Neural networks are not the path to building self-aware Artificial Intelligence. It's time we understand the role of emotions and values....

A thesis on how humans have evolved self-awareness, and what the journey for a machine will require: emotions, values and social community.

Neural networks are exceptional prediction machines. The mathematical machinery is trivial, modular and scalable, almost shockingly so. So far, these algorithms mostly outperform all other learning algorithms under suitable conditions (richness of data, supervised learning, stability in environment etc). 

Their performance with recognition tasks regard-less data form (voice, text, image) holds promise that they will be effective encoders of the gamut of abstract concepts utilised by humans. Indeed, there is a strong argument that prediction and concept-encoding are two sides of the same coin: it is when concepts are encoded that our brains can identify and anticipate these concepts to support our survival in this world.

Prediction and concept survival are critical for efficiency in data processing, and thus for efficiency in survival (for example, the occipital lobe processes much more data from other regions of the cerebral cortex vs. directly from our optical nerve). When it comes to survival, better to recognise and predict than to wait for a data processor to fire up.

Yet there is no promise that neural networks in our current applications in A.I. are anywhere close to engendering self-awareness. So, just how has the ‘neural network’ architecture of the brain brought about self-awareness in humans?

First, lets briefly cover the evolutionary why of self-awareness. We lean on Richard Dawkins: self-awareness evolved because a brain that could simulate the future for survival scenario-planning (read prediction machines) which had a notion of a self-entity in that future, became aware of this self-entity as a distinct ongoing entity with which it self-associated. This was an evolutionary by-product or a ‘mutation leap’ however you want to think about it. It was retained because it strengthened the ongoing survival of these beings in social co-operation (more on this below).

Now, we are prone to the age old chicken-or-egg conundrum: does it not take a conscious aware being first in order for said being to recognise self-awareness? 

The answer is no. Our escape lies in the fact that you can encode a concept of a self-aware being before that being actually practises its own self-awareness. Self-awareness can be an incremental and continuous journey - just as the construction of other abstract concepts - and not a miraculous epiphany. 

So self-awareness is justified in evolutionary terms and has plausible path. Now what is the machinery that enables self-awareness? What were the milestones on this path?

The crux of the argument here is that the interaction of the limbic system and the nervous system has brought about self-awareness. This interaction both spawned a self-aware being and reinforced its mode of survival and proliferation via social cohesion using abstract concepts endowed with meaning in art, culture, spirituality, even love.

The ‘selfish gene’ has twin purpose: survival and self-reproduction. A human being - as a vehicle for an army of approx. 25,000 symbiotic self-ish genes - does better towards this purpose with a ‘prediction machine powerhouse’ guiding its decision making.

Whilst a prediction machine can support the concept of a self-aware being, how does it sustain it, and self-associate with it?

Without any features and attributes to anchor its existence, the lone self-aware identity is vacuous. It has a philosophically pure existence but not a tangible one. How long can an animal-brain hold and retain such a self-aware identity if it cannot dress this identity and make it more tangible? As a parallel, consider how quickly cultures must anthropomorphize and literally dress their concepts of ‘God’ (an otherwise clean concept of the omnipresent identity) to make this concept palatable.

Kant argues (Critique of Pure Reason, 1781) that logical computation without inputs (assumptions, beliefs, values) is ineffective towards reaching conclusion. Values, or initialised belief systems, are the basis for a logical machine to start processing and applying functional operators.

Neural networks cannot provide values. Values have immediate intrinsic meaning in and of themselves. Only one machinery in the body provides force for this and that is emotion. Emotion is driven by the functioning of the limbic system. Why it uniquely provides such force is because the experience of emotion, in and of itself, provides impetus for action in animals without requiring consciousness, justification or logical basis. This is not the case for actions driven by neural network processing which rests on cleaner concepts, and the logical stringing of these concepts.

The limbic system evolved as one supporting means for homeostasis (Feldman-Barrett, Dalmasio, 2017). It is a more gradual and sustained driver of body maintenance and body transformation than the nervous system. At some point the advent of linguistic means (evolved for the communication of pre homo-sapien tribal groups) started to build concepts around these emotions i.e., labelings of fear, love, anger, hate.

For something to be a value, it must have a property of ‘valence’ - the affective quality which refers to the intrinsic attractiveness ("good"-ness) or averseness ("bad"-ness) of an event, object, or situation.

Limbic hormonal activity by itself cannot provide values. Limbic hormonal activity needs to be translated into emotions (abstract linguistic concepts) which have valence. Emotions with valence become the basis for values. Thus a value system is born based on translating the sensations fuelled by hormones, into recognisable emotional concepts with valence.

Having values enables the self-aware being to dress the concept of a self-aware identity and associate this representation with itself. The self-aware being has now made its identity richer and more palatable. It is capable of functional operation by using its values as inputs for decision-making and direction.

The biggest challenge with this point of view is you have to believe the following causal path: neural network prediction machine evolves concept of self-awareness -> workings of the limbic system paired with emotional concepts being used linguistically, and thus socially, enable a value system -> with a value-system and the concept of self-awareness, a being becomes consciously self-aware. 

The lynchpin is that you have to believe that linguistic and social communication of emotional concepts preceded a sense of self-awareness. I find this plausible as I consider non-human primates to have emotional concepts without the level of self-awareness of humans, but I expect some will dispute this.

The creation of emotional concepts further supported survival because abstract concepts endowed with values (love, family, God, art etc.) supported the bonding, cohesion and empathy of human groups. Thus, the limbic system was critical for the development of a rich self-awareness and for the survival of this self-awareness in social groups.

Furthermore, this value system paired with abstract concepts propelled the recognition of meaning in our world. Meaning becomes possible for self-aware beings via the use of emotionally-endowed abstract-concepts such as love, family, God, art etc. 

Without this sense of meaning in day-to-day life, self-aware beings are prone to existential crises en masse. The realisation that their survival and proliferation (of no less than their self-ish genes) is the sole end is insufficient to entertain the conscious stream of a self-aware being.

The field of A.I. research has been poorer, and had less progress, because of its near fetishism for neural network architecture. There has been inadequate focus on applying lessons from the role and the architecture of the limbic system in producing an intelligence with self-awareness. 

I hope in the coming years we see more research, practical application and testing in this vein. I strongly believe this is our most promising path to developing artificial beings with sustained self-awareness. Appreciate how the human system processes and sees meaning in culture, music, art etc. and you can build an appreciation for why this mechanism is critical for self-awareness in intelligent beings.

[Thanks to Rajiv Bhat, Kunal Shah, Sriyansa Dash, Soham Sen and Kshitij Batra for their comments and inputs in building this thesis]