Signaloid: Shaping the Future of Uncertainty Quantification in Computing
Uncertainty is an inherent part of our lives, from the movements of stocks to hitting sales predictions for the next quarter. No measurement is perfect, and many processes are noisy.
When it comes to data about the real world, it is typically a distribution of possible outcomes—some more likely than others—rather than a single point. Yet, computers make calculations with individual numbers that represent points rather than distributions; thus, estimating uncertainty becomes a lot of overhead. But what if a new way of implementing the maths within computers could inherently account for distributions and, thus, uncertainty?
Enter Signaloid, a startup founded by Phillip Stanley-Marbell in 2019, developing a processor that has created a new way of doing the maths within computers so that they can track and manage uncertainty during calculations. As a virtualization layer running on top of existing hardware, it seamlessly integrates into existing workflows while dramatically reducing the computational effort and costs spent on dealing with uncertainty. Signaloid has raised a €3.5M seed funding led by IQ Capital, joined by Type One Ventures and Creator Fund, in the spring of 2023, and went through the Intel Ignite startup program.
Learn more about the future of uncertainty quantification in computing from our interview with the founder and CEO, Phillip Stanley-Marbell:
Why Did You Start Signaloid?
I’ve been working for over twenty years in computing, going back and forth between industry and academia, working at Bell Labs and being appointed as the first Chair of Physical Computation as a professor at the University of Cambridge. Yet, people kept asking me why I wouldn’t turn my research into a company. I realized that if I didn’t go for it, someone else would, and I would regret it. So, I founded Signaloid in 2019.
How Do You Make Computers Quantify Uncertainty?
Every computer does calculations. But one important thing no computer currently can do inherently is to judge how much to trust the data they’re processing. There is always some uncertainty about real-world data, whether the data acquisition method is imperfect or the underlying real-world process is noisy.
At Signaloid, we built a new processor to help computers deal with uncertainty. If you provide information initially on how uncertain data is, it will keep track of this uncertainty during further processing. This can help, for example, with Monte Carlo simulations in computational finance, help autonomous robots estimate how far an object is and how much you trust that estimate, and even with topics in materials modeling and quantum computing.
Imagine your customer’s software is an orange Lego brick sitting on top of a normal CPU, a blue Lego brick. Suppose you want to predict the price of a financial stock with Monte Carlo simulations. You can change the software to implement a Monte Carlo simulation, and you will then have to run thousands or millions of iterations on normal computing hardware, which is a lot of effort and cost. Or, you can use our processor to do all of these things using new mathematical techniques we have implemented inside the processor’s design.
It’s like a transparent Lego brick sitting in between the blue and orange one, providing a virtualization layer on top of your existing computing hardware but making it look like a new kind of processor. Our customers just see the virtualization layer as a new backend, not what hardware is beneath it. So they can get it up and running more easily without having to rewrite their software, and still get high-quality results.
Since you can’t represent real numbers on a digital computer exactly, digital computers today deal with either integers or floating point numbers with a certain bit-level precision. We add a third data type that accounts for a distribution of values, representing the uncertainty around a data point and capturing what other outcomes may be possible, though not necessarily likely. Our challenge was creating an efficient representation of a distribution, as digital computers only have a limited number of bits to use to represent any distribution.
There are several naive ways you could try to solve the problem, such as representing a distribution by its first N moments, but such methods don’t work as well as the methods we came up with and ultimately implemented. Our approach is very efficient in representing any kind of distribution and then efficiently doing any sort of arithmetic with it. Our goal is really to make the best use of the bits we have and ensure calculations work fast and smoothly.
You can sign up to use our developer platform with just three clicks and try it out yourself. In addition to the proprietary applications we are working on with select customers, we have a few example applications shared with the public on GitHub, ranging from simple kernels to complete applications in materials engineering and hybrid quantum-classical computing. For those interested in the big-picture implications of our technology, earlier this year, we put together a concise high-level video about Signaloid, targeted at a non-technical audience.
How Did You Evaluate Your Startup Idea?
Starting as a faculty member at the University of Cambridge, I had the safety of the university environment to try out all the risky things that may not work and get the technology to a point where it was worth taking a leap of faith and founding a company.
The goal from day one was to validate the commercial potential of our solution as quickly as possible. This influenced our decision to implement the solution using a virtualization layer, allowing us to make any hardware acceleration or custom silicon we implement a purely operational detail that does not affect the abstraction layer seen by our customers. I think it is important to be product and customer focused and to have a product that is being used by paying customers as early in the company’s evolution as possible. That way, we can have greater confidence that we are solving the problems that paying customers care about.
We currently have over 1900 developers who have signed up organically for one of our product tiers (Free, Developer, Developer+API, and Enterprise). In addition to having a product that any individual or organization can sign up and use today, we have been working with both SMEs and large Fortune 500 corporations to run pilot proof-of-concept projects with the aim of translating those pilots into recurring revenue. In recent months, we have started an active outbound sales activity targeting potential customers in our priority verticals of finance and manufacturing, as well as working with horizontal channels like cloud service providers.
What Advice Would You Give Fellow Deep Tech Founders?
It’s very difficult to hire strong candidates. There are various strategies for dealing with this, and one of our angel investors, Catherine Lenson, has written a brilliant LinkedIn piece on this. In short, advertise on various platforms and be open to going remote-first. In Cambridge, we may only get a dozen or so applications, but someone in Barcelona might not be looking for a job in Cambridge. We had good experiences with going remote-first and advertising local copies of our job ads tailored to cities with strong universities like Cambridge, Munich, Copenhagen, Barcelona, and Eindhoven.
Last but not least, we are delighted that we had the chance to participate in the Intel Ignite startup program. It is amazing, and you should give it a look too.
Further Reading
Cambridge-based real-time uncertainty quantification platform Signaloid secures €3.5 million seed round – Press release about Signaloid’s seed round featured by EU-Startups
Intel Ignite Program Selects 10 European Startups for 4th Cohort – Signaloid was part of this!
The Laplace Microarchitecture for Tracking Data Uncertainty – Research article presenting a microarchitecture for representing probability distributions on a computer