Christoph Adami

Clifford Bohm, Vincent R. Ragusa, Arend Hintze et al.

A central challenge in analyzing multivariate interactions within complex systems is to decompose how multiple inputs jointly determine an output. Existing approaches generally operate on observed probability distributions and can conflate a system's intrinsic functional logic with statistical artifacts of limited data. As a result, distinct systems can yield identical observations, rendering information decomposition fundamentally underdetermined and obscuring true higher-order interactions. We introduce Functional Information Decomposition (FID), both a computational and theoretical framework, which defines informational components with respect to a system's complete input-output mapping, thereby addressing a core cross-scale inference problem: determining how information carried by individual components combines to shape system-level behavior. When the mapping is fully specified, FID provides a unique decomposition into independent and synergistic contributions. Crucially, given only partial observations, FID characterizes the entire space of consistent decompositions by sampling compatible functions, making inferential limits explicit. A complementary geometric perspective clarifies the structural origin of informational components. We demonstrate FID's interdisciplinary utility on canonical logical functions, Conway's Game of Life, and gene-expression-based prediction of cancer drug response, and provide an open-source implementation. By separating functional architecture from observational distribution, FID offers a principled foundation for analyzing multivariate dependence in both fully and partially observed complex systems.

Arend Hintze, Christoph Adami

Deep learning of artificial neural networks (ANNs) is creating highly functional processes that are, unfortunately, nearly as hard to interpret as their biological counterparts. Identification of functional modules in natural brains plays an important role in cognitive and neuroscience alike, and can be carried out using a wide range of technologies such as fMRI, EEG/ERP, MEG, or calcium imaging. However, we do not have such robust methods at our disposal when it comes to understanding functional modules in artificial neural networks. Ideally, understanding which parts of an artificial neural network perform what function might help us to address a number of vexing problems in ANN research, such as catastrophic forgetting and overfitting. Furthermore, revealing a network's modularity could improve our trust in them by making these black boxes more transparent. Here, we introduce a new information-theoretic concept that proves useful in understanding and analyzing a network's functional modularity: the relay information $I_R$. The relay information measures how much information groups of neurons that participate in a particular function (modules) relay from inputs to outputs. Combined with a greedy search algorithm, relay information can be used to identify computational modules in neural networks. We also show that the functionality of modules correlates with the amount of relay information they carry.

Ankit Gupta, Christoph Adami

Modern machine learning methods have been proposed to detect life in extraterrestrial samples, drawing on their ability to distinguish biotic from abiotic samples based on training models using natural and synthetic organic molecular mixtures. Here we show using Artificial Life that such methods are easily fooled into detecting life with near 100% confidence even if the analyzed sample is not capable of life. This is due to modern machine learning methods' propensity to be easily fooled by out-of-distribution samples. Because extra-terrestrial samples are very likely out of the distribution provided by terrestrial biotic and abiotic samples, using AI methods for life detection is bound to yield significant false positives.

Arend Hintze, Christoph Adami

The tragedy of the commons illustrates a fundamental social dilemma where individual rational actions lead to collectively undesired outcomes, threatening the sustainability of shared resources. Strategies to escape this dilemma, however, are in short supply. In this study, we explore how artificial intelligence (AI) agents can be leveraged to enhance cooperation in public goods games, moving beyond traditional regulatory approaches to using AI as facilitators of cooperation. We investigate three scenarios: (1) Mandatory Cooperation Policy for AI Agents, where AI agents are institutionally mandated always to cooperate; (2) Player-Controlled Agent Cooperation Policy, where players evolve control over AI agents' likelihood to cooperate; and (3) Agents Mimic Players, where AI agents copy the behavior of players. Using a computational evolutionary model with a population of agents playing public goods games, we find that only when AI agents mimic player behavior does the critical synergy threshold for cooperation decrease, effectively resolving the dilemma. This suggests that we can leverage AI to promote collective well-being in societal dilemmas by designing AI agents to mimic human players.

