An Analysis of the Epistemological Foundations of Machine Learning

Garibaldi Sarmento, Associate Professor in the Department of Philosophy at the Federal University of Paraíba (UFPB), João Pessoa, PB – Brazil.

Arão in his article titled Behind Artificial Intelligence: An Analysis of the Epistemological Foundations of Machine Learning (2024), states that “[…] it will be demonstrated how induction and mathematization work as the epistemological basis of artificial intelligence and how some of the limitations of this technology can be explained through the weaknesses of the methods that support it” (Arão, 2024, p. 01, italics mine). 

Nevertheless, the author is not, strictly speaking, presenting a demonstration in the said article, as we should, for this purpose, consider both a logical-mathematical analysis and a methodological refutation from the perspective of computer science, as emphasized in Sarmento (2004). In principle, the noted “weaknesses” are subject to correction if we a priori admit the construction of a maximally consistent system as the base of knowledge for the learning machines model. 

These observations are presented in the Commentary on “Behind Artificial Intelligence: An Analysis of the Epistemological Foundations of Machine Learning”, recently published in the journal Trans/Form/Ação, whose reflections will be summarized here.

In an article titled “Artificial Intelligence, Undecidability, and Trans-computability” (see Sarmento, 2004) I proposed a critical approach to the logical-epistemic foundations of Artificial Intelligence, in which some formal and technical aspects – that instantiate and delineate the field of investigations encompassing Artificial Intelligence – were addressed.

Without proposing a technical and formal discussion in this essay, I present, in general terms, some points directed towards the issue of using an inductive-probabilistic logic for the “Large Language Models” (LLMs) built based on “neural networks,” which, in turn, enable the machine learning process.

Firstly, some of the areas of computer science directly related to the development of LLMs are computational linguistics, computational neuroscience and quantum computing and, especially, the algorithmic theory of information. In all the highlighted areas, algorithmic methods of statistical inference are widely used. 

One of the subfields of the algorithmic theory of information is called Kolmogorov complexity, which deals with essential concepts for statistical-probabilistic methods of inference. This field of study involves a rigorous treatment of the concept of randomness, utilizing algorithmic methods such as, for example, the “Solomonoff probability” algorithm.

Another specific method for machine learning systems, via neural networks, is the use of the so-called Bayesian Logic. In general, there are various algorithmic methodological techniques that allow for a rigorous (and subject to correction) computational treatment of probabilistic-statistical approaches applicable to the development of AI systems.

Imagem: Freepik.

From a philosophical perspective, the theory of knowledge compatible with the development of these AI systems is the reductionist epistemic approach proposed by Lehrer (1990), in which the concept of knowledge is strictly reduced to the concept of information. Lehrer admits that the explanation of knowledge, in the sense of information, assumes that:

1. “P” accepts the information that “s”.
2. The information that “s” is correct
3. The acceptance of the information that “s” is justified.

Here, “P” and “s” are variables, “P” designates any individual, and “s” denotes any information that can be obtained. Thus, “P knows that s” expresses the “recognition of the correctness of the information.” The justification (or corroboration) of the information is the central aspect of the epistemic analysis, given that this procedure should, in principle, be understood as a “measure” between “strong reasonableness and not complete certainty,” as Lehrer argues.

This means that justification, from the perspective of inductive-probabilistic logic, is bounded between reasonableness and complete certainty (the concept of “certainty” here means strong evidence of a probative nature, given the sufficient conditions at a certain time “t”, for the confirmation of the correctness of the information). In probabilistic terms, the closer the correctness of the information is to 1, the higher the degree of certainty of its confirmation.

The algorithmic theory of information (a subfield of information theory), which lies at the intersection of Turing’s computability theory and Shannon’s classical information theory, could serve as a computational foundation for the approach advocated by Lehrer, when it comes to the concept of information and, consequently, justify that LLMs process (and generate) knowledge about the “physical world” in a manner analogous to humans.

The main philosophical issue, in this case, shifts to the level of autonomy that AI cybernetic systems could have to act independently of human intervention. This is a question that undoubtedly requires deeper reflection.

Up to the present moment, however, these cybernetic systems do not effectively exhibit a capacity for self-correction and decision-making in an interactive and autonomous manner.

Those who advocate for such an approach, i.e., that AI cybernetic systems can generate knowledge, cannot ignore that AI will always be limited — with respect to its cognitive capacity — due to Chaitin’s incompleteness theorem (a version of Gödel’s incompleteness metatheorem). And, consequently, AI cannot surpass the cognitive capacity of humans.

About Garibaldi Sarmento

He has a doctorate in philosophy from Unicamp, SP, and is an Associate Professor in the Department of Philosophy at the Federal University of Paraíba (UFPB), João Pessoa, PB – Brazil. Research Areas: Mathematical Logic and Philosophy of Mathematics.

References

ARÃO, C. Behind artificial intelligence: an analysis of the epistemological bases of machine learning. Trans/Form/Ação [online]. 2024, vol. 47, n. 3, e02400163 [viewed 15 August 2024]. https://doi.org/10.1590/0101-3173.2024.v47.n3.e02400163. Available from: https://www.scielo.br/j/trans/a/wKP3thTz35fmhG9pnmXNdMj/

LEHRER, K. Theory of Knowledge. New York. Westview, 1990.

SARMENTO, G. Inteligência artificial, indecidibilidade e transcomputabilidade. Perspectiva Filosófica. 2004, vol. I, no. 21, pp. 179-209.

To read the article, access

SARMENTO, G. Comentário a “Por trás da inteligência artificial: uma análise das bases epistemológicas do aprendizado de máquina”. Trans/Form/Ação [online]. 2024, vol. 47, no. 3, e02400221 [viewed 15 August 2024]. https://doi.org/10.1590/0101-3173.2024.v47.n3.e02400221. Available from: https://www.scielo.br/j/trans/a/Zk3GNmcnCHVmCkDRdWvXcRj/ 

External links

Trans/Form/Ação – TRANS: https://www.scielo.br/trans/

Trans/Form/Ação – Journal: Instagram | Facebook | Academia.edu

Garibaldi Sarmento – ORCID: https://orcid.org/0000-0002-9242-3945