welcome. i'm a fifth year masters student studying vision & intelligent systems at IIT Kharagpur.
i like computers. i really like computers that learn fast. here is my curriculum vitae.
let's get to know each other. please scroll down, or use the navbar above.
on mobile? got motion sickness? prefer retro? try the static version of the website .
I'm an undergrad at IIT Kharagpur. I'll graduate in 2025 with a B.Tech in ECE and M.Tech in Vision & Intelligent Systems. I'm broadly interested in MLSys. My overarching goal is to develop AI algorithms and systems which are democratic & usable. My feeble attempts at doing so thus far have been supported by Boeing, IITKGPF-USA and the NSF. This is me.
I am particularly interested in these closely related avenues towards "democratization and usability":
I see Neurosymbolic AI as a crucial frontier of ML research. Causal System 2 reasoning might help models become
leaner, generalize from less data and align with human values. Distilling neural agents and bootstrapping them with
solvers enables scientific discovery and safety-critical cyber-physical applications.
Distributed/Parallel/Pipelined Optimization will allow large-scale participation in ML training and enable
low-resource users to train their own state-of-the-art models. For that, it is necessary to investigate such
methods that maximize privacy and tolerate faults.
A common theme tying these together is brain-inspired computation. Now, I'm not one
to restrict our discrete, verifiable computer programs to follow natural, error-prone stochastic patterns. But they
can provide some select properties and heuristics. Linguists say we're born with a "universal grammar" that provides
a foundation that we wire up as we acquire language, allowing a distinct efficiency. Our neurons don't wait for
a supervisory signal to update, unlike backprop. Brain ontogenesis works though it
doesn't get to sample an exponential search space. Programs could benefit from such statistical modelling.
I spent my summers with Prof. David Inouye at
Purdue working on greedy & distributed optimization and at
AirLab, CMU working on ILP, LLMs and map-matching. At KGP, I lead the AGV undergrad group's DL team on
scene understanding, IRL and FL.
In my free time, I code goofy mini-projects, most of which never make it to my GitHub out of embarassment. I enjoy
playing and watching sports (especially basketball). I also spread vitriol online through my blog. I've represented
IITKGP twice in word games at the collegiate level. Recently, I've taken a liking to competitive programming and CTFs.
I did say I liked computers.
I grew up in the wonderful Hauz Khas, New Delhi, India. During breaks, you'll find me there hanging out with
old friends, my parents and the dog. I was born in Baltimore and spent some early years in Santa Clara. I guess
that makes me an expat/international student from India's perspective.
back (home) · detailed statement + interests · forth (papers)
If you scroll down, you'll find abstracts for my publications so far.
This includes most of my conference papers, journal papers, workshop papers, preprints.
back (about) · google scholar · skip papers (to current projects) · forth (paper 1: DVFL)
Abstract: Vertical Federated Learning (VFL) is an emergent distributed machine learning paradigm wherein owners of disjoint features of a common set of entities collaborate to learn a global model without sharing data. In VFL, a host client owns data labels for each entity and learns a final representation based on intermediate local representations from all guest clients. Therefore, the host is a single point of failure and label feedback can be used by malicious guest clients to infer private features. Requiring all participants to remain active and trustworthy throughout the entire training process is generally impractical and altogether infeasible outside of controlled environments. We propose Decoupled VFL (DVFL), a blockwise learning approach to VFL. By training each model on its own objective, DVFL allows for decentralized aggregation and isolation between feature learning and label supervision. With these properties, DVFL is fault-tolerant and secure. We implement DVFL to train split neural networks and show that model performance is comparable to VFL on a variety of classification datasets.
This work is currently under peer review. Preprint below (to be revised soon). I also presented an early sketch of the idea at the SURF Symposium at Purdue University.
This was my Bachelor's thesis. I completed a two-semester track thesis in one. I picked this problem, formulated the solution, designed & programmed the experiments and wrote the paper. Since the thesis submission, I've been trying to add more datasets and models and get it accepted at a conference. Thanks to Yash for helping out with some of the legwork since I didn't have as much time during the next semester to do it all alone. Thanks also to Profs. Inouye and Jithin R for their valuable guidance.
