Writing
- A small archive of ideas, experiments, and things worth slowing down enough to explain. + Hardware quirks, deep learning systems, and research problems + that needed more than the product page.
Archive
- {posts.length} published {posts.length === 1 ? 'essay' : 'essays'}. The archive stays selective. + {posts.length} {posts.length === 1 ? 'essay' : 'essays'} published so far. I publish when I have a real question to chase down.
Akshay Kolli / Research + Writing
+Akshay Kolli / UMass Lowell
- I'm a CS PhD student at UMass Lowell building world models for autonomous agents. + I'm a CS PhD student working on world models for reinforcement learning, + ML hardware, GPUs, and deep learning systems.
- Inferring the hidden graph behind a multi-agent system by training attention to predict what each agent does next. + We use next-state prediction to expose the interaction graph behind a multi-agent system.
Overview
- The paper studies a practical problem in networked multi-agent systems: the agents move, synchronize, or converge, but the interaction graph behind that behavior is not given. + In many multi-agent systems, you can watch agents move, synchronize, or settle, but you are not given who influences whom.
- The model is trained on state trajectories from consensus dynamics and Kuramoto oscillators. During prediction, attention scores between agent embeddings become an interpretable approximation of the adjacency matrix. + This project tests whether a predictor trained on trajectories can recover that missing adjacency matrix. On consensus dynamics and Kuramoto oscillators, the learned attention scores become a usable proxy for edges.
Method
- Instead of supervising the graph directly, the model learns to forecast the system. The attention matrix is then thresholded and compared with the true graph. + The model learns the dynamics first. After training, we threshold the attention matrix and compare the inferred edges with the true graph.
Results
- F1 link-prediction scores are above a random baseline for both tested dynamics. Consensus dynamics are easier than Kuramoto oscillators, while additional simulation data improves larger-agent inference. + F1 scores beat a random baseline for both dynamics. Consensus is the easier case; Kuramoto needs more data as the number of agents grows.
training behavior
+training trace
- Attention first learns the obvious self-dependencies, then gradually recovers inter-agent structure as training progresses. + Early training mostly finds self-dependencies. Later epochs start assigning weight to the agent-to-agent links.
Projects
- A compact index of paper pages, experiments, and systems work. + Most of my current work sits around ML hardware, deep learning models, + and world models for RL. This topology paper is the multi-agent exception.
@@ -77,7 +78,7 @@ export default function ProjectsPage() { href={project.href} className="font-medium text-ink underline decoration-line-strong underline-offset-4 transition-colors hover:text-accent hover:decoration-accent" > - Open project + Read project
Resume
- Research, engineering, and systems work across academia and industry. + My background across ML research, simulation tooling, and production software.