Ankit Gupta, Christoph Adami, Emily Dolson

Deep neural networks (DNNs) excel across image recognition tasks, yet continue to exhibit overconfidence on inputs that bear no resemblance to natural images. Revisiting the "fooling images" work introduced by Nguyen et al. (2015), we re-implement both CPPN-based and direct-encoding-based evolutionary fooling attacks on modern architectures, including convolutional and transformer classifiers. Our re-implementation confirm that high-confidence fooling persists even in state-of-the-art networks, with transformer-based ViT-B/16 emerging as the most susceptible--achieving near-certain misclassifications with substantially fewer queries than convolution-based models. We then introduce SPOOF, a minimalist, consistent, and more efficient black-box attack generating high-confidence fooling images. Despite its simplicity, SPOOF generates unrecognizable fooling images with minimal pixel modifications and drastically reduced compute. Furthermore, retraining with fooling images as an additional class provides only partial resistance, as SPOOF continues to fool consistently with slightly higher query budgets--highlighting persistent fragility of modern deep classifiers.

Clifford Bohm, Douglas Kirkpatrick, Victoria Cao et al.

Assessing where and how information is stored in biological networks (such as neuronal and genetic networks) is a central task both in neuroscience and in molecular genetics, but most available tools focus on the network's structure as opposed to its function. Here we introduce a new information-theoretic tool: "information fragmentation analysis" that, given full phenotypic data, allows us to localize information in complex networks, determine how fragmented (across multiple nodes of the network) the information is, and assess the level of encryption of that information. Using information fragmentation matrices, we can also create information flow graphs that illustrate how information propagates through these networks. We illustrate the use of this tool by analyzing how artificial brains that evolved "in silico" solve particular tasks, and show how information fragmentation analysis provides deeper insights into how these brains process information and "think". The measures of information fragmentation and encryption that result from our methods also quantify complexity of information processing in these networks and how this processing complexity differs between primary exposure to sensory data (early in the lifetime) and later routine processing.

Ali Tehrani-Saleh, Christoph Adami

To infer information flow in any network of agents, it is important first and foremost to establish causal temporal relations between the nodes. Practical and automated methods that can infer causality are difficult to find, and the subject of ongoing research. While Shannon information only detects correlation, there are several information-theoretic notions of "directed information" that have successfully detected causality in some systems, in particular in the neuroscience community. However, recent work has shown that some directed information measures can sometimes inadequately estimate the extent of causal relations, or even fail to identify existing cause-effect relations between components of systems, especially if neurons contribute in a cryptographic manner to influence the effector neuron. Here, we test how often cryptographic logic emerges in an evolutionary process that generates artificial neural circuits for two fundamental cognitive tasks: motion detection and sound localization. We also test whether activity time-series recorded from behaving digital brains can infer information flow using the transfer entropy concept, when compared to a ground-truth model of causal influence constructed from connectivity and circuit logic. Our results suggest that transfer entropy will sometimes fail to infer causality when it exists, and sometimes suggest a causal connection when there is none. However, the extent of incorrect inference strongly depends on the cognitive task considered. These results emphasize the importance of understanding the fundamental logic processes that contribute to information flow in cognitive processing, and quantifying their relevance in any given nervous system.

Ali Tehrani-Saleh, Thomas LaBar, Christoph Adami

A central goal of evolutionary biology is to explain the origins and distribution of diversity across life. Beyond species or genetic diversity, we also observe diversity in the circuits (genetic or otherwise) underlying complex functional traits. However, while the theory behind the origins and maintenance of genetic and species diversity has been studied for decades, theory concerning the origin of diverse functional circuits is still in its infancy. It is not known how many different circuit structures can implement any given function, which evolutionary factors lead to different circuits, and whether the evolution of a particular circuit was due to adaptive or non-adaptive processes. Here, we use digital experimental evolution to study the diversity of neural circuits that encode motion detection in digital (artificial) brains. We find that evolution leads to an enormous diversity of potential neural architectures encoding motion detection circuits, even for circuits encoding the exact same function. Evolved circuits vary in both redundancy and complexity (as previously found in genetic circuits) suggesting that similar evolutionary principles underlie circuit formation using any substrate. We also show that a simple (designed) motion detection circuit that is optimally-adapted gains in complexity when evolved further, and that selection for mutational robustness led this gain in complexity.

Arend Hintze, Douglas Kirkpatrick, Christoph Adami

Artificial neural networks (ANNs), while exceptionally useful for classification, are vulnerable to misdirection. Small amounts of noise can significantly affect their ability to correctly complete a task. Instead of generalizing concepts, ANNs seem to focus on surface statistical regularities in a given task. Here we compare how recurrent artificial neural networks, long short-term memory units, and Markov Brains sense and remember their environments. We show that information in Markov Brains is localized and sparsely distributed, while the other neural network substrates "smear" information about the environment across all nodes, which makes them vulnerable to noise.

