Understanding Memories of the Past in the Context of Different Complex Neural Network Architectures.

Deep learning (primarily using backpropagation) and neuroevolution are the preeminent methods of optimizing artificial neural networks. However, they often create black boxes that are as hard to understand as the natural brains they seek to mimic. Previous work has identified an information-theoretic tool, referred to as R, which allows us to quantify and identify mental representations in artificial cognitive systems. The use of such measures has allowed us to make previous black boxes more transparent. Here we extend R to not only identify where complex computational systems store memory about their environment but also to differentiate between different time points in the past. We show how this extended measure can identify the location of memory related to past experiences in neural networks optimized by deep learning as well as a genetic algorithm.

Using the Comparative Hybrid Approach to Disentangle the Role of Substrate Choice on the Evolution of Cognition.

Understanding the structure and evolution of natural cognition is a topic of broad scientific interest, as is the development of an engineering toolkit to construct artificial cognitive systems. One open question is determining which components and techniques to use in such a toolkit. To investigate this question, we employ agent-based AI, using simple computational substrates (i.e., digital brains) undergoing rapid evolution. Such systems are an ideal choice as they are fast to process, easy to manipulate, and transparent for analysis. Even in this limited domain, however, hundreds of different computational substrates are used. While benchmarks exist to compare the quality of different substrates, little work has been done to build broader theory on how substrate features interact. We propose a technique called the Comparative Hybrid Approach and develop a proof-of-concept by systematically analyzing components from three evolvable substrates: recurrent artificial neural networks, Markov brains, and Cartesian genetic programming. We study the role and interaction of individual elements of these substrates by recombining them in a piecewise manner to form new hybrid substrates that can be empirically tested. Here, we focus on network sparsity, memory discretization, and logic operators of each substrate. We test the original substrates and the hybrids across a suite of distinct environments with different logic and memory requirements. While we observe many trends, we see that discreteness of memory and the Markov brain logic gates correlate with high performance across our test conditions. Our results demonstrate that the Comparative Hybrid Approach can identify structural subcomponents that predict task performance across multiple computational substrates.

Information Fragmentation, Encryption and Information Flow in Complex Biological Networks.

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.

The effects of social rank and payoff structure on the evolution of group hunting.

Group hunting is common among social carnivores, and mechanisms that promote this behavior are a central topic in evolutionary biology. Increased prey capture success and decreased losses from competitors are often invoked as factors promoting group hunting. However, many animal societies have linear dominance hierarchies where access to critical resources is determined by social rank, and group-hunting rewards are shared unequally. Despite this inequality, animals in such societies cooperate to hunt and defend resources. Game theoretic models predict that rank and relative rewards from group hunting vs. solitary hunting affect which hunting strategies will evolve. These predictions are partially supported by empirical work, but data needed to test these predictions are difficult to obtain in natural systems. We use digital evolution to test how social rank and tolerance by dominants of subordinates feeding while sharing spoils from group hunting influence which hunting strategies evolve in digital organisms. We created a computer-simulated world to reflect social and hunting dynamics of spotted hyenas (Crocuta crocuta). We found that group hunting increased as tolerance increased and as the relative payoff from group hunting increased. Also, top-ranking agents were more likely to group hunt than lower-ranking agents under despotic sharing conditions. These results provide insights into mechanisms that may promote cooperation in animal societies structured by dominance hierarchies.