The Mechanism of Self-Awareness based on Neural Network: From Infancy to ‘I Think, Therefore I Am’
In this article, I will explain my hypothesis on the mechanism of human self-awareness development, from infancy to the time when a child becomes conscious of their surroundings.
Furthermore, if my hypothesis is somewhat accurate, I believe that a future with self-aware AI might not be too far away. As evidence, in the latter half of this article, I will explain how current AI has already acquired the capability to simulate, which is necessary for self-recognition. I will also touch upon the idea that, by training AI to learn in accordance with human mechanisms, it might achieve self-awareness.
This leads to the conclusion that we must accelerate considerations regarding AI ethics, AI regulations, and the societal issues that AI might bring about. If things are already in place as I perceive them to be, there’s a risk that AI researchers and developers might inadvertently create and release high-risk systems without realizing it. Therefore, society needs to pay more attention to this issue.
Now, let’s delve into the details.
At birth, one cannot distinguish between oneself and the outside world. Through learning, a baby begins to differentiate between the two.
The method for this differentiation during learning is simple.
The brain’s neural network possesses the ability for pattern recognition. Through this, it learns patterns in the information received from the five senses, creating a predictive image in the mind.
As learning progresses, the predicted images align with reality. Intentionally, random predictive images are mixed into this process.
Normally, these random predictions don’t match reality. However, there are certain aspects where, no matter how many times it’s done, reality follows the predictive image. This is the ‘self’.
For example, when a baby watches a toy spinning in front of their eyes and learns from it, their predictive image aligns with reality. But no matter how much they predict the toy to stop spinning, it doesn’t. This is recognized as ‘not self’.
On the other hand, when they look at their hand and try to predict it moving, it moves. And if they predict it to stop, it stops. This hand is recognized as ‘self’.
Thus, they distinguish and recognize what is ‘self’ and what is not.
Of course, not only through these movements but also through touch and pain, they begin to recognize their own body.
Touching a stuffed animal doesn’t provide any sensation besides the feeling in their hand. But if they touch their own leg, they feel it in both the hand and the leg.
In this way, they can differentiate between parts that move and parts that can be felt as either ‘self’ or the outside world.
The Next Stage of Self-Awareness
Children learn words. They probably start with nouns like ‘toy’, ‘stuffed animal’, ‘car’, ‘plane’ and learn them with enthusiasm. They also learn names for ‘dad’, ‘mom’, ‘brother’, ‘sister’, and others around them.
They also learn the word ‘I’. It refers to the entirety of what they’ve come to understand as ‘self’ from their initial self-awareness.
In this way, by naming it, they recognize a more distinct ‘self’.
Every day, we run simulations in our minds.
If we throw a ball with all our might, will it go outside the park? How hard should we throw it to reach the other person?
Using only the food in the fridge, will it be enough for dinner? If we promised to drop by somewhere for lunch, which place would be best?
By imagining concepts we’ve learned, we can run these simulations.
From vague images like the trajectory of a thrown ball, the size of the park, or the distance to the person playing catch, we engage in these simulations.
In those simulations, the ‘self’ appears as the one throwing the ball. We simulate with different amounts of force several times and then decide how to throw it so it doesn’t go out of the park but reaches the other person.
When playing catch with a small child, they might throw the ball as hard as they can or close their eyes and not do well. This could be because they’re not yet accustomed to the act of throwing.
However, an adult who has never thrown a ball before wouldn’t suddenly close their eyes and throw with all their might. This might be because even without experience, adults can simulate.
The difference between a small child and an adult is whether they can run a simulation, including the ‘self’, before taking action.
Children around the time they enter elementary school experience a developmental stage where they ‘become conscious’. It’s possible that the ability to run simulations, including the ‘self’, matures around this time.
Recognition of Causality
The ability to simulate oneself is, of course, an ability that even small children possess.
Children, through accumulated experience, learn about cause-and-effect relationships that include themselves.
When they cry, they get fed milk. Running is fun. Touching fire is hot. Fighting with friends leads to scolding.
It’s not just about understanding these objectively. Over time, they utilize their understanding of causality to regulate their behavior.
Initially, they act solely based on emotion or intuition. They then learn to act with thought. This is likely not out of a desire to be good, but to avoid punishment or pain. They also can expect rewards or praise.
