AI & My Cat Z: An Obituary

Disclaimer: This post is not on yet another cat image recognition in AI. However, on something within the core of the matter: correlation and inference in AI.

Past three months, ut has been both emotionally and intellectually painful. Due to some unknown cause, I have lost my cat Zeytin Zidane (Z).?He was a one-eye-patched-pirate looking yet a very sweet and affectionate companion. He has been all around with me within the last 14 years of my life: Three different countries, five homes, several each long and beautiful relationships and a lot more that we went through together.

There have been plenty enough experiences to understand his learning patterns, adaptation skills and emotions.?He would have very consistent traits that he would form through recurrence and persistence. Trust was one of them. We established it mutually, and it was always there. We would forgive each other easier, sharing the space would calm us and give peace. However, in the last three months of his life, he was continuously updating his other behaviors to avoid pain, which often times made me think about intelligence.

Correlations dominated his inferences. He would change his attitude toward objects, spots to hangout within the house or garden, even toward the taste or the smell of his food that he used to like would change.?For instance, when his pain attacks would start coinciding with his current resting spot, after just a few number of times in a row, he would avoid that very spot.

Such learning pattern is very similar to the statistical learning principle that dominates the vast majority of current AI. In most of these AI design systems, training data sets are formed to be able to count co-occurrences of two given events or attributes, which in return are used to form correlation matrices. Unfortunately, these correlations are used to infer causal relationships.

Deep learning in AI may sound different but it is essentially within such a paradigm. It does a better job however by adding more variables and layers. Variables are represented by number of nodes in neural networks. Layers enable to sequence events or variables preceding an outcome, such as the pain that Zeytin was suffering. In other words, deep learning to a large extent is just a sophisticated statistical inference. Instead of looking at correlations of two events at a time, it counts observed combinations and sequences of events that leads to an observed symptom or any other type of outcome.

Reasoning in some past and present times humans are not so different than the reasoning my cat Z had under extreme stress of physiological pain. Under anxiety led short-sightedness jumping to immediate conclusions via associations often leads to useless or damaging fears. Some groups in our current societies associate an increase in unemployment rates and arrival of refugees and conclude that refugees are taking jobs from the locals. In my experience, it was so hearth breaking to see that Z started to refuse food and water due to proceeding experience of pain.?

As AI and algorithms are becoming more and more prevalent that has significant effect on our societies and nature, we need to invest more in other forms of learning. For instance, at two of Lucidminds AI’s recent projects (Trees AI and DataVille), we employ agent based paradigm to include simple rule sets, social norms, or models from natural sciences. The paradigm enables us to make sense of the complex systems where labeled data do not exist, or problematic. Using the paradigm we run social and environmental scenarios relevant to data economy or for instance air quality within urban areas.

We will be sharing more insights and learnings from our firsthand experiences.

Bulent Ozel, PhD.