In early 2011, Feng Niu removed the locks. He and Christopher Ré were working on a machine-learning implementation at the University of Wisconsin-Madison. Its processors took turns changing a shared model, a sensible precaution against interfering with one another. Niu tried letting them work without that precaution. The program ran faster, and the answers held up. A small act of computational bad manners had produced a question worth investigating.
The experiment became Hogwild!, a paper with Benjamin Recht, Ré and Stephen Wright. Nine years after its 2011 publication, it received the NeurIPS Test of Time Award. By then, Niu had completed his doctorate and co-founded Alation and Lattice Data. Apple had acquired Lattice. His graduate-school work was still travelling under its own steam.
Today, Niu is CEO and co-founder of Evidently. Read his career through the companies alone and it looks like a sequence of new beginnings. Read the research alongside them and a more particular story emerges: a continuing interest in what prevents computers, and the people using them, from making useful progress with information.
The experiment that stopped waiting
Hogwild! concerns a practical difficulty in teaching a machine from examples. A model is adjusted repeatedly, using small pieces of data. This method, stochastic gradient descent, can require substantial computation. Dividing the job among processors offers an obvious attraction. The awkward part arrives when several workers want to update the same model.
Locks impose order. One worker changes something while others wait. The original Hogwild! paper explored letting processors make updates to shared memory without locking it, even though they might overwrite one another’s work. It established conditions under which that arrangement could remain effective. The result depended on sparsity: individual updates touched relatively small portions of the model.
Imagine several editors working on a very long document. If they mostly change different passages, requiring everyone to wait while one person corrects a comma would be an expensive form of politeness. That is an analogy, rather than a description of the software, but it captures the attraction of reducing coordination when interference is limited.
Concurrent updates
to shared memory
There is a pleasing order to the discovery. First came an implementation and an unexpected observation. Then came the task of explaining it. Ré brought the question to Recht; the collaboration connected the experiment with optimization analysis. The four names on the paper belong to that combined effort. Removing a mechanism was the beginning of the work, with a proof still to do.
An award with a long memory
The 2020 award recognized a paper that had remained influential well beyond its first conference appearance. For Niu, it also reached back to a particular phase of life: graduate school, when three professors who would become his co-authors were at Wisconsin at the same time.
From publication to the NeurIPS
Test of Time Award
His own recollection is disarmingly unceremonious. He calls Hogwild! a side project and expresses surprise at its later recognition. An award can make a career look inevitable in retrospect. His account preserves some of the surprise.
Even the punctuation has a history. Ré supplied the name, complete with its exclamation mark. Academic titles often promise a framework, a method or an approach. This one brought an animal and some enthusiasm. The paper underneath was precise about the conditions that made the apparent disorder work.
A database meets a difficult sentence
Hogwild! was one part of a wider research life. In 2011, Niu also co-authored Tuffy with Ré, AnHai Doan and Jude Shavlik. The project combined statistical and logical reasoning with the machinery of relational databases. Its question was practical: how could a powerful reasoning framework handle data volumes that existing implementations struggled to manage?
Tuffy used database optimization and a hybrid architecture to make inference more scalable. The applications included extracting information from text and resolving whether different references point to the same entity. A useful idea in artificial intelligence can be constrained by the system that must execute it. This work addressed the execution problem directly.
Then came DeepDive. In a 2012 demonstration paper, Niu, Ce Zhang, Ré and Shavlik described building a knowledge base from hundreds of millions of web pages. The raw material was imperfect and varied; statistical learning and inference helped combine it. Scale was central to the project, rather than a detail to resolve after a neat example worked.
A sentence may contain a fact without presenting it in the orderly form a database expects. Extracting that fact, connecting it with other evidence and handling uncertainty are separate jobs. Niu’s research kept entering that space between the material people produce and the structures machines can use. Ordinary prose turns out to give software quite a lot to think about.
How much can more data do?
