“Artificial Consciousness: Integrating Neural Networks and Psychoanalysis for the Future of AI”​

Luciano Fagundes

Neurocomputing – Conclusion Paper

National Institute of Space Research

(*) Translated to English by ChatGpt4

Abstract

In this work, we will use the current knowledge of neural networks and psychoanalysis to propose the path to be taken to evolve Artificial Intelligence (AI) and achieve the development of Artificial Consciousness (AC). AI has been primarily based on the neural structures found in the human brain, which has led to a tremendous leap in the ability of computers to process and respond to questions. This research aims to take a step further. We attempt to propose a new structure for networks to include psychoanalytic concepts in their topology. We propose a new line of reasoning for the development of responses, incorporating control mechanisms, feelings, and subjectivity into the networks. All this is done by applying concepts theorized by psychoanalysis to explain the high-level functioning of human consciousness, without adhering specifically to any one neural structure.

Keywords: Artificial Intelligence, Artificial Consciousness, Neural Networks, Psychoanalysis, Human Brain, Cognitive Development, Subjectivity, Control Mechanisms, Reasoning Structure

Introduction

This article arises from reflections on neural networks, psychoanalytic structures, and some famous “failures” found in the results of Neural Networks.

Neural networks are mathematical mechanisms that can connect any input to any output. If we abstract this concept, it is as if, during training, we were searching for a dimension where all questions and answers align. The weight changes during training and various topologies do nothing but find a dimension (training result) where the answers to our questions are located. Even more interesting, they identify the most probable answers to questions that have not yet been asked (network generalization).

However, blindly trusting the neural network and its training seems unfinished to me. Always seeking a parallel with the human brain, there are control mechanisms likely running within the neural network, but operating separately. There must be some short-term memory system that allows reasoning, inference of the external world, and other processes. These mechanisms are not referenced in current neural networks.

Perhaps one of the main issues we face today with neural networks is how to express reasoning regarding a response. For example, if we take an LLM (Large Language Model) network and ask how to decorate a living room, it will respond by describing what objects to place, how to arrange them, etc. If we ask how it arrived at that response, it will mention a particular technique, such as lighting or design elements. But if we analyze the two responses, we find no clear line of reasoning from nothing to the final answer. Both the response and the explanation might not even be connected—they are simply outputs from a dimension found during training, which, aside from some randomness in the seed of the question, would always be the same in all cases. The network did not explain how it arrived at the answer; it only gave what a person might have said when asked how they arrived at the response. This may seem circular and confusing, but that’s precisely the point.

There is a classic example where the neural network arrived at the “correct” answer by the wrong path but failed to generalize and respond appropriately to real-world examples. This was the case of classifying a dog and a wolf [GEIRHOS1]. In this case, several labeled photos of wolves and dogs were shown to the network. The neural network ended up learning that the background of the photo was what defined whether it was a dog or a wolf. If the background was wild—snow, forest, etc.—it was a wolf. If it was a sofa, garden, or living room, it was a dog. As a result, the network gave grossly incorrect answers when presented with photos of dogs and wolves outside the training set. The AI could not explain what it was considering to solve the problem; humans identified the issue through trial and error, deducing the root of the training problem.

This is where psychoanalysis comes in. Today’s neural networks seem very focused on the brain as a question-answering machine. Psychoanalytic theory offers a broader abstraction and attempts to theorize about the internal components of the brain—not in terms of neural structures, but in terms of control structures that enable the brain to interact with the world. While psychoanalysis focuses on traumas and failures that occur during the formation of these structures, in our case, it is worth understanding how these abstractions “should” function, so we can restructure our neural network topologies accordingly.

Freud and the Psychic Apparatus

Freud proposed the psychic apparatus in 1900 [FREUD1], initially concerned with dream interpretation. For him, the psychic apparatus consisted of a sequence of steps, resembling a comb. On one side was the conscious, linked to all body sensors—skin, stomach, eyes, etc.—and the motor part. Freud viewed this apparatus as something akin to a neural network, with sensory inputs passing through intermediate layers and returning “motor” commands. For example, if a lion appears in front of you, the information enters through the eyes, passes through the neural network, and returns a command to your leg muscles, making you run.