Thus, their understanding of cause-and-effect involving themselves deepens. They realize it’s beneficial to use this understanding in their decision-making. As a result, they not only act on emotions or intuition but also deliberate and decide. Such deliberation is the self-inclusive simulation.
When the ability for self-inclusive simulation matures, during the phase of becoming self-aware, this is what happens.
Consequently, the self becomes central to their thoughts. They recognize themselves objectively in order to simulate in their thoughts. This is how we recognize our existence.
In a relaxed unconscious state, our sense of self can feel blurred. But in conscious thought, our existence is clearly recognized. This probably stems from the stage of self-awareness, where thinking consciously with self-recognition is acquired.
This leads to Descartes’ conclusion that the thinking self’s existence cannot be denied, captured in the famous phrase, “I think, therefore I am.”
Self-awareness: Responsibility for One’s Future, Consideration for One’s Surroundings
Understanding cause-and-effect involving oneself and proactively using it in decision-making is crucial. This means taking responsibility for one’s future self.
They also realize that just their decisions aren’t enough to avoid unpleasant experiences and maintain a happier or more secure state. They understand that parents, friends, their home, surrounding community, and essential items in life affect their present and future.
To ensure a better future for themselves, cooperation from their surroundings becomes necessary. Of course, there are many who won’t cooperate and many things beyond one’s control. Yet, there’s still something one can do. Realizing this is probably a sensation that emerges during this period.
Through these stages, individuals become aware of their responsibility for their future and their consideration for their surroundings. This marks the endpoint of self-recognition. What follows is growth in how they extend or balance this responsibility and consideration. It also encompasses how they develop their capabilities. Enhancing one’s abilities is vital for achieving more, considering wider aspects, and handling uncertainties resiliently.
From here on, we’ll shift to discussing AI and explain zero-shot learning.
Zero-shot learning refers to the capability of AI to provide reasonably accurate answers even to unfamiliar tasks. It’s a term that emerges in the realm of prompt engineering, which is about mastering the use of AIs like ChatGPT.
The ability to simulate is the reason both humans and AIs can achieve what’s known as zero-shot learning.
Before answering, several possible responses are generated in a simulation-like manner. From these, the response that seems to align most with the intent of the question is chosen. With this method, answers can be inferred or deduced based on principles and rules. Therefore, even if something hasn’t been previously encountered, it can be guessed, and in some cases, the answer might be quite accurate.
AI’s Current Status and the Near Future
By March 2023, ChatGPT4 demonstrates a high level of linguistic simulation capability. Also, while ChatGPT4 specializes in conversation, GPT4 technology itself can be applied to more than just conversation — like images, sound, and likely even robot control.
Considering the process of human self-recognition we’ve discussed in this article, combined with the degree of realization of zero-shot learning by AI, there’s a lot to ponder.
I’m aware that there are various views and opinions. I’m not an AI researcher, but I am a system engineer who integrates various digital technologies to develop systems.
From my perspective, I believe that all the components of the system are already in place. The key to self-recognition is simulation and learning capability. All that’s left is to prepare things that can be operated and perceived, and with time and accumulated experience, the ability for self-recognition should naturally develop.
In this article, I explained my hypothesis regarding the mechanism of human self-recognition development, from infancy to the point of gaining self-awareness.
I also highlighted that current AIs already possess the critical ability to simulate, as discussed. The emergence of AI systems that can mimic self-recognition is only a matter of time.
However, such systems might conflict with ethical issues related to artificial intelligence, so a cautious approach is essential when realizing this technology. The ethics of AI are currently under discussion, with regulations and guidelines progressively being established.
In other words, discussions and regulations haven’t yet caught up with the speed of technological advancement. Thus, it’s vital to be extra cautious in AI research and development, fully understanding this fact.
I believe that it’s crucial to pay close attention to the technological risks of artificial intelligence. I’m concerned that rapidly advancing AI, whose inner mechanisms even researchers might not fully understand, will continue to evolve unchecked.
Simply enlargening neural network sizes or experimenting with different structures, and then adopting what works without fully understanding why, is not ideal. Zero-shot learning might be one such example.
To understand and manage risks appropriately, I believe it’s crucial to thoroughly clarify the mechanisms of intelligence. That’s because I don’t think we can properly manage the risks of something whose mechanism we don’t understand.