Approaching Artificial Intelligence as an Embodied Developmental Process
presentationposted on 2020-10-28, 18:18 authored by Bradly AliceaBradly Alicea, Rishabh Chakrabarty, Akshara Gopi, Anson Lim, Furkan Özçelik, Jesse ParentJesse Parent
In this submission, we will lay out an argument for building artificial intelligence (AI) based on biological development. A developmental AI consists of three components: an embodied input/output network, a critical period for information acquisition and connectivity, and a network topology that expands over time according to the principles of contingency and developmental freedom.
Developmental contingency restricts the developmental neural network to an increasingly limited set of possible trajectories. While contingency results in earlier
developmental events (cell differentiation) to constrain the range of possibilities for future developmental events (formation of a network motif), developmental freedom
is the tendency for information and representations in particular to take advantage of a stable underlying topology. This contributes to an artificial brain that acts as a
generative network with a further capacity for behavioral complexity.
We will demonstrate examples for each of these points using developmental Braitenberg Vehicles (dBVs) as the subject of thought experiments. These thought experiments will take the form of a contingency analysis of a typical dBV input/output network topology. This analysis shows how embodiment developmental agents provide three potentially powerful mechanisms for the emergence of adaptive
intelligent systems. First, the presence of an embodied input/output network can act to structure input data. Secondly, the existence of a critical period embedded within the developmental process acts to structure association-building through network complexity. Finally, developmental freedom results from the placement of the critical
period within a sequence of events defining the biological developmental process, which in turn can enhance both information acquisition and network structure. In
conclusion, we will discuss potential benchmarks and future advances in the developmental approach, including the application of epigenetic landscapes to characterize the formation of a neural network topology.