Report (preprint, under peer review)
back (all papers) · skip papers (to current projects) · forth (paper 2: entity augmentation)
Abstract: Vertical Federated Learning (VFL) is a machine learning paradigm for learning from vertically partitioned data (i.e. features for each input are distributed across multiple "guest" clients and an aggregating "host" server owns labels) without communicating raw data. Traditionally, VFL involves an "entity resolution" phase where the host identifies and serializes the unique entities known to all guests. This is followed by private set intersection to find common entities, and an "entity alignment" step to ensure all guests are always processing the same entity's data. However, using only data of entities from the intersection means guests discard potentially useful data. Besides, the effect on privacy is dubious and these operations are computationally expensive. We propose a novel approach that eliminates the need for set intersection and entity alignment in categorical tasks. Our Entity Augmentation technique generates meaningful labels for activations sent to the host, regardless of their originating entity, enabling efficient VFL without explicit entity alignment. With limited overlap between training data, this approach performs substantially better (e.g. with 5% overlap, 48.1% vs 69.48% test accuracy on CIFAR-10). In fact, thanks to the regularizing effect, our model performs marginally better even with 100% overlap.
Report (IJCAI '24 GLOW, Archival)
back (paper 1: dvfl) · all papers · skip papers (to current projects) · forth (paper 3: internet learning)
Abstract: Distributed machine learning has grown in popularity due to data privacy, edge computing, and large model training. A subset of this class, Vertical Federated learning (VFL), aims to provide privacy guarantees in the scenario where each party shares the same sample space but only holds a subset of features. While VFL tackles key privacy challenges, it often assumes perfect hardware or communication (and may perform poorly under other conditions). This assumption hinders the broad deployment of VFL, particularly on edge devices, which may need to conserve power and may connect or disconnect at any time. To address this gap, we define the paradigm of Internet Learning (IL), which defines a context, of which VFL is a subset, and puts good performance under extreme dynamic condition of data entities as the primary goal. As IL represents a fundamentally different paradigm, it will likely require novel learning algorithms beyond end-to-end backpropagation, which requires careful synchronization across devices. In light of this, we provide some potential approaches for the IL context and present preliminary analysis and experimental results on a toy problem.
This was the first bit of my work as a Summer Undergraduate Research Fellow at Purdue University. I helped Prof. Inouye design and program the experiments, and wrote the appendix and relevant sections of the paper. I also wrote an intermediate revision of the submission from scratch. This work was accepted to an ICML 2023 workshop.
back (paper 2: entity augmentation) · all papers · skip papers (to current projects) · forth (paper 4: ynet transfer and reproduction)
Won Machine Learning Reproducibility Challenge 2022. Accepted to ReScience C and invited to NeurIPS 2022 special reproducibility track poster session. Successfully reproduced the results and showed significant transfer learning abilities.
Abstract: Human trajectory forecasting is an inherenty multimodal problem. Uncertainty in future trajectories stems from two sources: (a) sources that are known to the agent but unknown to the model, such as long term goals and (b) sources that are unknown to both the agent & the model, such as intent of other agents & irreducible randomness in decisions. This stochasticity is modelled in two major ways: the epistemic uncertainity which accounts for the multimodal nature of the long term goals and the aleatoric uncertainity which accounts for the multimodal nature of the waypoints. Furthermore, the paper extends the existing prediction horizon to up to a minute. The aforementioned features are encompassed into Y-Net, a scene compliant trajectory forecasting network. The network has been implemented on the following datasets : (a) Stanford Drone (SDD) (b) ETH/UCY (c) Intersection Drone. The network significantly improves upon state-of-the-art performance for both short and long prediction horizon settings.
Report (ReScience C Vol. 8 No. 3)
back (paper 3: internet learning) · all papers · forth (current projects) ·
In this section you'll find projects that I'm hacking away at right now.
What to expect here: My personal projects are pretty heavily influenced by my research interests. Usually there'll be deployments or integrations of topics I'm researching with other stuff I find interesting. There is also a slightly heavier systems bent here because I love toying with systems and I am a bit of a security/privacy nut.
Please scroll down or use the links below to navigate/skip.
back (mlrc) · github · skip current projects (to more projects) · forth (project 1: info theory nesy learning)
This project is still ongoing. I will update this section once we submit our paper.
