Artificial intelligence and Darwinian evolutionism
The latest innovations in the field of artificial intelligence are programs such as AutoML-Zero, which allows a population of randomly generated algorithms to compete with each other and identify the best in each cycle or iteration or “evolutionary phase”.
The human being occupies an important place in the production cycle of artificial intelligence (“Human-in-the-loop Artificial Intelligence” – HitAI). As a result, the output is still conditioned by the creativity and progress of the data scientists, and affected by their prejudices (Gent, 2020). However, in frontier research in the field of artificial intelligence (AI), the “man productive factor” tends to vanish and be progressively replaced by the machine. Human input, in this context, becomes obsolete compared to other productive factors.
Along this direction, a promising new field of application is Darwinian-type evolutionary biology: AI can be instructed to evolve on its own, that is, without human input.
The evolutionary theory of natural selection, developed at the beginning of the second half of the 19th century by Charles Darwin, could thus represent a starting point for the creation of more advanced AI, capable of evolving on their own to arrive at results never before achieved.
In general, it is known that building an algorithm takes a long time. Consider, for example, a type of machine learning used for the development of autonomous driving techniques. Neural networks imitate the structure of the human brain in an elementary way and learn what to do through the “training” data, thus strengthening the connections between their artificial neurons. As a rule, we proceed by designing neural “sub-networks” dedicated to specific tasks – for example understanding road signs – which are then connected together to collaborate avoiding – it is appropriate to say – path accidents. But the road is long …
A Google computer scientist, Quoc V. Le, along with other researchers, tried to find a faster and more efficient way to implement the algorithms. The AI program called AutoML-Zero – with zero “human inputs” – has managed to replicate decades of AI research in just a few days (Gent, 2020). Thanks to mechanisms of variation, inheritance and selection inspired by biological evolution, Le’s AutoML-Zero is able to improve itself, replicating itself from generation to generation in versions more and more suitable for carrying out a specific task assigned to it – to example, distinguish a cat from a truck (Saturday, 2020). The AutoML-Zero program selects the algorithms through an approximation of the evolutionary process in nature. In Darwinian terms, algorithms capable of adapting better to the external environment survive thanks to progressive mutations (“survival of the fittest”). The environment is complex, uncertain, lacking in information, evolves over time, is subject to shock, is subordinated to the complex of geoclimatic conditions, depends on the set of other beings with which each one comes into contact and interacts. Genetic mutations have enabled human beings – and more generally all forms of life – to survive shocks such as drought, famines, diseases and disasters of various kinds, namely what Darwin called “living conditions”. A very current example is offered by the study (2020) conducted by the Cnr, together with other institutions, and published in the Advances Sciences. It shows how our organisms have evolved innate cellular immunization processes capable of “hacking” the genetic code of Sars-CoV-2 through a particular process known as “editing” of RNA (in chemistry it is the acronym for ribonucleic acid, an enzyme involved in various biological roles of coding, decoding, regulation and expression of genes. In molecular biology, editing constitutes a set of molecular processes that result in a chemical modification of RNA, adapting it to new survivors events). Therefore, the living being tries to adapt to changes in the external environment through a long and tiring process of “groping”, “trial and errors” which result in genetic mutations. When Darwin’s idea is transposed into the field of AI,
The idea is to make a population of randomly generated algorithms compete with each other and identify the best in each cycle or iteration or “evolutionary phase”.
In particular, through very simple mathematical operations, the software starts by creating a population of 100 candidate algorithms to evolve. AutoML-Zero verifies them by making them perform elementary activities, such as recognizing if a certain image corresponds to that of a mouse or a truck (Gent, 2020). This process unfolds in cycles, analogous to the various phases of the biological evolutionary process or, in terms of Game Theory, to the various rounds of a repeated game. The first tests based on the recognition of some images seem to confirm the effective functioning of the new system.
For each cycle, the AutoML-Zero program compares the performance of these algorithms compared to the performance of hand-designed algorithms and selects the best performing within the first group. Through random mutations, the program makes copies of the latter. The new algorithms that arise from this process – that is, those belonging to the next generation – feed the starting population, to the detriment of the more obsolete algorithms of said population. This iterative process therefore continues with a new cycle – just like in an evolutionary process of society – which will follow the same pattern and with a population that is constantly changing. Thus AI programs improve from generation to generation without human external instructions. In the long run the self-generating algorithm can become the best available, surpassing those designed by hand. This succession of generations therefore evolves towards an optimal solution of the assigned problem.
The same admits that today this approach behaves uncertainly on a number of classic machine learning techniques. The solutions he identifies are simple with respect to the most advanced algorithms that already exist, but his study (Real et al., 2020, on the arXiv online archive, which collects the studies pending approval and publication by the scientific community. ) is intended to be a conceptual demonstration, prodromal to new more complex AI that can develop along two directions: the first is to focus on smaller problems instead of an entire algorithm, the second is to expand the battery of mathematical operations and devote more computing resources to AutoML-Zero.
The potential results are currently difficult to predict, but we can certainly say that we are spectators of rapid advances that raise increasingly pressing ethical, philosophical and anthropological reflections. A trade-off is indeed in sight: progress on the “human-in-the-loop”, with the disappearance of man in the process, if they can constitute an indicator of a gain in terms of AI efficiency, could get out of hand , outclassing our own intellectual abilities and with unpredictable effects. And how long will it take to get to all this? Beyond the current and future controversies on this essential issue, the answer to the question is even more urgent: if a superhuman AI appeared, would it be a good thing? Hence a fascinating analysis of the possible scenarios: a superintelligence that coexists with humans; that is to replace them entirely (Tegmark, 2018). To what interactions and dynamics will this coexistence lead? Also in this case, through an iterative process of adaptation of the human factor, Darwinian evolution can give an answer.