Ecosystems in industry are commonly composed of various data types in terms of data modalities or feature distributions. Heterogeneous graphs (HGs) present these multimodal data systems in a unified view by defining multiple types of nodes and edges — for instance, e-commerce ne…
TL;DR:
Historically FIFO-based algorithms are thought to be less efficient (having higher miss ratios) than LRU-based algorithms.
In this blog, we introduce two techniques, lazy promotion, which promotes objects only at eviction time, and quick demotion, which evicts most new ob…
Part of the challenge (and fun) of low-level systems code is in the optimizations they employ:
developers might use manual memory management, they might use bit-packing and bit-twiddling optimizations,
or they might use multi-threading to speed up their code.
When dealing with su…
What if you could run untrusted code and still be able to sleep at night, safe and sound?
Disclaimer: our award-winning work [1] can only calm your unsafe-software related fears; we recommend complementing this by additionally checking for monsters under your bed, and leaving …
Figure 1: diagram showing the code conversion process in a distributed storage system.
Today’s society is data-driven, and many of the applications that society relies on require storing ever-increasing amounts of data.
To this end, distributed storage systems have become the …
The physicist’s method is a powerful framework for cost analysis that
many a computer scientist will learn at some point in their undergraduate career.
However, its high-level description leaves some practical gaps, especially concerning
how to actually bookkeep its finer details…
As machine learning systems become increasingly implemented in safety-critical applications, such as autonomous driving and healthcare, we need to ensure these systems are reliable and trustworthy. For example, we might wish to determine whether a car’s camera-based autopilot sys…
This blog post is based on a research paper with the same title, authored by Anilesh Krishnaswamy, Haoming Li, David Rein, Hanrui Zhang, and Vincent Conitzer, published at AAAI 2021.
TL;DR: We investigate a classification problem where each data point being classified is controll…
An extended version of this post appears on my personal site.
Suppose you’re building a collaborative app, along the lines of Google Docs/Sheets/Slides, Figma, Notion, etc., but without a central server. One challenge you’ll face is the actual collaboration: when one user cha…
Blockchains are a powerful technology which allow decentralized agreement with an immutable history. Since transactions can be added, but not removed, blockchains allow distributed banking as a trustworthy alternative to central banking.
A vast amount of cryptographic research on…