Joel Lehman, Jeff Clune, Dusan Misevic et al.

Biological evolution provides a creative fount of complex and subtle adaptations, often surprising the scientists who discover them. However, because evolution is an algorithmic process that transcends the substrate in which it occurs, evolution's creativity is not limited to nature. Indeed, many researchers in the field of digital evolution have observed their evolving algorithms and organisms subverting their intentions, exposing unrecognized bugs in their code, producing unexpected adaptations, or exhibiting outcomes uncannily convergent with ones in nature. Such stories routinely reveal creativity by evolution in these digital worlds, but they rarely fit into the standard scientific narrative. Instead they are often treated as mere obstacles to be overcome, rather than results that warrant study in their own right. The stories themselves are traded among researchers through oral tradition, but that mode of information transmission is inefficient and prone to error and outright loss. Moreover, the fact that these stories tend to be shared only among practitioners means that many natural scientists do not realize how interesting and lifelike digital organisms are and how natural their evolution can be. To our knowledge, no collection of such anecdotes has been published before. This paper is the crowd-sourced product of researchers in the fields of artificial life and evolutionary computation who have provided first-hand accounts of such cases. It thus serves as a written, fact-checked collection of scientifically important and even entertaining stories. In doing so we also present here substantial evidence that the existence and importance of evolutionary surprises extends beyond the natural world, and may indeed be a universal property of all complex evolving systems.

Christoph Adami

The present document is an excerpt of an essay that I wrote as part of my application material to graduate school in Computer Science (with a focus on Artificial Intelligence), in 1986. I was not invited by any of the schools that received it, so I became a theoretical physicist instead. The essay's full title was "Some Topics in Philosophy and Computer Science". I am making this text (unchanged from 1985, preserving the typesetting as much as possible) available now in memory of Jerry Fodor, whose writings had influenced me significantly at the time (even though I did not always agree).

Arend Hintze, Jeffrey A. Edlund, Randal S. Olson et al.

Markov Brains are a class of evolvable artificial neural networks (ANN). They differ from conventional ANNs in many aspects, but the key difference is that instead of a layered architecture, with each node performing the same function, Markov Brains are networks built from individual computational components. These computational components interact with each other, receive inputs from sensors, and control motor outputs. The function of the computational components, their connections to each other, as well as connections to sensors and motors are all subject to evolutionary optimization. Here we describe in detail how a Markov Brain works, what techniques can be used to study them, and how they can be evolved.

Randal S. Olson, Jason H. Moore, Christoph Adami

Pattern recognition and classification is a central concern for modern information processing systems. In particular, one key challenge to image and video classification has been that the computational cost of image processing scales linearly with the number of pixels in the image or video. Here we present an intelligent machine (the "active categorical classifier," or ACC) that is inspired by the saccadic movements of the eye, and is capable of classifying images by selectively scanning only a portion of the image. We harness evolutionary computation to optimize the ACC on the MNIST hand-written digit classification task, and provide a proof-of-concept that the ACC works on noisy multi-class data. We further analyze the ACC and demonstrate its ability to classify images after viewing only a fraction of the pixels, and provide insight on future research paths to further improve upon the ACC presented here.

Ali Tehrani-Saleh, Christoph Adami

Flies that walk in a covered planar arena on straight paths avoid colliding with each other, but which of the two flies stops is not random. High-throughput video observations, coupled with dedicated experiments with controlled robot flies have revealed that flies utilize the type of optic flow on their retina as a determinant of who should stop, a strategy also used by ship captains to determine which of two ships on a collision course should throw engines in reverse. We use digital evolution to test whether this strategy evolves when collision avoidance is the sole penalty. We find that the strategy does indeed evolve in a narrow range of cost/benefit ratios, for experiments in which the "regressive motion" cue is error free. We speculate that these stringent conditions may not be sufficient to evolve the strategy in real flies, pointing perhaps to auxiliary costs and benefits not modeled in our study

Randal S. Olson, Arend Hintze, Fred C. Dyer et al.

