Workshop

9:00AM - 9:20AM Check-in

Goal-directed behavior: overarching perspectives (chair: Richard Löffler)

9:20AM - 9:45AM Miguel Aguilera: introduction and overview of the “Goal-directed behavior and the origin of life” project

9:45AM - 10:30AM Xabier Barandiaran: What is goal directed behaviour? From biological autonomy to intentional action
 

10:30AM - 11:00AM Coffee break

11:00AM - 11:45AM Ben de Bari: Emergent normativity: origins of goal-directedness in far-from-equilibrium self-organization

11:45AM - 12:30PM Tomás Veloz: Modeling the origins and evolution of goal-directed systems: a reaction network approach

12:30PM - 1:30PM LUNCH

Droplets and active matter  (chair: Christian Euler)

1:30PM - 2:15PM Richard Löffler: Observation of complex behaviors in simple dissipative systems

2:15PM - 3:00PM Yuki Koyano: Self-propelled motion of camphor on a water surface interacting with system boundaries

3:00PM - 3:30PM  Coffee break

3:30PM - 4:15PM Ágota Tóth: Synchronization of hydrogel beads

4:15PM - 5:00PM Jitka Cejkova: Manipulating goal-directed behaviour in Dictyostelium discoideum with external signals

5:00PM - 7:00 PM  Free time, discussions, walks, etc.

7:00PM  Pintxos downtown (optional)

Viability (chair: Martin Biehl)

9:00AM - 9:45AM Nathaniel Virgo: Goal-directed behaviour and meaningfulness: towards a mathematical framework

9:45AM -  10:30AM Connor McShaffrey: Foundations of viability space

10:30AM - 11:00AM coffee break

11:00AM - 11:45AM Denizhan Pak: Metastable goals and their possible mechanisms

11:45AM - 12:30PM Yusuke Himeoka: A theoretical basis for cell deaths

12:30PM - 1:30PM    LUNCH

Cells and metabolism (chair: Omer Markovitch)

1:30PM - 2:15PM Daniele De Martino: The metabolic goals of cells: lesson from single cell flux analysis

2:15PM - 3:00PM Christian Euler: Thermodynamics constrains information flow in metabolism

3:00PM -  3:15PM coffee break

3:15PM - 4:00PM Eran Agmon: Emerging goals: A compositional framework for modeling biological agents

Discussion / collaborative time                                            

4:00PM - 5:30PM      Open discussion / collaborations

Goals and Agency (chair: Artemy Kolchinsky)

9:00AM - 9:45AM Manuel Baltieri: On the interplay between goals and action-orientedness

9:45AM - 10:30AM Fernando Rosas: Exploring the informational architecture of agency

10:30AM - 11:00AM coffee break

11:00AM - 11:45AM Hiroyuki Iizuka: The origin of intentions: from others to self through superposition

11:45AM - 12:30PM Simon Mcgregor:  The case for “pansolipsism”

12:30PM - 1:30PM LUNCH

Goals in computational systems (chair: Eran Agmon)

1:30PM - 2:15PM Hiroki Kojima: Persistence of complex patterns in Lenia and their physical interpretation

2:15PM - 3:00PM Kazuya Horibe: Acquisition of implicit world models via domain adaptation in homeostatic meta-reinforcement learning

3:00PM - 3:15PM coffee break

3:15PM - 4:00PM Paul Colognese: ​​Goal detection for AI catastrophe prevention

Wrap up                                                                                                                                          

4:00PM - 5:30PM Open discussion / collaborations

Basque Center for Applied Mathematics

  Practical information: link

  Address (Google maps)
  Alameda Mazarredo 14
  48009 Bilbao
 

WhatsApp group for social events

 https://t.ly/3wufj

Title: Emerging Goals: A Compositional Framework for Modeling Biological Agents

Abstract: Traditional systems biology models often lack the modular and compositional nature inherent in their biological counterparts. Simulators are typically written as monolithic, preset structures, and the models they simulate focus on the dynamics of specific biological subsystems, often in controlled conditions. To address this, I have proposed a compositional framework for systems biology that emphasizes the interfaces, interconnections and orchestration of subsystems, allowing for the natural emergence of system-level behaviors from lower-level interactions.  Compositional systems biology aims to connect models of different subsystems into integrative multiscale simulations, bridging processes across molecular, cellular, and organismal levels, each with distinct spatial and temporal dynamics.

