Hans, an Arab stallion from Russia, had proven himself to be a clever horse indeed. He had learned to use his hoof to tap out numbers written on a blackboard. Much to Von Osten’s (a German high school mathematics instructor) delight, a “3” on the blackboard would prompt a tap-tap-tap from Hans. “What is the square root of sixteen?” Four taps. Von Osten taught Hans simple arithmetic, tell the time and date, and even understand German. In 1891, he travelled around Germany with Hans, showing off his clever horse to the public. Wilhelm von Osten was convinced that animals possess an intelligence comparable to humans.
Clever Hans gained popularity and attracted the attention of Germany’s board of education. They formed the Hans Commission to investigate horse’s intelligence further. In 1904 they concluded, “that there was no trickery involved in Hans’ responses; as far as they could tell, the horse’s talents were genuine.”
However, Oskar Pfungst, a student at the Psychological Institute at the University of Berlin, was not convinced and finally unravelled the mystery. Pfungst noticed that every questioner’s breathing, posture, and facial expression involuntarily changed each time the hoof tapped, showing an ever-so-slight increase in expectation and tension. Once the “correct” tap was made, that subtle underlying tension suddenly disappeared from the person’s face, which Hans took as the cue to stop tapping. These facial cues were not present when the questioner was unaware of the correct answer, leaving Hans without the necessary feedback. Clever Hans was not dipping into a reservoir of intellect to work out the answers. He was merely receptive to the subtle, microscopic, and unconscious cues universally present in his human questioners.
Clever indeed.
Pfungst learned to signal or send cues to Hans by slightly raising his eyebrows and get the correct answers (or the answers he wanted) from the horse. This later came to be referred to as the “Clever Hans effect” and explains how someone could take unintentional cues of desired response by the questioner.
The Clever Hans effect has relevance today in studies of the mental capacities of animals, humans and feature identification for machine learning methods. Being aware of the phenomenon makes us cautious of unintentionally leading subjects through subtle cues to give answers that seem right to the researchers and data scientists.
Learning Models in AI
We encounter the Clever Hans phenomenon while training statistical and deep learning models referred to as spurious correlation in the statistical literature (first introduced by Karl Pearson in 1897). Tyler Vigen, a JD student at Harvard Law School and the author of Spurious Correlations, demonstrates this extensively with examples on his website. Additionally, Vigen has also programmed his site so that anyone can find and chart absurd correlations in large data sets. An example from Vigen (illustrated below) correlates between deaths by getting entangled in bedsheets and cheese consumed. The datasets are highly correlated but have no practical purpose of making predictions in real life.
The Clever Hans effect occurs when a feature in the data set is highly correlated with the desired outcome (such as calculations by Clever Hans) but is not the cause (unconscious cues given by the questioner) of the correct answer. In other words, we can see the “Hans effect” in AI learned models when the system produces correct predictions based on wrong features/parameters. Such effects usually go undetected in standard validation techniques and come to light only when the ML model starts operating across actual and more generalized data.
There are many examples where ML models have learned to predict correctly (during learning) based on wrong or highly correlated input parameters such as recognizing a “boat” based on “water” around it or linking “smiling” to the “age” of the person.
An article published in Nature Communications clarifies with examples the problem that the “Clever Hans” phenomenon creates in deep learning (non-linear) models. “The classifier trained on the PASCAL VOC 2007 data set focuses on a source tag (see yellow box in the pictures below) present in about one-fifth of the horse figures. Removing the tag also removes the ability to classify the picture as a horse. Furthermore, inserting the tag on a car image changes the classification from car to a horse.”
Clever Hans never had a chance to meet Jim Key. In 1897, Jim Key from Tennessee became the most educated horse. Jim was born in 1895 to Lauretta (with Egyptian lineage) and Tennessee Volunteer, the number one racing horse at that time. Bill Key trained him for 3 – 4 years before he started seeing real signs of intelligence. There is a quote from a newspaper “…Jim grasped a piece of chalk in his mouth and scrawls’ Jim’ on the blackboard”. Jim Key performed many tricks and amazed his audiences, e.g., he could read, count from 1 to 25, and tell time. Jim died in 1912. Unlike in the case of Hans, no one (including a group of Harvard professors) was ever able to prove the act was fake or that the horse was taking cues from his trainer instead of following the audience’s instructions.
It shows that Cicero (106–43 BC) was not right when he said that horses (and other animals) are without intelligence.
Further Reading
- Clever Hans: The True Story of the Counting, Adding, and Time-Telling Horse by Kerri Kokias
- Beautiful Jim Key: The Lost History of the World’s Smartest Horse by Mim E. Rivas
- Beware Spurious Correlations, Harvard Business Review, Managing Organizations
- Analyzing Classifiers: Fisher Vectors and Deep Neural Networks
- Explaining Deep Neural Networks and Beyond: A Review of Methods and Applications
Leave a Reply