Motivation

It's exciting to think that nature has come up with a simple mechanism that can make sense of any kind of sensory information (about the animal or the world in which it lives), adapt to any kind of physical body, and then navigate the world in it.

In general, humans hold the idea that their brain is the seat of intelligence, and exists for the animal's survival, with preprogrammed capabilities to enable communication and navigating the world. Oddly, because we do not observe other species communicating with the same communication modalities, or solving problems that we would solve, we assume that we alone are intelligent, and other species may not even be self-aware. As we look for artificial intelligence (AI) in computer systems, we rank highly those systems that can exhibit communications (such as writing, speaking, translating one natural language to another, and creating music or visual art). However, we must be careful of this novelty-- LLMs generate results because they are trained on large corpi of human-generated data, and naturally can exhibit the breadth and depth of this behavior. But does it make them sentient?

Starting more simply, we seek to understand how animals exhibit behavior, through the direction of their central nervous system (CNS), which in all cases so far studied has consisted of a network of cells called neurons, interconnected with a high degree of density (a given neuron typically has thousands of connections with neurons). Many different types of neurons have been studied in animal tissues, and their properties well-documented. While we understand the mechanics of neurons at this level, there is a large gap of understanding about how various types of neurons may be interconnected in patterns to create a brain from which real behavior emerges.

Nature has shown us billions of species over millions of years that have adopted different brain architectures that all successfully generated behavior that allowed those animals to navigate their environment. Further, the individual animals developed with their own brain wiring within their species' brain plan architectural constraints. It is astonishing to come to the understanding that, if billions of species, each with at least billions of individuals, produced an enormous range of individual brain wirings that exhibited behavior, there is not one fixed wiring plan of spiking neurons that generates behavior. Instead, there must be a set of salient architectural elements of these brain plans that enable the brain to generate behavior.

The term behavior is used intentionally broadly. The behavior of an ant is quite different from a fruit fly, rodent, human, or octopus. We can say, however, that any behavior that arises from the brain simply interacting with the body and the environment which enables the organism to navigate its world and lead a functional life, is evidence of sentience.

What is sentience though? In different cultures and over time, the term sentience has come to mean different things, largely because humans use it in conjunction with the seat of human consciousness or emotions, and this is by its very nature anthropomorphically limiting. At NeuroSynthetica, we come back to behavior and what it must mean for the experience of the sentient animal, which armed with a body and brain alone in a new world, is able to familiarize itself with its world and successfully navigate it throughout its life. The emergence of this type of meaningful and relevant behavior is an indication that the animal is sentient-- that it senses its world around it (as well as its internal state), and makes decisions about taking actions, not because it has been hardwired to do so, but because it has adapted to do so.

In a general embodiment, the feedback loop associated with sensory inputs stimulating the neural network in one area of the animal's brain can affect the neural networks in other areas of the animal's brain that generate actions, causing a change in sensory input, which may give rise to more changes in actions. Though this feedback loop is not itself sufficient, we must remember that connections between neurons have strengths that are adjusted in real time as those connections encounter signals. It is likely that the network of neural networks in a typical brain can encode uncountable mathematical attractors that allow sensory inputs to give rise to behavior automatically at the lowest level, and sequences of patterns of attractors to arise under certain conditions. We believe that a paradigm similar to this may give rise to behavior.

NeuroSynthetica's goal is to identify the salient components of the biological mechanism that gives rise to sentience. We do this in simulo by creating models of brains, starting from small ones that exhibit small behaviors, and simulate them using our own simulator called the Sentience Engine, which runs as a server on COTS (Commercial Off The Shelf) computers.

We use an interactive Workbench to write code in the SOMA descriptive programming language, allowing us to define a brain's architecture. Then from the Workbench, we compile the model and generate it on the server. Finally, we use the Workbench to start and stop the simulation, and observe it with instrumentation while it interacts with robotic sensors and movement actuators.

We'll know success when we see real animal-level behavior emerge from a synthetic model of a brain described purely in SOMA, simulated in real-time, interacting with the real world. An example might be seeing walking emerge from a quadruped robot-- the coordinated movement of 12 distinct actuators in response to the environment, not because a gait was programmed a certain way in a programming language, but because the robot discovered how to ambulate by itself.

The understanding of how sentience emerges in animal brains may be assisted partially by reverse engineering real brains, but it is unlikely to produce real results because of the complexities of natural brains, which may be constructed with ad hoc architectural optimizations that won't make sense to us. Understanding first design principles in real time spiking neural networks that interact with their environment to produce meaningful behavior is much more likely with a steady stream of experiments-- starting small, and building on that, just as we built on automotive principles, or the transistor.