In this talk, I will present a compositional perspective on building multiscale models of minimal biological agents, and demonstrate how this can account for goal-directed behavior. The framework emphasizes the importance of designing models that capture the underlying principles of biological systems, including how goals and viability boundaries are dynamically managed through the cellular interface. By rethinking model structures to allow goals to emerge naturally, compositional systems biology enhances our understanding of cellular functionality, improving our ability to predict and influence biological outcomes in complex environments.

Title: On the interplay between goals and action-orientedness

Abstract: Action-oriented models are a construction posited to (partly) capture notions of embodiment in probabilistic frameworks used to model agents in situations involving various degrees of uncertainty.

While of interest, their definition is still quite informal, with works in philosophy of mind struggling to settle on clear requirements, and simulations in artificial life/reinforcement learning focusing on specific, mostly idiosyncratic examples that don't necessarily capture more general patterns.

In this talk I will discuss some work in progress where I try to formally define a class of action-oriented models using a construction from theoretical computer science, bisimulations, that provides different ways to compress problems, or tasks, presented as (PO)MDPs. Some of these ways include specific accounts of actions, policies and the goals of an agent, and in light of this, I believe bisimulations may thus provide a formal understanding of action-orientedness.

Furthermore, while typically hard to compute, different definitions of bisimulations can be approximated and implemented in modern (deep) reinforcement learning frameworks, where we find already several examples of agents with the ability to solve tasks by compressing a problem to only consider "task-relevant" information, thus making it, I would argue, not only an appealing theoretical construction but also a useful definition with direct practical applications.

Title: What is goal directed behaviour? From biological autonomy to intentional action

Abstract: Goal-directed behavior is a central theme in understanding both biological and cognitive processes. This talk explores the concept from an organismic perspective, focusing on how autonomy underpins the emergence of goal-directed behaviour. Drawing from contemporary enactive and organizational apporaches to autonomy, we will examine how goal-directedness is a natural phenomenon that can be understood to span from simple metabolism-driven behaviour to complex intentional action. Starting with the foundational notion of autonomous systems, I will trace the evolution of adaptive behaviour from biological self-maintenance, to sensorimotor habitual to fully intentional (yet pre-linguistic, non-intelectualist) action, emphasizing the dynamic presuppositions that govern sensorimotor coordination patterns. Key to this analysis is the hierarchical and nested structure of goals within biological organisms, which allows for adaptability and error correction. By examining both normative functionality and behavioral errors, we uncover the minimal requirements for defining teleological behavior in organisms.

Title: Emergent Normativity: Origins of Goal-Directedness in Far-From-Equilibrium Self-Organization

Abstract: Systems open to flows of energy and matter will sometimes display spontaneous spatial or temporal pattern formation. These emergent dynamics are called dissipative structures, maintained by irreversible entropy-producing processes. Some such systems display a quality we call emergent normativity, a sensitivity to or "preference" for different states of affairs. When a system displays emergent normativity, these intrinsic norms serve as implicit goals of the system, and we can reliably predict and explain their behavior in terms of end-directedness towards these goals. This talk will briefly outline the circumstances under which emergent normativity is manifest, with examples from non-living dissipative structures. Contrasts will also be made between this emergent normativity and other forms of end-directedness in physics.

Title: Comparing Goal-Directed Behavior in Dictyostelium discoideum and Liquid Robots: Chemotaxis, Shape-Changes, and Collective Dynamics

Abstract: I will focus on comparing the behavior of living organisms, specifically the slime mold Dictyostelium discoideum, with liquid robots in the form of droplets. Both systems exhibit goal-directed behaviors, including movement, self-organization, shape-changing, and collective dynamics. While Dictyostelium relies on chemotaxis to aggregate under stress, liquid droplets exhibit lifelike characteristics, including movement, self-division, and coordinated group behavior akin to living organisms. The discussion will focus on both the similarities and differences between these systems.

​​Title: Goal Detection for AI Catastrophe Prevention

Absrtact: As AI agents advance in capabilities, they promise immense benefits but also pose significant risks. If we build powerful AI agents that pursue misaligned goals, the consequences could be catastrophic. Training AI to consistently pursue aligned goals is challenging for various reasons, including the potential for deceptive AI agents to actively mislead oversight processes. This talk explores whether goals might be represented in the computational substrates running the agent and how we might leverage these representations for goal oversight, thereby contributing to the prevention of potential AI-driven catastrophes.