A common idiom in biology education states, "Eyes in the front, the animal hunts. Eyes on the side, the animal hides." In this paper, we explore one possible explanation for why predators tend to have forward-facing, high-acuity visual systems. We do so using an agent-based computational model of evolution, where predators and prey interact and adapt their behavior and morphology to one another over successive generations of evolution. In this model, we observe a coevolutionary cycle between prey swarming behavior and the predator's visual system, where the predator and prey continually adapt their visual system and behavior, respectively, over evolutionary time in reaction to one another due to the well-known "predator confusion effect." Furthermore, we provide evidence that the predator visual system is what drives this coevolutionary cycle, and suggest that the cycle could be closed if the predator evolves a hybrid visual system capable of narrow, high-acuity vision for tracking prey as well as broad, coarse vision for prey discovery. Thus, the conflicting demands imposed on a predator's visual system by the predator confusion effect could have led to the evolution of complex eyes in many predators.

Randal S. Olson, Patrick B. Haley, Fred C. Dyer et al.

Despite the fact that grouping behavior has been actively studied for over a century, the relative importance of the numerous proposed fitness benefits of grouping remain unclear. We use a digital model of evolving prey under simulated predation to directly explore the evolution of gregarious foraging behavior according to one such benefit, the "many eyes" hypothesis. According to this hypothesis, collective vigilance allows prey in large groups to detect predators more efficiently by making alarm signals or behavioral cues to each other, thereby allowing individuals within the group to spend more time foraging. Here, we find that collective vigilance is sufficient to select for gregarious foraging behavior as long there is not a direct cost for grouping (e.g., competition for limited food resources), even when controlling for confounding factors such as the dilution effect. Further, we explore the role of the genetic relatedness and reproductive strategy of the prey, and find that highly related groups of prey with a semelparous reproductive strategy are the most likely to evolve gregarious foraging behavior mediated by the benefit of vigilance. These findings, combined with earlier studies with evolving digital organisms, further sharpen our understanding of the factors favoring grouping behavior.

Anya Elaine Johnson, Eli Strauss, Rodney Pickett et al.

Within the context of evolution, an altruistic act that benefits the receiving individual at the expense of the acting individual is a puzzling phenomenon. An extreme form of altruism can be found in colicinogenic E. coli. These suicidal altruists explode, releasing colicins that kill unrelated individuals, which are not colicin resistant. By committing suicide, the altruist makes it more likely that its kin will have less competition. The benefits of this strategy rely on the number of competitors and kin nearby. If the organism explodes at an inopportune time, the suicidal act may not harm any competitors. Communication could enable organisms to act altruistically when environmental conditions suggest that that strategy would be most beneficial. Quorum sensing is a form of communication in which bacteria produce a protein and gauge the amount of that protein around them. Quorum sensing is one means by which bacteria sense the biotic factors around them and determine when to produce products, such as antibiotics, that influence competition. Suicidal altruists could use quorum sensing to determine when exploding is most beneficial, but it is challenging to study the selective forces at work in microbes. To address these challenges, we use digital evolution (a form of experimental evolution that uses self-replicating computer programs as organisms) to investigate the effects of enabling altruistic organisms to communicate via quorum sensing. We found that quorum-sensing altruists killed a greater number of competitors per explosion, winning competitions against non-communicative altruists. These findings indicate that quorum sensing could increase the beneficial effect of altruism and the suite of conditions under which it will evolve.

David B. Knoester, Heather J. Goldsby, Christoph Adami

Building self-adaptive and self-organizing (SASO) systems is a challenging problem, in part because SASO principles are not yet well understood and few platforms exist for exploring them. Cellular automata (CA) are a well-studied approach to exploring the principles underlying self-organization. A CA comprises a lattice of cells whose states change over time based on a discrete update function. One challenge to developing CA is that the relationship of an update function, which describes the local behavior of each cell, to the global behavior of the entire CA is often unclear. As a result, many researchers have used stochastic search techniques, such as evolutionary algorithms, to automatically discover update functions that produce a desired global behavior. However, these update functions are typically defined in a way that does not provide for self-adaptation. Here we describe an approach to discovering CA update functions that are both self-adaptive and self-organizing. Specifically, we use a novel evolutionary algorithm-based approach to discover finite state machines (FSMs) that implement update functions for CA. We show how this approach is able to evolve FSM-based update functions that perform well on the density classification task for 1-, 2-, and 3-dimensional CA. Moreover, we show that these FSMs are self-adaptive, self-organizing, and highly scalable, often performing well on CA that are orders of magnitude larger than those used to evaluate performance during the evolutionary search. These results demonstrate that CA are a viable platform for studying the integration of self-adaptation and self-organization, and strengthen the case for using evolutionary algorithms as a component of SASO systems.

Arend Hintze, Randal S. Olson, Christoph Adami et al.