My work is to advance a project that aims to forecast and analyze aircraft activity at airports. My goal is: given a long and complex natural-language rulebook and a view of the airport, to reliably determine if any rule is about to be broken.
Existing work on this project includes a large-scale dataset and a software repository which processes the data, predicts trajectories, performs visualizations and simulations etc. Besides my work on rule-checking, I am also adding a few features to the software repository: semantic grounding for data and semantic-aware trajectory refinement.
back (all projects) · forth (current project 2: mcpvpnesyrl)
Performing fixed tasks in the open-world sandbox game Minecraft is a popular benchmark for RL agents. As someone who used to enjoy a spot of Minecraft as a kid, I had the fantastic idea of training an agent to do a slightly more difficult but fun task: PvP combat.
This is a slightly more challenging task than, say, the MineRL benchmark. Firstly, our goal is to learn purely from visual + input data, viz. without using the internal game state. Second of all, Minecraft PvP dynamics are a little weird, and the agent needs to learn a very specific technique from an open-ended and continuous search space.
I had originally started this project when I was home during the pandemic, and had to give it up because I couldn't bring my PC to campus when offline class restarted. Now, I am armed with more money (so I can go back home occasionally to hack away at this project) and insight (I want to try imparting symbolic primitives as an expert Minecraft PvPer and allow the model to learn its parameters in order to help the policy zone in on the "intricate" PvP mechanics).
back (current project 1: info theory nesy learning) · all projects · forth (current project 3: inverse rl traj pred)
Trajectory Prediction has become a benchmark task with the ostensible goal of deployment in autonomous vehicles. But are models that get the best ADE/FDE really the best for downstream tasks such as path planning?
We wish to investigate two prongs: given an environment that evolves based on some probabilistic logic, can an inverse RL agent extract the logic? If not exactly, how far are real trajectories from those rolled-out by the discovered logic? Further, can we use the logic (or better yet, the whole reward map) to learn a better planner than the SotA forecasting methods?
And, can we predict paths using the learned reward map that are performant on typical trajectory forecasting benchmarks?
back (current project 2: mcpvprl) · all projects · forth (more projects)
I have done far too many silly little projects to list. In this section [WIP], I'm going to enumerate some of the more interesting ones that didn't appear above. These include unsubmitted research projects, personal projects and a smattering of course projects.
What to expect here: I've been programming in C/C++ since the 8th grade and Python since my first year of undergrad. I also have a penchant for niche technologies. Sometimes I go out of my way to find cool stuff to implement in unpopular frameworks and languages. But since most of my significant work is related to machine learning, my GitHub is sadly whitewashed in Python.
I guess I also kinda familiar with Prolog now. I also know some bash, asm (8051, x86, hopefully some RISC stuff soon) and ostensibly even Verilog. When I get time I want to get better at OCaml, Rust and Scala.
I'm comfortable with C++'s standard libraries and actively studying some implementations of low-level libraries in C. I also intend to study CUDA programming in the near future. I'm well-versed in the typical ML/data pipeline for Python with PyTorch (including Geometric and Lightning), scikit, Pandas and what have you. I also use some security tools for CTFs like radare2, pwntools and so on.
I'd also like to think I'm a good technical writer. I've got a few peer reviewed works under my belt.
My e-mail is avi.amalanshu@kgpian.iitkgp.iitkgp.ac.co.in..com I anticipate your hate mail.
My email and LinkedIn are probably the easiest way to get a hold of me.
I also have a Twitter (avi_amalanshu) but I don't really use it.
Check out my blog: malansh on Medium. I plan to start writing on it again when I'm a little more free, say Winter 2024. Topics I have in mind cover a mix of general advice, opinion pieces, and academic nerdery. (Information Geometry, anyone?)
I'm always on the lookout for interesting puzzles and research problems, especially stuff that's interdisciplinary or niche (underappreciated). Let's talk if you have something interesting and can use my contributions.
I love handing out advice and mentoring (to the extent that I often do so unsolicited). So feel free to solicit if you're interested! Be it JEE prep, handling acads + research at IIT, and of course, breaking into fields I'm working in.
I'm graduating in 2025 and I'm looking for opportunities. Please reach out.
back (key project 3: internet learning) · back to top