Title: The metabolic goals of cells: lesson from single cell flux analysis

Abstract: Is cellular metabolism optimized, for instance, to maximize growth? This assumption is common in system-level approaches such as Flux Balance Analysis (FBA), which provides useful insights and is theoretically supported by competitive evolutionary dynamics. However, I will demonstrate that this hypothesis is not backed by flux data, even under ideal conditions in E. coli experiments. Moreover, it conflicts with the well-established observation of single-cell heterogeneity. By incorporating this into models using statistical methods like maximum entropy (MaxEnt), we can quantitatively explain the flux data. Extending this approach to full inverse modeling reveals that the linear objective function framework of optimization offers limited predictive power and comes at the expense of explainability. Heterogeneity, on the other hand, allows for intercellular exchange interactions, a crucial factor often overlooked in standard models. I will present recent experimental findings from single-cell flux analysis that demonstrate this point:

1) Tumor-stroma co-cultures: These systems either collectively maintain the medium's homeostasis or fail to do so, manifesting the Warburg effect. Statistical physics modeling of this data reveals that the Warburg threshold represents a formal phase transition, where the key driver of acidification is a lack of coordination, rather than local hypoxia or mitochondrial saturation.

2) Cyanobacterial metabolic ecology: In a simple flask, cyanobacterial cells exhibit  complex metabolic behaviors, forming sub-clusters that alternate between optimal nitrogen and carbon yields. There are also acid exchanges, and extreme cells with metabolic rates double the average, as well as ‘fixer’ cells, which exhibit slow growth but maintain a high metabolic rate.

These findings highlight the limitations of traditional optimization approaches and the need to account for cellular interactions to really understand metabolic systems.

DDM et al (2018) Nat. comm., 9(1), 2988.
Muntoni et al (2022) Biophysical Journal 121.10: 1919-1930
Onesto et al (2023) ACS Nano, 17, 4, 3313–3323
Narayanankutty et al. (2024) arXiv:2405.13424

Title: Thermodynamics constrains information flow in metabolic networks

Abstract: Metabolism is an open thermodynamic system characterized by mass, energy, and information exchanges with the environment. Many environmental and internal signals are initially processed by metabolic reactions before they can influence transcription and translation. Yet, the capacity of metabolic networks to process and transmit information remains largely unexplored. Here, the linear noise approximation (LNA) is applied to a reversible, enzyme-catalyzed reaction to estimate its channel capacity and constraints thereon. It is shown that perturbations to metabolite concentrations cannot be propagated through reactions operating away from equilibrium. As a result, thermodynamically unfavourable reactions - such as those required for anabolism - are structurally constrained to limit information flow in metabolic networks. However, the substrates of such reactions can maximally transfer information via non-reactive binding, as in regulatory interactions. This provides a mechanism by which information transfer can be maintained across metabolic conditions. Bioinformatic evidence supporting the hypothesis that metabolism has evolved to conserve information flow in varied conditions will be presented, and implications of these results for goal-directedness within the metabolism-first hypothesis of the origin of life will be discussed.

Title: A theoretical basis for cell deaths

Abstract: Comprehending cell death is one of the central topics of biological science. Currently, the criteria for microbial cell death are purely experimental, based on PI staining and regrowth experiments. In the present project, we aimed to develop a mathematical framework of cell death based on the metabolic state of the cells.

Our attempt is to develop a theoretical framework of “death” for cellular metabolism [1]. We start by defining dead states as cellular metabolic states that are not returnable to the predefined “representative living states”; regardless of the modulation of enzyme concentrations and external nutrient concentrations. The definition requires a method to compute the restricted, global, and nonlinear controllability, for which no general theory exists. We have developed “The Stoichiometric Rays”, a simple method to solve the controllability computation. This allows us to compute how the enzyme concentration should be modulated to control the metabolic state from a given state to a desired state.

Using the stoichiometric rays, we have computed the returnability of the non-growing state emerging in an in silico metabolic model of E. coli [2] to the growing state of the model, and found that the non-growing state is indeed a “dead” state. Furthermore, we have quantified “the Separating Alive and Non-life Zone (SANZ) hypersurface” [3] which divides the phase space into the living- and non-living regions.

In this talk, I would like to present our framework for cell death, including stoichiometric rays, and what we can learn from quantifying the SANZ hypersurface.

[1] Himeoka et al., (2024), arXiv, March. https://arxiv.org/abs/2403.02169.
[2] Boecker et al., (2021), Mol. Syst. Biol., 17 (12): e10504.
[3] The Sanzu hypersurface is derived from a mythical river in the Japanese Buddhist tradition, the Sanzu River that represents the boundary between the world of the living and the afterlife.

