mila and Université de Montréal and then at Microsoft Research. His papers cover neural audio generation, fast adaptation, reinforcement learning and the reliability of language models. His preprints are on his arXiv author page.
| Year | Paper | Venue | What it is about |
|---|---|---|---|
| 2017 | SampleRNN: An Unconditional End-to-End Neural Audio Generation Model. Soroush Mehri, Kundan Kumar, Ishaan Gulrajani, Rithesh Kumar, Shubham Jain, Jose Sotelo, Aaron Courville, Yoshua Bengio. First author. | ICLR 2017 | A model that generates raw audio one sample at a time, for music, speech and sound. arXiv. Case study |
| 2017 | Char2Wav: End-to-End Speech Synthesis. Jose Sotelo, Soroush Mehri, Kundan Kumar, João Felipe Santos, Kyle Kastner, Aaron Courville, Yoshua Bengio. | ICLR 2017, Workshop Track | One neural model that turns written text straight into speech audio. Case study |
| 2018 | Deep Complex Networks. Chiheb Trabelsi, Olexa Bilaniuk, Ying Zhang, Dmitriy Serdyuk, Sandeep Subramanian, João Felipe Santos, Soroush Mehri, Negar Rostamzadeh, Yoshua Bengio, Christopher J Pal. | ICLR 2018 | Building blocks that let neural networks work with complex numbers, which suit audio and signal data. arXiv. Case study coming soon |
| 2018 | Rapid Adaptation with Conditionally Shifted Neurons. Tsendsuren Munkhdalai, Xingdi Yuan, Soroush Mehri, Adam Trischler. | ICML 2018 | A way for a model to adapt to a new task on the fly from a few examples. arXiv. Case study |
| 2021 | Successor Feature Sets: Generalizing Successor Representations Across Policies. Kianté Brantley, Soroush Mehri, Geoff J. Gordon. | AAAI 2021 | Helps reinforcement learning agents carry what they learned over to new goals and rewards. arXiv. Case study |
| 2021 | Increasing Robustness to Spurious Correlations using Forgettable Examples. Yadollah Yaghoobzadeh, Soroush Mehri, Remi Tachet des Combes, T. J. Hazen, Alessandro Sordoni. | EACL 2021 | Uses the examples a model finds hard to learn to make language models rely less on shortcuts. arXiv. Case study |
| 2021 | Benchmarking Bias Mitigation Algorithms in Representation Learning through Fairness Metrics. Charan Reddy, Deepak Sharma, Soroush Mehri, Adriana Romero Soriano, Samira Shabanian, Sina Honari. | NeurIPS 2021, Datasets and Benchmarks Track | Compares methods that reduce bias in deep learning on shared data and shared fairness metrics. OpenReview. Case study |
The forgettable-examples paper is the closest to our work today. It studies when a language model answers from a shortcut it learned in training instead of from the evidence in front of it, which is one of the reasons an AI answer can describe a company wrongly.
Which papers has Dr. Maziar Gomrokchi published?
Dr. Maziar Gomrokchi, Co-Founder and CEO, did his PhD at McGill University and Mila. His papers cover reinforcement learning, privacy in machine learning, and how reliably results can be reproduced. His earlier papers, written at Concordia University in Montreal, include work on trust and reputation between software agents. His full record is on Google Scholar.
| Year | Paper | Venue | What it is about |
|---|---|---|---|
| 2016 | Differentially Private Policy Evaluation. Borja Balle, Maziar Gomrokchi, Doina Precup. | ICML 2016 | How to evaluate a decision-making policy from data without exposing the individual records it learned from. arXiv |
| 2017 | Reproducibility of Benchmarked Deep Reinforcement Learning Tasks for Continuous Control. Riashat Islam, Peter Henderson, Maziar Gomrokchi, Doina Precup. | Reproducibility in Machine Learning Workshop, ICML 2017 | Shows how settings and run-to-run variation change reported results, and gives guidance for consistent comparisons. |
| 2021 | Locally Persistent Exploration in Continuous Control Tasks with Sparse Rewards. Susan Amin, Maziar Gomrokchi, Hossein Aboutalebi, Harsh Satija, Doina Precup. | ICML 2021 | A way for learning agents to explore more usefully when feedback is rare. arXiv |
| 2023 | Membership Inference Attacks Against Temporally Correlated Data in Deep Reinforcement Learning. Maziar Gomrokchi, Susan Amin, Hossein Aboutalebi, Alexander Wong, Doina Precup. | IEEE Access, 2023 | Tests whether an attacker can tell which data a trained agent learned from. arXiv |
The reproducibility paper's concern, that one run is not a result, is the same one behind how we measure AI engines today: in our research measurements we ask every question at least three times, and repeat it on another day before we believe a change.
Which papers has Dr. Mehrdad Mehrbod published?
Dr. Mehrdad Mehrbod (Dr. Med), Co-Founder and Chief Product Officer, earned his PhD at UC Berkeley, where his research used computer models of molecules to study how cells attach and signal. He is also the first-listed inventor on US patent 10,508,305 B2.
| Year | Paper | Venue | What it is about |
|---|---|---|---|
| 2013 | Localized Lipid Packing of Transmembrane Domains Impedes Integrin Clustering. Mehrdad Mehrbod, Mohammad R. K. Mofrad. | PLOS Computational Biology | A computer simulation of how the membrane around a cell receptor affects whether receptors group together. |
| 2013 | On the Activation of Integrin αIIbβ3: Outside-in and Inside-out Pathways. Mehrdad Mehrbod, Stephen Trisno, Mohammad R. K. Mofrad. | Biophysical Journal | Simulations of the two routes by which a cell receptor switches on. |
Which papers has Mojtaba Ebadi published?
Mojtaba Ebadi, Co-Founder and Chief Operating Officer, is a senior engineering leader with more than 15 years of experience. Early in his career he co-authored research on automated trading agents with Dr. Maziar Gomrokchi.
| Year | Paper | Venue | What it is about |
|---|---|---|---|
| 2008 | Evaluation of PersianCAT Agent's Accepting Policy in Continuous Double Auction, Participant in CAT 2007 Competition. Sina Honari, Amin Fos-hati, Mojtaba Ebadi, Maziar Gomrokchi. | 13th CSI Computer Conference (CSICC 2008), Communications in Computer and Information Science, Springer | Tests a rule for which buy and sell offers an automated market should accept, using an agent entered in the CAT 2007 market design competition. |
Why does a company working on AI search list research papers?
Because peer review is independent evidence of expertise that we cannot write ourselves. An AI answer engine is a learning system, and the founders' work on how such systems learn, decide and can be measured is the groundwork for measuring AI engines today.
The same habits carry into how we work for clients: we ask each question more than once, we never pool engines, we publish a finding only with its scope, and we record which model answered. You can read more on our company page.
Work with us
If you want the AEO work done for you, our managed AEO services follow the same research habits: Book a call. If you would rather see where you stand first, Lenz gives you a free analysis of how ChatGPT and Gemini answer your buyers' questions: Get your free AI visibility report.