Dialectek is a research and development LLC involved in a variety of activities.
T. E. Portegys, "Training sensory-motor behavior in the connectome of an artificial C. elegans",
Neurocomputing (2015), pp. 128-134. DOI: 10.1016/j.neucom.2015.06.007
The C. elegans nematode worm is a small well-known creature, intensely studied for decades. Its entire morphology has been mapped cell-by-cell, including its 302 neuron connectome. The connectome is a synaptic wiring diagram that also specifies neurotransmitters and junction types. It does not however specify the synaptic connection strengths. It is believed that measuring these must be done in live specimens, requiring emerging or yet to be developed techniques. Without the connection strengths, it is not known how the nematode's nervous system produces behaviors. Discovering these strengths as a set of weights is a challenging and important problem: an artificial worm embodying the connectome and trained to perform a set of behaviors taken from measurements of the actual C. elegans would behave realistically in its environment. This is a crucial step toward creating a functional artificial creature. Indeed, knowing the artificial weights might cast light on the actual ones. In this project a genetic algorithm was used to train the entire connectome, a large space of 3680 synapse weights, to learn behaviors defined as sensory-motor sequences. It was found that utilizing the topology of the connectome for local optimization and crossover significantly boosts the performance of the genetic algorithm. Using a network of artificial neurons, random sequences involving the entire connectome were successfully trained. Additionally, for locomotion training, sinusoidal body postures were observed when sensory touch neurons were stimulated. Locomotion training was done using a Fourier Transform fitness function. Finally, using the NEURON tool to simulate a biologically higher fidelity network, the pharyngeal assembly of neurons was successfully trained.
T. E. Portegys, "Training artificial neural networks to learn a nondeterministic game",
ICAI'15: The 2015 International Conference on Artificial Intelligence, 2015.
It is well known that artificial neural networks (ANNs) can learn deterministic automata. Learning nondeterministic automata is another matter. This is important because much of the world is nondeterministic, taking the form of unpredictable or probabilistic events that must be acted upon. If ANNs are to engage such phenomena, then they must be able to learn how to deal with nondeterminism. In this project the game of Pong poses a nondeterministic environment. The learner is given an incomplete view of the game state and underlying deterministic physics, resulting in a nondeterministic game. Three models were trained and tested on the game: Mona, Elman, and Numenta's NuPIC.
Game play video.
Thomas Portegys, Gabriel Pascualy, Richard Gordon, Steve McGrew, Bradly Alicea,
"Morphozoic: cellular automata with nested neighborhoods as a metamorphic representation of morphogenesis",
in "Multi-Agent Based Simulations Applied to Biological and Environmental Systems", ISBN: 978-1-5225-1756-6, February 2017
When an embryo develops, its cells exhibit interactions with other cells ranging from nearest neighbors to hormonal effects that can reach anywhere in the body, and are thus global in range. Some of these interactions are mechanical, and of intermediate range. Some involve propagating waves and some are hypothesized to involve diffusion gradients. We can generalize the concept of morphogenetic fields to encompass this full range of cell-cell interactions. These morphogenetic fields would then include the morphogenesis of cellular structures such as tissues and organs and the differentiation of the cells in them. Although there is a preponderance of evidence for the existence of morphogenetic fields of various types, there is also much that is unknown about many of them, including how they are expressed, sensed, and acted upon. Here we investigate a cellular automaton model, Morphozoic, which may be used to investigate the computational power of morphogenetic fields to foster the development of structures and cell differentiation. We are therefore assuming some equivalence between a living cell in an embryo and a cellular automaton cell. A cellular automaton (CA) is a network of elements (cells) that pass signals to one another, then send signals to other cells, depending on the signals that are received. Here we confine cells to a square grid as in ordinary cellular automata. The term "morphogenetic field" is used here to describe a generalized abstraction: a cell signals information about its state to its environment and is able to sense and act on signals from a subset of the other cells. The received set of signals determines the field value for the cell. The subsets are defined by nested neighborhoods that can represent local to global morphogenetic effects. Neighborhood signals are compacted into aggregated quantities, forming a precision gradient that caps the amount of information exchanged: signals from smaller, more local neighborhoods are thus more finely discriminated, while those from larger, more global neighborhoods are less so. Our model allows a cell to use this signaling scheme to determine and perform metamorphic actions, such as division, death, and state change (differentiation). An assembly of cells can thus cooperate to generate spatial and temporal structure. Unlike traditional cellular automata, such as the Game of Life, Morphozoic was found to be robust, redundant, and noise tolerant. Morphozoic is thus an enrichment of the space of cellular automata, which may give us some insight into the unsolved problem of how an embryo builds itself. The evidence for interactions over various distances in real embryos, up to global, is briefly reviewed. Applications of Morphozoic presented here include: (1) Conway's Game of Life, (2) cell regeneration, (3) evolution of a gastrulation-like sequence, (4) neuron pathfinding, and (5) a simulation of Turing's reaction-diffusion morphogenesis. An artificial neural network implementation also provides a noise tolerant generalization capability. Morphozoic provides a tool for exploring relationships between the whole and parts in many realms.