A different 2012 collaboration posed a question that remains easy to grasp. When teaching a system to extract relationships from text, how much does it help to give it more documents? How much does it help to ask people to label examples? Niu explored the comparison with Zhang, Ré and Shavlik in Big Data versus the Crowd.
Their experiments used corpora of up to 100 million documents and tens of thousands of examples labelled by people. They found a statistically significant improvement in quality from increasing the corpus used for distant supervision. Human feedback helped too, although its measured effect was smaller in those experiments. The findings belonged to a particular study, with particular methods and tasks.
The distinction matters. A research result becomes more useful when its boundaries stay attached. Here, the team measured different ways of supplying a system with evidence. Beneath the large document count sat a fairly economical question about effort: which additional input improved the output, and by how much? It fits comfortably beside Niu’s other work on making demanding information tasks tractable.
Four founders, one missing answer
Niu earned his Wisconsin PhD in 2012, following a computer science degree at Tsinghua University in 2008. Wisconsin records Google as his initial placement. The transition brought his research interests into contact with the daily problems of organizations that had already accumulated plenty of data.
Alation dates its founding to 2012. Its four founders approached a shared difficulty from different directions. Satyen Sangani brought experience from Oracle. Aaron Kalb brought design and language expertise. Venky Ganti and Niu brought database and machine-learning research. People inside companies struggled to find information, understand it and decide whether to trust it.
The company’s early data catalog combined machine learning with human curation. That combination gave the technical work a social dimension. An organization’s knowledge includes what its people understand about the data, as well as the data itself. A catalog had to help connect those pieces. It is easy to admire a warehouse; someone still needs to know which shelf to visit.
In March 2015, investor Greg Sands singled out Niu and Ganti’s technical leadership when discussing Costanoa’s investment in Alation. His assessment covered both solving difficult problems and attracting technical colleagues who wanted to work on them. It adds a concrete outside view of Niu’s contribution: building a team around the engineering, alongside building the engineering itself.
“I’m lucky to have built two successful companies”
Feng Niu, reflecting on Alation and Lattice Data
The dark-data years
In 2015, Niu co-founded Lattice Data. Its founding group included Ré, Michael Cafarella and Raphael Hoffmann. The company worked on converting unstructured material into structured information. Text and images could contain something valuable while remaining awkward for conventional analysis tools to process. Lattice took that awkwardness as its business.
- 2012Wisconsin PhD
Alation founding - 2015Lattice Data
co-founded - 2017Apple acquires
Lattice Data - 2020Hogwild! receives
Test of Time Award
Apple acquired Lattice in 2017, a milestone also recorded by its investor Madrona. The acquisition put another company name on Niu’s route. It also extended an existing collaboration. In a later technical keynote, Ré described continuing to build production applications with Niu at Apple, involving extraction, search and integration.
That account keeps the work visible after the transaction. A company can disappear into an acquisition announcement while its engineers continue facing specific problems. Ré’s description returns to code and applications, including the difficulty of building complex machine-learning systems. The research partnership had found another setting in which to operate.
When the adjective disappears
Niu’s next company chapter is Evidently, where he serves as CEO and co-founder. In December 2024, the company announced a $15 million Series A led by DN Capital, with FRAMEWORK, Clear Ventures, Fellows Fund and existing investors participating. His role now includes the company-building responsibilities that sit around the technical work.
His stated ambition offers a quieter measure of success than a funding round. He wants decisions informed by data to become so ordinary that announcing the fact would resemble pointing out that an appliance uses electricity. The comparison is funny because the adjective suddenly sounds unnecessary. The information has become useful enough to be taken for granted.
That is a fitting destination for a career that began, in one memorable episode, with processors waiting for permission to proceed. The intervening years contain proofs, papers, collaborators and companies. Again and again, Niu’s work has approached the friction between available information and practical use. Sometimes the obstacle needs a new system. Once, in Wisconsin, it needed someone willing to remove a lock and see what happened.