Funded Project (2023-2025) BP Pragmatist Intelligence
Collectives like atoms, cells, swarms, and societies feature a remarkable capacity to organize and create entities with emergent and smart properties. Using different system components to reach a common goal evokes a concept currently vividly debated: intelligence. Intelligence is enabled through a complex interplay between molecular and elementary components, as notably in neural networks in the human brain. Yet, the definitions of said interplay are often quite loose. Creating artificially intelligent systems, however, requires a robust concept of intelligence.
Artificial intelligence aims to emulate abstracted structures of the brain, targeting not only brain-like algorithms but, eventually, also brain-like or neuromorphic hardware. Current research revolves around the question of whether functional materials could give rise to biological intelligence, thus creating intelligent matter, capable of, for instance, emulating artificial neural networks (ANNs). This emulation, however, focuses on structures identified as behavior in form of a functionality, creating an obvious difference between brain-like materials and the brain.
Traditionally, intelligence has been associated with reasoned, and therefore human behavior. Now, intelligence as the aforementioned emulation is increasingly applied to non-cerebral and non-living systems. This raises the question whether this evidently different machine intelligence produces a functionality that deviates from the desired behavior. The way ANNs select and process features of the perceived objects has a crucial influence on the system’s functionality. Selection and processing govern significantly what the system can eventually perceive though its epistemic access to reality.
Our aim is to investigate a viable approach that characterizes this epistemic access. In a first step, we focus on the lack of feature binding in ANNs and elucidate how this lack influences a machine’s representation of reality and the capability to extrapolation. The second part of the project concerns the principles behind the selection of features introducing pragmatist approaches. Philosophical pragmatism is often the implicit paradigm in engineering. Features selected in ANNs follow the principle of usefulness, i.e., they dismiss all parameters or features not useful or endangering the approximation toward the optimal sate. In viewing ANNs that way, we can shed light on the proximity between pragmatist theories and the paradigm governing the functionality of this kind of intelligence: consequentialism. Consequentialism sanctions as right those actions that have the (relatively) best consequences lining these theories up with theories of optimization.
Bridging the gap between physics of machine learning and philosophy, within this interdisciplinary project we aim to develop a novel approach that describes the selection, processing, and integration of features (e.g., a visuo-haptic perception), and the kind of intelligent matter it creates.
