Causal disentanglement is the next frontier of AI

A major challenge for artificial intelligence (AI) is the ability to see past superficial phenomena to guess the underlying causal processes. A new study by KAUST and an international team of leading experts has developed a new approach that goes beyond the detection of superficial models.

Humans have an extraordinarily refined sense of intuition or inference that allows us, for example, to understand that a purple apple can be a red apple lit by a blue light. This sense is so developed in humans that we are also inclined to see patterns and relationships where there are none, which gives rise to our propensity for superstition.

This type of understanding is so difficult to codify in artificial intelligence that researchers still do not know where to start – yet it is one of the most fundamental differences between natural and artificial thinking.

Five years ago, a collaboration between KAUST's affiliated researchers, Hector Zenil and Jesper Tegnér, as well as Narsis Kiani and Allan Zea from the Karolinska Institutet in Sweden, began to adapt the theory of algorithmic information to biology networks and systems to solve the fundamental problems of genomics and molecular circuits. . This collaboration led to the development of an algorithmic approach to deduce causal processes that could form the basis of a universal model of AI.

"Machine learning and artificial intelligence are becoming ubiquitous in the industry, science and society," said Professor KAUST, Tegnér. "Despite recent progress, we are still far from achieving versatile machine intelligence with the ability to reason and learn in different tasks. Part of the challenge is to go beyond the superficial detection of models to techniques for discovering the underlying causal mechanisms that produce the models. . "

This causal disentanglement, however, becomes very difficult when several different processes are linked, as is often the case in molecular and genomic data. "Our work identifies the parts of the data that are causally related, eliminating parasitic correlations, and then identifies the different causal mechanisms involved in producing the observed data," said Tegnér.

The method is based on a well-defined mathematical concept of probability of algorithmic information serving as a basis for an optimal inference machine. The main difference from previous approaches, however, lies in the shift from a view of the observer-centric problem to an objective analysis of phenomena based on deviations from randomness.

"We use algorithmic complexity to isolate several interacting programs, then we search for all the programs that can generate the observations," explains Tegnér.

The team has demonstrated its method by applying it to interacting results of several computer codes. The algorithm finds the shortest program combination that could construct the convoluted output string of 1 and 0.

"This technique can equip current machine learning methods with advanced complementary capabilities to better manage abstraction, inference and concepts, such as cause and effect, that Other methods, including in-depth learning, currently can not handle, "says Zenil.

Explore further:
The machine learning reveals a turtle motif hidden in a quantum firework

More information:
Hector Zenil et al. Causal deconvolution by algorithmic generative models, Nature Machine Intelligence (2018). DOI: 10.1038 / s42256-018-0005-0