Welcome to the Avi Portal...█

Greetings, traveler of the digital realms. I am Avi Amalanshu of CMU (MSML), former dual-degree (5-year B.Tech + M.Tech) at the ECE department, IIT Kharagpur. My vision is that one day you and I can train and infer from reasonable AI agents aligned to our own values without bearing the cost to throw half the internet on a bazillion GPUs for a whole year.

Towards that, I'm exploring the theory of representation learning. My broad hypothesis is that domain expertise should inform structural priors, not knowledge priors; given that, we might aim to uncover latent aleatoric structure. Finetuning to learn epistemic behavior within that framework is, then, akin to theories behind generalizable and efficient human learning from the linguists' universal grammar school of thought.

My sweet spot is the boundary between state-of-the-art and esoteric. I like to call it avant-garde.

An Information-Theoretic Bridge Between Neural and Symbolic AI

Neural agents are great at System 1 thinking: fast, intuitive, statistical. Not so much at System 2 reasoning, which is slow, deliberate and logical.

Our brains can learn from remarkably few samples and make interpretable, verifiable decisions based on sound logical reasoning. The best of foundation models even after being trained on unfathomable amounts of data are not remotely reliable, even with fancy feedback-loop-chain-of-thought prompt tricks.

So how do we get there?
Optimizing over discrete logic typically implies a combinatorial search is needed somewhere-- probably NP hard, not good. Can we make things better by modeling probabilistic logical distributions?

Manuscript under preparation.

Amelia

Intent Prediction for Airport Surface Operations.

My role in this large-scale Boeing project was to come up with an efficient map-matching algorithm to induce GIS information into a trajectory prediction model. Also, a way to use English rules via LLM as a heuristic for procedural bias in Popper, an inductive logic programming system.

Entity Augmentation

Imagine a database whose columns belong to different organizations.

How do you learn something meaningful in this system? Sure, you could collect all the features at a central location and use standard machine learning. But what if they don't share?
Vertical Federated Learning is intricately coordinated. You need to figure out rows everyone knows some features for (without leaking the keys). Then, on every training iteration you need to arrange matters so that everyone is passing their features pertaining to the same key, so that the aggregator knows what to predict.

Or, you could just use Entity Augmentation.
GLOW @ IJCAI 2024 (Archival) and HoTDiML @ ICDCS 2025

Decoupled Vertical Federated Learning

Vertical Federated Learning is not easy.

All it takes is one connection or participant to fail and the whole thing crashes. Also, if one of your participants is curious, they can learn a lot about others' data from gradients the aggregator sends them. And of course there's the whole bore of entity alignment (and therefore limitations on sample size). It's hard enough cross-silo. Forget about scaling up.

Try Decoupled Vertical Federated Learning instead.
arXiv (short version in SSL @ NIPS 2024)

Internet Learning

We're wasting the Internet by merely passing data around. Why don't we compute, too?

Internet Learning is a paradigm for learning over decentralized networks. Our baseline proposal is a collaborative backpropagation over the whole network. But you are encouraged to propose something better.
Learning on dynamic networks with unreliable and unavailable edge computation is not trivial. Here's a first step towards fixing that.
LLW @ ICML 2023