My name is Ben, and I am currently a computer science student at Northeastern University interested in generative models.

Projects

Equivariant Proximal Policy Optimization With Behavioral Cloning

Equivariance has been shown to increase sample efficiency in many different reinforcement learning algorithms. These models are particularly relevant for classic control and robotic manipulation learning problems, where state spaces can be thought of as symmetric under rotation. My project builds on previous work, examining the model architecture for equivariant actor-critic methods, and how symmetry can burnish the Proximal Policy Optimization (PPO) algorithm.

A snippet of the environment

"Better Together", Large Graph Embeddings with Scalable representation Learning

Last summer, I worked on Ken Church's team at the JSALT Speech and NLP workshop hosted by Johns Hopkins. Our overall aim was to build an academic search engine for papers and authors in Semantic Scholar (S2). My focus was on large-scale graph embeddings, implementing and refining traditional linear algebra methods and graph neural networks to produce embedding files on CPU and GPU. I focused on the PRONe algorithm (Zhang et al., 2019), which utilizes spectral clustering, Chebyshev iterations and Fourier transforms to produce embeddings. I worked in python and C, toying with the low-level linear algebra libraries to optimize compute efficiency on our limited hardware.

Search engine computing related works for Attention is All You Need

Publications

MEANT: Multimodal Encoder for Antecedent Information

Benjamin Irving, Annika Marie Schoene

EMNLP 2024

We introduce MEANT, a multimodal model architecture with a novel, temporally focused self-attention mechanism. The model effectively processes stock market data across multiple modalities - price information, social media text, and graphical data. Our research demonstrates that MEANT improves performance on existing baselines by over 15%, with textual information showing significantly more impact than visual information in time-dependent tasks.

Additionally, we release TempStock, a new dataset containing 1.7M+ Tweets and price information from S&P 500 companies, specifically designed for sequential processing across varying lag periods.

Related Work is All you Need

Rodolfo Zevallos, John E. Ortega, Benjamin Irving

LREC-COLING 2024

In modern times, generational artificial intelligence is used in several industries and by many people. One use case that can be considered important but somewhat redundant is the act of searching for related work and other references to cite. As an avenue to better ascertain the value of citations and their corresponding locations, we focus on the common "related work" section as a focus of experimentation with the overall objective to generate the section.

In this article, we present a corpus with 400k annotations that distinguish related work from the rest of the references. Additionally, we show that for the papers in our experiments, the related work section represents the paper just as well, and in many cases, better than the rest of the references. We show that this is the case for more than 74% of the articles when using cosine similarity to measure the distance between two common graph neural network algorithms: Prone and Specter.

Published in: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 13874-13878, 20-25 May, 2024, Torino, Italia.