The Neuromorphic AI

Evolution has created the many flavours of life on this planet, higher animals among them reveal the phenomenon of consciousness most evidently. By definition it means to identify with an internal state which is different from the external unalive universe, to evidently be able to practice agency in world. A monkey, elephant and a human are all conscious, but only humans among them are considered sentient which is the inner dialogue. Consciousness is the provider of qualia, to taste a fruit and know its sweet is not simply equal to concluding the fact that it is high in fructose (which a machine can do). We are also certain in the fact that this is a function of the brain as a organ, given basic modern biology. Because neurochemical states of the brain results in altered consciousness, during deep sleep it is extinguished when the brain rests, and it is heightened in presence of environmental stressors.

Since, this is a property of the gross brain as a organ, then the wide many varieties of the brain in the natural world might feel it fundamentally differently, because the neuron density varies (highest in humans). But what does the brain do? It processes information coming from the senses, it maintains a state of the world around the animal and lets it take good decisions. This is exactly what an algorithm does, it is a calculation, the inputs are converted into an output which drives other algorithms. The Convolutional neural networks of AI are incredibly similar to how our own human eyes operate - the first layer is always photons on a detector array, then that information is distilled to find lines and patterns, and finally to the context of what is actually present in the image both in the brain and in AI. Thus we establish that the brain is a conscious computer! what we want to investigate is that if the computer and conscious parts are separate or the computation gives rise to consciousness.

First we look into the architecture choice of nature vs man-made, and look to establish that if both of them are similar processes happening on dis-similar substrates. Then what is the main differentiator. One that we can immediately think of the bat is - distributed processing paradigm

In the brain we do not have a RAM or a CPU, all is CPU plus the RAM, even up to the level of individual neuron. The processing happens right on place where memory is stored, this is sharply different from our classical computers which store the memory in a strictly separate space than its processing head.

The separation of memory and processing is a design choice made by traditional digital computers. Physical systems that combine these two aspects together can provide a much more energy efficient solution to neural processing.

The Church Turing thesis suggests that anything that can be computed can be computed by a Turing machine. This can be said to be equivalent to the law of causality in physics. Generally algorithms are chains of cause and effect. A Turing machine is a universal calculator that is essentially a state machine, and every particle in the universe is a state machine. All computation in this universe therefore will be Turing complete as it strictly abides by the law of causality.

This is specific to Von Neumann hardware.

The Church Turing thesis is about what can be computed. Many systems equivalent to a Turing machine don’t have a memory/processing split. The Game of Life by Conway and other cellular automata are Turing complete. All cells contain some memory.

This comes from the fact that the computer is using a stored program design. This means there is a processing unit and a memory unit connected via a bus. This was a design choice made by computer designers. However, neuromorphic chips, in-memory computing (such as memristor crossbars), dataflow machines and FPGAs all reduce this bottleneck. They are all Turing equivalent or can be stimulated by a Turing machine.

The tape and head of a Turing machine may seem like there is a memory/processor split. However, this is just an abstraction to prove theorems. Complexity classes are robust to different models of computation up to polynomial factors. This means the bottleneck is actually due to constant factors, energy and latency - which is what neuromorphic hardware tries to solve.

\