Unfortunately, most articles covering AGI misrepresent consciousness as a state of ‘self-awareness’. This is a common misrepresentation, particularly among people with conventional secular or religious views. I have found myself debating with more than a few passionate people in pubs, encouraging them to delete the ‘self’ from ‘self-awareness’ when defining consciousness.

As we dive into this discussion, I ask you to consider that consciousness is a more fundamental state than self-awareness. For example, my dog does not need to contemplate self-actualization for her to be a conscious entity capable of subjective experience, she just is. She might not be self-aware, but if someone is mean to her then a moral consequence arises. Many humans are terrible at being self-aware, yet I believe they possess consciousness.

Consciousness is less on the level of self-awareness and more on the level of existence through subjective experience. Can I validate my own existence to myself? Is there such a thing as me? Do my thoughts and experiences matter even when there is no external significance? Does a moral consequence arise when I become the object of pain or harm? Such questions are not asked in a corpuscular sense (of course your body exists and can be harmed), but from an inner subjective sense. From this perspective, we can define consciousness as the emergence of a unified self from a capacity for subjective experience.

Descartes is famously cited for the phrase, “I think, therefore I am”, but in many ways one’s ego is better confirmed by, “I exist because I am aware”, logic that applies equally to my dog as to a person. Buddhists embrace the Principle of Interbeing, which states that “both the subject and object of perception manifest from consciousness”. This principle implies that both the observer and the observed are connected. The experience of you seeing a tree confirms the existence of both you and the tree. If that strikes you as being far too philosophical for science, we can always reframe the general concept in terms of quantum mechanics and the collapse of the wave function. Same paradox, different math.

Encoding

If consciousness arises from, or manifests through, subjective experience, then what physical process causes a subjective experience? How does ordinary matter (dirt) become sentient? The scientific quest for consciousness, as Koch framed the title of his 2004 book, begins with an inquiry into the encoding schemes used to transform neural impulses into phenomenal experiences. For example, blue light is nothing but an arbitrary band in the electromagnetic spectrum, there is nothing ‘blue’ about light in that wavelength (500 nm ±30). Given that fact, what causes the phenomenal experience of blueness? Where does blueness come from?

When the human retina is exposed to light, a bit of neural circuitry generates electric signals in the optic nerve bundle. A similar process occurs in a digital photosensor. However, in both cases the signals do not encode the light’s frequency. Visible light oscillates in the 400 THz to 800 THz band, which is an exotically high frequency, not present in neurons or most electronic circuits. To make a practical implementation, image information is always encoded in summary form, whether from an eye or a digital photosensor.

In the case of a digital photosensor, the encoding is a pixel-level photon count, not the frequency of light. The photon count is only treated as blue if the receptor had a blue filter, thus the count is assigned to a blue channel. In the optic nerve, electrical signals use analog encoding schemes such as ‘center-surround’ encoding. In both cases, nothing in the resulting signal explicitly carries the frequency of the received light. Whatever your mind’s eye experiences from seeing blue is constructed from the encoded signal, not the original source.

Language gives us the label ‘blue’, but the label is arbitrary and unnecessary. One does not need language to fully experience and appreciate that something is blue. Moreover, having the label does not ensure one can experience the color even when subjected (sorry colorblind people).

One starting point common to consciousness research is to probe the neurological mechanisms by which an unremarkable electrical signal in the optic nerve manifests the phenomenal experience of seeing blue.

If we start with how a computer handles this, it needs explicit taxonomies to preserve abstract concepts, like color.

The brain likely employs a substantially different encoding system than a computer. A computer simply quantizes the intensity of each red, green, and blue subpixel to express image information. Does the brain even encode pixel information? Does it use methods of encoding that are intrinsically more meaningful and less dependent on convention? What might those be? Is the encoding itself fully responsible for the phenomenal experience, or is the encoding just a first step in a larger workflow? Does the blue seen by our mind’s eye exist outside the brain as a fundamental construct in nature? If so, does that require new physics to explain?

Artificial Neural Networks (ANNs)

ANNs rely on digital encoding and quantization to represent abstract constructs such as colors. A neural network never sees blue, it only processes number arrays. It is up to engineers to keep values that represent blue separated from values that represent red. Whatever encoding the human brain uses to transform microvolt electrical impulses into an experience of seeing blue is not present in today’s ANN architectures, there is no mind’s eye.

To make the leap from information processing to consciousness, data must acquire non-symbolic meaning. If we assume consciousness arises from neural anatomy, and not some spiritual origin, then the process likely depends on a more elaborate coding scheme than that used by ANNs. When the human brain creates the phenomenal experience of blue, the underlying processes must do more than simply assign a taxonomical label to a tensor. What computational process, if any, produces the subjective experience? Moreover, even if there is something spiritual or quantum mechanical to blame for this, we still have the matter of how the brain’s neural hardware interfaces with that. Regardless, the encoding problem is the gateway to probing consciousness. ANNs are simply not doing any of this.

Holographic Encoding

With Hodgkin-Huxley inspired ANNs dominating AI research in recent years, alternative architectures are getting less attention. At the risk of opening Pandora’s Box, one alternative encoding scheme that has been suggested for decades is holographic memory (distinct from holonomic brain theory).

A hologram differs from a conventional photograph in several ways. To create a photograph, one uses a lens to direct rays of light from points in a scene to pixels on a sheet of photosensitive film, or a photosensor, in such a way that each pixel only receives light from a corresponding part of the scene. Pixels are essentially directional photon counters. If you tear away half of a photograph, you get half of the scene. With a hologram there is no focusing lens, every pixel is exposed to the entire scene, not to directional rays. The higher the pixel density, the better it records details. If you tear away half of a hologram, the entire scene is still encoded in the remaining portion, but at lower resolution. A holographic pixel records raw photon counts but as sampled from the entire scene that is visible from that pixel’s position. The incoming light arrives from every point in the scene at once. Each pixel therefore records the resulting interference pattern, usually against a coherent reference beam. Neighboring pixels record a slightly different interference pattern, and so on. Every holographic pixel records some information from the entirety of the scene. This gives holograms a certain superpower to encode enough information to reconstruct the scene from any viewing angle. Holograms are robust against damage and can suffer large numbers of dead pixels without losing the general features of the scene. Moreover, the scene is encoded in the frequency domain, which is a fundamentally different storage paradigm.

Holographic memory is a concept based on a similar principle. Instead of quantizing individual datum or image pixels, the memory system processes frequency domain interference patterns. In the same way that a hologram records a complex scene from many perspectives, a holographic memory cell might have the capacity to capture a world model that encodes many perceived but related domains together (sight, smell, language, etc.) I use the word ‘might’ because holographic memory has not been fully explored from a practical engineering perspective, certainly not to the extent other architectures have.