Arjun Karuvally

PhD. Candidate

BiNDS Laboratory, UMASS

Hello and welcome! I’m a scientist with a curiosity for unraveling the mysteries that link the human mind and artificial intelligence. But let’s not get ahead of ourselves; allow me to walk you through my journey—a tale of math, code, and a quest for understanding that goes beyond mere equations.

The Allure of Many Fields Since the start of my academic journey, I have been enamored by a multitude of disciplines—ranging from the intricacies of human intelligence to the abstract beauty of mathematical models and the thought-provoking landscapes of scientific philosophy.

Memory Models After diving headfirst into artificial intelligence, I found my interest in memory models. In my opinion, memory can potentially decode how intelligence—both artificial and human—functions. I believe that memory isn’t just a storehouse of information; it’s the foundational bedrock upon which intelligence is built. Understanding memory in its most fundamental manner unlocks the secrets of cognition and intelligent behavior.

Current Focus: Beyond The Black Box Now, if you’ve ever delved into artificial neural networks, you’ve likely heard them referred to as ‘black boxes,’ mysterious entities that perform complex calculations yet resist easy explanation. My current research aims to shed light into them. I focus on what is called mechanistic interpretability, which is a fancy way of saying I want to know the “how” and the “why” behind these computational models, not just the “what.”

I am trying to uncover universal computations—common threads in various models that may be integral to their functioning. Think of it this way: A car mechanic doesn’t merely know how to change the oil; they understand the entire engine, inside and out. That’s the level of comprehension I aim for with neural networks. If we’re going to trust these systems with increasingly complex tasks, we should understand their inner workings to the same degree that a mechanic understands a car engine.

The Horizon: A Mechanic for The Mind? As we gaze into the future, I envision a world where artificial neural networks and human cognition are not mystical phenomena, but well-understood systems. Systems that can be diagnosed, tuned, and enhanced, much like the car at the mechanic’s shop. As we draw back the curtains on the “black boxes,” my hope is that this clarity propels us toward more efficient, ethical, and insightful applications that benefit us all.

Feel free to explore the site to obtain insight into my projects, publications, and ongoing research. If you have any comments, thoughts or queries, you can reach me at arjun.k018@gmail.com.

news

Oct 4, 2023	New preprint released! We applied the Episodic Memory Theory to mechanistically intepret RNNs. We fully describe the behavior of RNNs trained on simple tasks and provide a method to interpret the learned parameters and hidden states. Check it out at https://arxiv.org/abs/2310.02430
Jul 28, 2023	Presenting Episodic Memory Theory of Recurrent Neural Networks at TAG-ML workshop in ICML 2023, Hawaii. Check-in for an approach to mechanistically interpret RNNs using ideas from the Memory Modeling literature.
Jul 21, 2023	Presenting General Sequential Episodic Memory Model at ICML 2023, Hawaii. Looking forward to meet all the memory researchers there.
Jul 3, 2023	Presentation and discussions at the Institute for Machine Learning, Johannes Kepler University on GSEMM and EMT. Thanks to Günter, Daniel, Sebastian and the others for a wonderful summer of research.
Nov 21, 2021	Check out our new preprint Energy-based General Sequential Episodic Memory Networks at the Adiabatic Limit for a new memory modeling paradigm where the energy surface is not stationary but have temporal characteristics.

latest posts

Oct 18, 2023	Fractals
Sep 14, 2023	Memory and the Energy Paradigm

selected publications

2023

General Sequential Episodic Memory Model

Arjun Karuvally, Terrence J. Sejnowski, and Hava T. Siegelmann

In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, 2023

Abs PDF

The state-of-the-art memory model is the General Associative Memory Model, a generalization of the classical Hopfield network. Like its ancestor, the general associative memory has a well-defined state-dependant energy surface, and its memories correlate with its fixed points. This is unlike human memories, which are commonly sequential rather than separated fixed points. In this paper, we introduce a class of General Sequential Episodic Memory Models (GSEMM) that, in the adiabatic limit, exhibit a dynamic energy surface, leading to a series of meta-stable states capable of encoding memory sequences. A multiple-timescale architecture enables the dynamic nature of the energy surface with newly introduced asymmetric synapses and signal propagation delays. We demonstrate its dense capacity under polynomial activation functions. GSEMM combines separate memories, short and long sequential episodic memories, under a unified theoretical framework, demonstrating how energy-based memory modeling can provide richer, human-like episodes.
Episodic Memory Theory of Recurrent Neural Networks: Insights into Long-Term Information Storage and Manipulation

Arjun Karuvally, Peter DelMastro, and Hava T. Siegelmann

In Annual Workshop on Topology, Algebra, and Geometry in Machine Learning (TAG-ML) at the 40th International Conference on Machine Learning ICML 2023, 2023

Abs PDF

Recurrent neural networks (RNNs) have emerged as powerful models capable of storing and manipulating external information over long periods in various domains. Yet, the mechanisms that underly this behavior remain a mystery due to the black-box nature of these models. This paper addresses this question by proposing an episodic memory theory of RNN dynamics, enabling a more comprehensive understanding of the RNN weights as memories and inter-memory interactions. This approach sheds light on the inner workings of RNNs and connects to existing research on memory representation and organization. The theory extends the current linearization approaches by providing alternative interpretations of the eigenspectrum and its connection to the long-term storage and manipulation of information. We discuss how the segregation, representation, and composition of the variable binding problem—a fundamental question in cognitive science and artificial intelligence—can be mechanistically interpreted within the theory. Using an elementary task - repeat copy, we demonstrate the validity of the theory in experimental settings. Our work represents a step towards opening the black box of RNNs, offering new insights into their functionality and bridging the gap between recurrent neural networks and memory models.