Risk aversion is a common behavior universal to humans and animals alike. Economists have traditionally defined risk preferences by the curvature of the utility function. Psychologists and behavioral economists also make use of concepts such as loss aversion and probability weighting to model risk aversion. Neurophysiological evidence suggests that loss aversion has its origins in relatively ancient neural circuitries (e.g., ventral striatum). Could there thus be an evolutionary origin to risk avoidance? We study this question by evolving strategies that adapt to play the equivalent mean payoff gamble. We hypothesize that risk aversion in the equivalent mean payoff gamble is beneficial as an adaptation to living in small groups, and find that a preference for risk averse strategies only evolves in small populations of less than 1,000 individuals, while agents exhibit no such strategy preference in larger populations. Further, we discover that risk aversion can also evolve in larger populations, but only when the population is segmented into small groups of around 150 individuals. Finally, we observe that risk aversion only evolves when the gamble is a rare event that has a large impact on the individual's fitness. These findings align with earlier reports that humans lived in small groups for a large portion of their evolutionary history. As such, we suggest that rare, high-risk, high-payoff events such as mating and mate competition could have driven the evolution of risk averse behavior in humans living in small groups.

Randal S. Olson, David B. Knoester, Christoph Adami

Animal grouping behaviors have been widely studied due to their implications for understanding social intelligence, collective cognition, and potential applications in engineering, artificial intelligence, and robotics. An important biological aspect of these studies is discerning which selection pressures favor the evolution of grouping behavior. In the past decade, researchers have begun using evolutionary computation to study the evolutionary effects of these selection pressures in predator-prey models. The selfish herd hypothesis states that concentrated groups arise because prey selfishly attempt to place their conspecifics between themselves and the predator, thus causing an endless cycle of movement toward the center of the group. Using an evolutionary model of a predator-prey system, we show that how predators attack is critical to the evolution of the selfish herd. Following this discovery, we show that density-dependent predation provides an abstraction of Hamilton's original formulation of ``domains of danger.'' Finally, we verify that density-dependent predation provides a sufficient selective advantage for prey to evolve the selfish herd in response to predation by coevolving predators. Thus, our work corroborates Hamilton's selfish herd hypothesis in a digital evolutionary model, refines the assumptions of the selfish herd hypothesis, and generalizes the domain of danger concept to density-dependent predation.

Randal S. Olson, Arend Hintze, Fred C. Dyer et al.

Swarming behaviors in animals have been extensively studied due to their implications for the evolution of cooperation, social cognition, and predator-prey dynamics. An important goal of these studies is discerning which evolutionary pressures favor the formation of swarms. One hypothesis is that swarms arise because the presence of multiple moving prey in swarms causes confusion for attacking predators, but it remains unclear how important this selective force is. Using an evolutionary model of a predator-prey system, we show that predator confusion provides a sufficient selection pressure to evolve swarming behavior in prey. Furthermore, we demonstrate that the evolutionary effect of predator confusion on prey could in turn exert pressure on the structure of the predator's visual field, favoring the frontally oriented, high-resolution visual systems commonly observed in predators that feed on swarming animals. Finally, we provide evidence that when prey evolve swarming in response to predator confusion, there is a change in the shape of the functional response curve describing the predator's consumption rate as prey density increases. Thus, we show that a relatively simple perceptual constraint--predator confusion--could have pervasive evolutionary effects on prey behavior, predator sensory mechanisms, and the ecological interactions between predators and prey.

Lars Marstaller, Arend Hintze, Christoph Adami

Representations are internal models of the environment that can provide guidance to a behaving agent, even in the absence of sensory information. It is not clear how representations are developed and whether or not they are necessary or even essential for intelligent behavior. We argue here that the ability to represent relevant features of the environment is the expected consequence of an adaptive process, give a formal definition of representation based on information theory, and quantify it with a measure R. To measure how R changes over time, we evolve two types of networks---an artificial neural network and a network of hidden Markov gates---to solve a categorization task using a genetic algorithm. We find that the capacity to represent increases during evolutionary adaptation, and that agents form representations of their environment during their lifetime. This ability allows the agents to act on sensorial inputs in the context of their acquired representations and enables complex and context-dependent behavior. We examine which concepts (features of the environment) our networks are representing, how the representations are logically encoded in the networks, and how they form as an agent behaves to solve a task. We conclude that R should be able to quantify the representations within any cognitive system, and should be predictive of an agent's long-term adaptive success.