Machine understanding and deep learning representation

Abstract

Practical ability manifested through robust and reliable task performance, as well as information relevance and well-structured representation, are key factors indicative of understanding in the philosophical literature. We explore these factors in the context of deep learning, identifying prominent patterns in how the results of these algorithms represent information. While the estimation applications of modern neural networks do not qualify as the mental activity of persons, we argue that coupling analyses from philosophical accounts with the empirical and theoretical basis for identifying these factors in deep learning representations provides a framework for discussing and critically evaluating potential machine understanding given the continually improving task performance enabled by such algorithms.

Towards a Benchmark for Scientific Understanding in Humans and Machines

“…Although our focus is on scientific understanding in the natural sciences (e.g., physics), we anticipate that our framework can be applied to other scientific disciplines as well. We break with the traditional view (Dellsén, 2020;Wilkenfeld, 2013;Searle, 1980;Johnson-Laird, 2010;Nersessian, 1992) by arguing that understanding should be conceptualized in terms of abilities rather than internal mechanics (Marcus, 2018;Chollet, 2017) or representations (Tamir & Shech, 2023;Wilkenfeld, 2013). Specifically, we contend that scientific understanding is a skill-based capability that relies on an agent's ability to perform specific actions, rather than a subjective mental state.…”

“…We maintain that the evaluation of understanding in any agent, including artificial ones like LLMs, should follow the same principles used to assess human scientific understanding -that is, it should be based on their abilities to perform relevant tasks. Incidentally, we are not the only ones who relate understanding to an ability (see Krenn et al, 2022;Tamir & Shech, 2023). For example, Tamir and Shech (2023) have argued that practical abilities (such as reliable and robust task performance) can be seen as key factors indicative of understanding in the context of deep learning.…”

“…Incidentally, we are not the only ones who relate understanding to an ability (see Krenn et al, 2022;Tamir & Shech, 2023). For example, Tamir and Shech (2023) have argued that practical abilities (such as reliable and robust task performance) can be seen as key factors indicative of understanding in the context of deep learning. While we think this is a good start, we argue that a more comprehensive and rigorous evaluation of understanding as an ability is needed.…”

Reliability and Interpretability in Science and Deep Learning

“…based on well-defined assumptions). Hence, it is largely unrelated to the literature that investigates the sociological basis for the trust in ML models, or analogies between ML and human behaviours [see e.g., Clark and Khosrowi (2022), Duede (2022), Tamir and Shech (2023)]. Further comparisons with the literature are provided in the main text.…”

Искусственный интеллект и технологии Big Data (больших данных)

“…Утверждая революционный характер влияния на общество и человека AI и Big Data, нельзя не отдавать себе отчета в том, что понятия искусственного интеллекта, глубокого обучения (deep learning), нейронных сетей и т. п. в значительной мере были и еще будут довольно долго оставаться метафорами. Хотя алгоритмы глубокого обучения являются едва ли не ключевыми в создании AI и использовании Big Data, но методы понимания текста (и тем более контекста) основаны на механическом переборе словаря, который должен постоянно расширяться (Tamir & Shech, 2023). В мозге человека, способном действительно понимать, 85 миллиардов нейронов, а на настоящий момент наиболее успешная программа нейронной сети, призванная смоделировать функции мозга, способна воссоздать мозг примитивного червя, который состоит из чуть более чем 300 нейронов.…”