Artificial Life	Neural Dynamics	Ph.D. Dissertation	By Institution

Toward decentralized, autonomous systems

The trend toward segmentation and distribution of large systems over a multitude of smaller and relatively simpler components is both a growing need and an inevitable fact in many domains of computer science & engineering, artificial intelligence, and robotics. Faced with an explosion in size and complexity of computing systems at all levels of organization, including hardware (integrated parts), software (program modules), and networks (applications and users), engineers are gradually led, more or less willingly, to rethink these systems in terms of complex systems. Computing systems must transition from a state of exogenously imposed order toward increasing organizational and functional autonomy. Instead of rigidly designing systems in every detail, engineers need to step back and only "meta-design" these systems, i.e., focus on the generic conditions allowing their endogenous growth, function, and evolution. Looking around, we can observe an abundance of autonomous systems in the environment, whether in nature (cells, organisms, societies of organisms) or spontaneously emerging human superstructures (societies, economy, Internet). Decentralized, unplanned, emergent systems are probably the most pervasive and efficient type of systems. It is centrally planned systems that are unique and costly to build, as they require another intelligent system to exist and operate. In fact, "complex" systems might well turn out to be the simplest. In this context, natural adaptive systems, in particular biological (developmental, neural, evolutionary) and social (complex networks), have an important role to play in providing a powerful source of inspiration for emerging technologies. Understanding these systems could help create a new generation of artificial systems based on paradigms still largely absent from traditional engineering, such as decentralization, autonomy and adaptation.

Complex systems at multiple scales

Complex systems are generally defined as large networks of elements that interact locally and produce an emergent collective behavior at a macroscopic scale. They are characterized by a high degree of decentralization, and the ability to self-assemble and self-regulate. Most complex systems are also adaptive, in the sense that they are able to "learn" and evolve toward further innovation by feedback from their external fitness to their internal architecture. The agents or parts composing a complex system follow local rules that can be more or less sophisticated (see Table). Often, these rules are themselves internally structured as networks of smaller entities. For example, one cell can be modeled as a self-regulatory network of genetic switches, one social agent (ant, software process) as a network of decision rules, one neural unit as a local assembly of neurons (dual excitatory/inhibitory oscillator system, synfire chain). Conversely, agents can also interact collectively at the level of clusters or subnetworks (organs, assemblies, cliques) that combine in a modular fashion to form larger collectives. Thus, from both perspectives, complex systems can often be described as "networks of networks" on several hierarchical levels. The higher levels connecting elements or clusters of elements are generally spatially extended (cell tissues, cortical areas, ant colonies, computer network), whereas the lower levels inside elements are generally nonspatial (gene nets, neural assemblies, rule trees). Elements follow the dynamics dictated by their inner networks and also influence neighboring elements through the emission and reception of signals (chemical, electrical, software packets).

Reintroducing programmability

To be able to apply complex systems to engineered systems, the most important challenge is not simply to allow self-organization to happen, but first and foremost to guide it. Past a certain fascination for spontaneous "free" order (such as random patterning or flocking), the next critical research question concerns the reintroduction of programmability and reproducibility into self-organization, i.e., the attempt to re-engineer emergence. In other words, the relevant models for future emergent ICT (Information and Communication Technologies) will probably not be found in the traditional "statistical" complex systems, such as random patterning or flocking, but rather in "morphological" complex systems, such as biological development and social structures. The difference between these two classes of systems resides in the relative sophistication of the elements and their ability to combine in sufficiently various ways to form precise and reproducible architectures (see Figure, green and orange frames, respectively).

Naturally, this seems to lead to paradoxical objectives: Can autonomy be planned? Can decentralization be controlled? Can evolution be designed? The answer to this puzzle lies in a change of scale: instead of a top-down enforcement of macroscopic structures, the new controls take the form of local instructions inside every microscopic agent of the system. These instructions can also diversify, depending on the agent types and positions, introducing the degree of heterogeneity in the system required for any complex behavior beyond patterning and flocking. It is an often underappreciated ability of complex systems to exhibit controllable properties, at the same time (or despite the fact) that they are self-organizing. It seems that "complex" is too commonly construed as "homogeneous", "monolithic" and/or "random". Yet, there can be a wide diversity of agents and heterogeneity of patterns, via positions; a complex system can be modular, hierarchical, and architecturally detailed at multiple scales; it can also consist of reproducible patterns arising from programmable agents.

My research

My work currently addresses two domains: (a) neural computation, where a central question is to understand how the symbolic level of cognition (AI) can arise from the underlying complex dynamical system of the brain (neural networks); and (b) artificial life, preoccupied with explaining how complexity and fitness can spontaneously develop and evolve without the need for a higher symbolic level. In other words, the cognitive challenge I pursue consists of reconstructing the emergence of the human symbolic faculty to help create intelligent machines. The engineering challenge is rather about removing the symbolic human bias from intelligent system design and creating autonomous efficient systems that could grow and adapt without explicit programming. These two challenges are naturally closely linked and their concepts exchangeable

Summary of Research Topics

Artificial Life — Biologically Inspired Engineering "Meta-designing" the development, function and evolution of self-organized complex systems that do not use a symbolic level How do embryonic cells construct an entire organism without a blueprint map? How do complexity, innovation and fitness spontaneously evolve (without a "watchmaker")? How can biological organisms inspire a novel engineering paradigm based on decentralized, self-adapting collectivities of agents, instead of explicit rules and external design? How can we create a network of agents that would spontaneously diversify, multiply and self-organize to work collectively on a given task (e.g., swarm robotics, immune security)? Keywords: artificial development, self-assembly, pattern formation, spatial computing, evolutionary computation multi-agent and cellular automata models of morphogenesis, based on gene regulation networks decentralized but programmable pattern formation, network self-assembly and shape development spatially extended cellular automata models of population genetics, evolution and ecology Neural Dynamics — Large-Scale Spiking Neural Networks Understanding and reconstructing the emergence of a symbolic level from a complex dynamical system How is the infinite diversity of analog (visual, auditory) stimuli segmented, grouped and reduced to a few logical categories? How is discrete symbolic meaning "carved out" from the continuous physical environment? How are neural signals organized in the brain and what kind of complex coordinated (and reproducible) spatiotemporal patterns do they form? How does this pattern formation of a spatiotemporal kind provide the basis for structured "mental objects" and their hierarchical composition? Keywords: mesoscopic level, segmentation, schematization, categorization, perception, language, ontology mesoscopic emergence and interaction of spatiotemporal patterns of activity and connectivity based on: stochastic-firing, excitable, oscillatory and subthreshold neuron models creating: synchronization, traveling waves, coherence induction, synfire chains, compositionality for: segmentation, schematization, pattern recognition & categorization in perception and language