Title: Acquisition of Implicit World Models via Domain Adaptation in Homeostatic Meta-Reinforcement Learning

Abstract: In this talk, we discuss the possibility of world models and active exploration as emergent properties of open-ended behavior optimization in autonomous agents. In discussing the source of the open-endedness of living things, we start from the perspective of biological systems as understood by the mechanistic approach of theoretical biology and artificial life. From this perspective, we discuss the potential of homeostasis in particular as an open-ended objective for autonomous agents and as a general, integrative extrinsic motivation. We then discuss the possibility of implicitly acquiring a world model and active exploration through the internal dynamics of a network, and a hypothetical architecture for this, by combining meta-reinforcement learning, which assumes domain adaptation as a system that achieves robust homeostasis.

Title: The Origin of Intentions: From Others to Self through Superposition

Abstract: This presentation explores the origins of goal-directed intentions using deep learning simulation models. Building upon our previous work on the superposition mechanism for understanding self and other, we now extend this concept to investigate how intentional behaviors emerge. The most crucial idea of the superposition mechanism is the premise that self and other are processed internally by the same model without distinction. Our simulation results demonstrated that the superposition mechanism enables agents to naturally share internal states and engage in perspective-taking without explicit information about others, simply by predicting their own future sensations. Based on these results, we hypothesize that intentional behaviors may arise through a process of "contagion" from observing others' actions. This talk will discuss the implications of our findings for understanding the development of goal-directed actions. We will explore how this unified model of self and other contributes to the emergence of intentional behavior, offering new insights into the cognitive processes underlying goal-directed actions.

Title: Persistence of Complex Patterns in Lenia and Their Physical Interpretation

Abstract: Lenia is a two-dimensional continuous cellular automaton conceived as an extension of the Game of Life. It generates localized, moving patterns with diverse and complex shapes that persist over time. This life-like behavior appears qualitatively different from patterns observed in physical systems such as reaction-diffusion (RD) systems, although some patterns in Lenia resemble those in RD systems. Motivated by this observation, we explore how Lenia differs from physical systems by examining the underlying dynamics. By identifying similarities and differences, we aim to understand how Lenia's specific features lead to the maintenance of complex patterns, which we regard as its form of goal-directed behavior.

Title: Self-propelled motion of camphor on a water surface interacting with system boundaries

Abstract: To achieve the goal directed behaviour, self-propelled objects that perceive the surrounding environment and have a protocol to determine where to move next can be a candidate system. A camphor particle, a sublimable solid substance, on a water surface is one of the simple examples of such a system. When a camphor particle is put on the water surface, it releases the camphor molecules on a water surface. Since the camphor molecules change the surface tension of the water surface, the camphor particle can move in a direction with higher surface tension. Since the time development of the camphor concentration field is affected by the water surface boundaries, the camphor particle "feels" the system boundaries through the camphor concentration field, and changes its direction of motion. Furthermore, a camphor particle can push or pull floating objects on the water surface. In my talk, I would like to introduce two topics: a camphor particle motion confined in certain boundaries and the behaviour of a camphor particle that can open and close a gate on the water surface.

Title: Observation of complex behaviors in simple dissipative systems

Abstract: Even very simple and small-scale dissipative systems can express relatively complex spatio-temporal patterns with seemingly life-like properties. Furthermore, In this talk I will showcase a few examples of laboratory experiments with both soft and condensed matter systems that are self-propelled on aqueous surfaces. These systems demonstrate that with only minor increases in system complexity, the behavioral complexity can vary or increase significantly. Furthermore, it is striking how easy it feels for the observer to assign imaginary non-trivial goals to these clearly non-living objects. This may indicate that these dissipative systems operate on the verge of goal-directed behavior.

Title: The Case For “Pansolipsism”

Abstract:  I describe a novel conceptual approach to the scientific study of embodied minds, which concerns itself with the relations between different embodied agents. The underlying intuition is that there are “gaps” between minds, corresponding to the difficulty involved in translating from one cognitive “reference frame” to another. In this framework, there is no limit on how alien one mind can be to another. This allows for a nuanced form of panpsychism, in which very few physical systems can be understood as possessing minds that are meaningfully “accessible” to humans, but every physical system can potentially be understood as possessing a mind “in the abstract”.

