If you haven’t heard the term ‘artificial swarm intelligence’ yet, you soon will. Louis Rosenberg, CEO and Chief Scientist at Unanimous AI, the organisation that pioneered the technique, explains all

Artificial Swarm Intelligence, or ‘Swarm AI’, is a recently developed technique for generating accurate forecasts, decisions, and diagnoses by capturing and amplifying the insights of networked human groups. A combination of real-time human input and AI algorithms, the technology connects groups of human participants over the internet to form intelligent systems modeled on swarms in nature.

Current implementations use a simple graphical interface to enable users to answer questions together by working as a ‘virtual swarm,’ each person using their mouse or touchscreen to convey their changing sentiments to the underlying AI algorithms. The process enables the human group to quickly explore an issue, pushing and pulling on each other until they converge on a solution that is generally more accurate than the individuals could generate using traditional methods.

For example, a recent study published by researchers at Stanford University School of Medicine and Unanimous AI tasked small groups of radiologists with diagnosing chest x-rays by either (i) working alone or (ii) diagnosing together as a real-time Artificial Swarm Intelligence. The test was conducted using the ‘Swarm’ software platform from Unanimous AI and employed experienced medical professionals from Stanford School of Medicine and Duke School of Medicine as the human participants. The published results show a 33% reduction in diagnostic errors when small groups of radiologists worked together as an Artificial Swarm Intelligence as compared to working alone or by majority vote. In another recent study published by researchers at MIT and Unanimous AI, groups of financial traders were tasked with predicting the weekly change in price of oil, gold, and the S&P 500 over 20 consecutive weeks. The results showed that when working together as an Artificial Swarm Intelligence, the groups increased their forecasting accuracy by an average of 36% as compared to working alone or majority vote.

While the two studies above used subject-matter experts as the human participants (i.e. radiologists and financial analysts) the technology of Artificial Swarm Intelligence has proven successful when the participants are ‘enthusiasts’ rather than professional practitioners. For example, a study published in 2018 by researchers at Oxford University and Unanimous AI, showed that groups of self-identified ‘sports-fans”’ could outperform professional betting markets when forecasting sporting events as an Artificial Swarm Intelligence. The study tasked randomly selected groups of fans from the UK with predicting the outcome of a set of 50 consecutive games in the English Premier League over a 5 week period. The average individual in the study was 55% accurate when predicting the outcome. But when working together in swarms, the sport-fans boosted their collective accuracy to 72%, thereby achieving a 31% increase in overall performance.

While Artificial Swarm Intelligence has been shown to amplify the intelligence of human groups, there is a drawback compared to traditional methods of aggregating group intelligence. Specifically, because Artificial Swarm Intelligence involves groups collaborating as unified systems, the method requires real-time synchronous participation. This makes swarming more ‘hands on’ than traditional group-based methods. For example, when capturing the input from groups using votes, polls, or surveys, the participants can provide their input asynchronously over an extended period. When forecasting using swarm-based techniques, the participants must log in to a central server at the same time and provide their input to the swarming algorithms in unison. Fortunately, the process is fast, with each decisions, forecasts, or diagnoses being converged upon in only 20 to 60 seconds.

Comparing Artificial Swarm Intelligence to traditional machine learning, we find each has unique strengths which make it appropriate for different types of problems. Specifically, machine learning has proven extremely powerful when large and reliable datasets exist that accurately represent the task at hand. Unfortunately, there is practical weakness in machine learning that is often overlooked – namely that for many real world problems, historical data is not available to accurately represent the problem. In some cases, the historical data exists but quickly gets out of date, rendering it ineffective. In other cases, the historical data does not exist in an accessible format, instead being represented primarily as the wisdom, insight, and intuition of human experts. In such situations, Artificial Swarm Intelligence can be a powerful supplement or replacement to traditional machine learning.

Consider the example of financial forecasting, where large data-sets exist but go stale quickly as market conditions change. Human analysts, on the other hand, are quite skilled at considering current conditions and extrapolating into the future. In this way, amplifying the insights of human analysts using Artificial Swarm Intelligence has proven successful in predicting rapidly changing environments like financial markets. Similar effects have been observed in medicine, where historical databases can be quite good at documenting commonly occurring presentations of medical conditions, but not as good at representing rare presentations. In such cases, the knowledge, wisdom and intuitions of human experts can be tapped using Artificial Swarm Intelligence. In fact, a recent study published in Nature, Digital Medicine by researchers at Stanford University and Unanimous AI showed that by combining the output of a deep learning system and an Artificial Swarm Intelligence system, diagnostic accuracy could be achieved that exceeded either technique alone.

In conclusion, Artificial Swarm Intelligence is a powerful technique for generating accurate forecasts, decisions, and diagnoses by combining the wisdom of human practitioners with the amplification benefits of AI. It is most useful for domains where large historical datasets either don’t exist or fall out of date too quickly to be effective. It is also useful for domains like medicine, where large data exists for common occurrences but not for rare occurrences where accuracy is also required. In such situations, the knowledge, wisdom, and intuition of experts can be harnessed and amplified using swarm-based techniques, either used alone or in combination with traditional machine learning.