Title: Foundations of Viability Space

Abstract: As a general principle, a precondition for all goals is survival, at least until the point that said goals have been achieved. To accomplish this, organisms must metabolically and behaviorally achieve dynamics to balance their state within a complex set of physiological constraints. The question of how organisms do this falls under the study of viability, and, respectively, the physiological limits that distinguish life from death are known as viability constraints. Although it is common to see agent dynamics being modeled in tandem with viability constraints within computational biology, personalized medicine, and Artificial Life, analyzing these systems has remained an open challenge. In this talk, we demonstrate the method of Viability Space Decomposition, which allows us to partition an agent’s state space into regions that will maintain viable dynamics and regions that result in death on a finite time horizon. We first demonstrate this by analyzing a simple protocell model in a fixed environment. Then, we allow the protocell to navigate its environment according to its metabolic needs and show how the method can still be applied to this more complex case. We finish by pointing out some other essential details regarding viability at large, such as how interacting agents influence each other’s survival outcomes and how viability constraints emerge in the first place.

Title: Metastable Goals and Their Possible Mechanisms

Abstract: The mechanistic basis of goal-directed behavior is a seeming contradiction, since causal mechanisms must chronologically precede achieved goals. The post-cybernetic concept of stability offers a diachronic mechanism that synthesizes past this contradiction. However, even stability is insufficient because it misses the fundamentally transient nature of agential behavior. The concept of metastability promises to overcome this gap. In this talk, I will describe several approaches that have been taken to formalize metastability, paying particular attention to their conceptual commitments, mathematical differences, and empirical predictions.

Title: Exploring the informational architecture of agency

Abstract: In this talk we will explore recent advances that are opening the way towards novel approaches to characterise agents. We will first review how methods to disentangle qualitatively distinct types of information reveals a peculiar informational architecture that is distinctive of the human brain. This architecture allows to coordinate robust localised processing for ensuring reliable input-output functions, while enabling distributed processing in associative cortices associated with complex cognitive capabilities. We will then explore applications of computational mechanics to reverse-engineer perception-action loops from data, and discuss criteria for discriminating when modelling based on “blankets” is appropriate. We will conclude reviewing open question and future research directions related to these approaches.

Title: Synchronization of hydrogel beads

Abstract: A drop of chitosan solution into a pool of sodium hydroxide solution results in passive hydrogel beads. When surface active ethanol is added to the chitosan solution the evolving active beads exhibit self-propulsion on the air-liquid interface. When several active beads are formed along with a single passive bead, the passive bead serves as a coordinator and the active beads assemble around it in a rotatory way.  When three or more fuelled beads are placed close to each other, synchronized oscillatory patterns develop due to the inherent Marangoni convection and the arising capillary attraction.

Title: Modeling the Origins and Evolution of Goal-Directed Systems: A Reaction Network Approach

Abstract: Traditional approaches to modeling complex adaptive systems often struggle to adequately represent crucial systemic concepts such as emergence, autonomy, and purpose. In this talk, I will present a novel framework based on reaction network modeling that shifts the focus from static objects to dynamic processes of change, offering new insights into the origins and evolution of goal-directed systems.

Building on the foundational concept of autopoiesis—the self-maintaining and self-reproducing nature of living systems—this approach leverages chemical organization theory to formalize key ideas around agency, autonomy, and purpose in general systems. I will explain the basics of this framework, scope its applicability, and trace a modeling roadmap from basic self-maintenance to purposeful collective intelligence.

In particular, I will focus on the computation of a hierarchy of autopoietic systems from a general reaction network, and how to study the evolution of complexity/goals in this setting, as well as introducing concepts that link the behavior of the system with its cognitive-like capabilities. Additionally, I will show some modeling examples built on a library currently under development.

Title: Goal-directed behaviour and meaningfulness: towards a mathematical framework

Abstract: A system and its environment can always be regarded as merely two coupled physical/dynamical systems. However, if our system is performing some kind of task - an organism staying alive, for example, or a robot solving a maze - then we often regard it as more than that: its internal states seem, sometimes, to carry meaning about the environment. The nature of this 'meaning' has been hotly debated for decades. Here we propose a simple mathematical framework that relates (i) the idea of a _viability boundary_ as found in the works of Ashby, Maturana and Varela, Beer, and others, with (ii) the notion of an _interpretation map_ from our own previous work, which can be regarded as assigning meaning to the states of a system. Our goal is not to take a strong philosophical position but to unpick the subtle relationship between viability and meaning; we believe that the mathematical insights are useful regardless of what philosophical position one takes. This is joint work with Martin Biehl, Manuel Baltieri and Matteo Capucci, although any philosophical opinions expressed are mine.