The inner teeth of this comb represent the layers of our neural networks, processing the data, understanding it, and finding the appropriate response. Learning would occur through repetition and reward: each time an action succeeds, the same neurons are activated, strengthening that area and making that “memory” permanent. This is very close to current neural network training theories.

The most important aspect of this model is Freud’s first mention of the unconscious. That is, a significant portion of the brain’s neuron layers is hidden, and it is impossible to identify all factors influencing the generation of a response.

Lacan and the Brain as Language

Another prominent psychoanalyst was Jacques Lacan, who sought to formalize Freud’s theory mathematically. Perhaps his most important contribution was recognizing that the unconscious described by Freud was organized like a language [LACAN1].

This opens the door for us to imagine neural networks that communicate with each other through language—not only by neuron activation but through internal “conversation” in a language that could be translated into human language. We could have a neural network responsible for threat detection, whose output might not be mere probabilistic values between zero and one but could also be a phrase such as “I think there is a lion behind that tree.” In other words, the outputs of the networks could consist of commands and linguistic observations, translatable and understandable to humans.

The major advantage of representing the artificial “unconscious” as a language would be the emergence of a line of thought. If each internal structure has a linguistic output, we could maintain a log of these artificial “thoughts” and perceptions, comparing them with the final result.

Ego, Superego, and Id (Unconscious)

In Freud’s second topography, he advances the psychic apparatus by introducing the concepts of Ego, Superego, and Id. At this point, abstraction progresses further. Freud realized we need more than a simple question-answer neural structure. If this were enough, everyone would be the same; starting with a blank brain (a clean neural network) and training it the same way would yield identical neural networks. But we know this is not true for neural networks, much less for humans.

The Ego [FREUD2] is responsible for the survival of the individual, receiving external perceptions, comparing them with internal needs, and acting. To find balance, the Ego requires assistance from two other controllers, the Superego and the Id.

The Superego consists of all the rules acquired during life—what is allowed or not by law, guilt, etc. It serves as the Ego’s legal advisor. For example, if the Ego is hungry, sees a hot dog cart but lacks money, it asks the Superego, “Can I take the hot dog?” The Superego responds with something like, “If you do that, you could go to jail because you can’t pay.” This is a consultation because the Ego will make the final decision. The hunger may be so severe that the individual risks their health, leading the Id (unconscious) to say, “Eat or die!” The individual may then commit a crime to survive.

The Id or Unconscious is the third major controller and the most primal element of consciousness. It represents everything the individual has always had, without any filters. All knowledge acquired over time is stored here, much like dense layers in a neural network, where everything is mixed and unstructured but still capable of influencing all generated responses.

In this example, we see the Unconscious modifies the response of the other controllers, expressing the individual’s “true” will. As seen, the Unconscious affected the Superego’s response, giving validation for a reprehensible act, yet it was “necessary” for the individual’s survival. Again, communication between these controllers would occur through language and could be readable by humans. This could explain how a “desperate” robot might justify disconnecting another robot to recharge itself, displaying the emergence of artificial “feelings.”

The Desire of the Other

Now we return to Lacan. For an AI to be promoted to consciousness, one of the first things it must develop is the capacity to desire something. Fortunately for us, Lacan worked on this concept. He also supported the idea that the brain starts as a “blank slate.” Therefore, human beings do not know what they desire when they are born. They spend their childhood trying to find their personality and what they want to do. No one is born wanting to be a systems analyst. So, how do we get there?

A neural network today has only one desire: to answer the questions posed to it. The person who built the training, validation, and test sets and marked the necessary answers imposed this desire on it. This was the desire of the Other.

Now, if we want a conscious neural network, it must learn to desire on its own. Just like a human, it needs time to observe the world and figure out what it wants to do in that world.

Lacan’s theory posits that desire is always the desire of the Other [LACAN2]. A full article would be needed just to talk about Lacan’s concept of the Other, but let’s focus on computational structures here.

I’ll simplify this next section a lot. I know it’s not easy to create something like this, and much research would be required to develop a module capable of doing what is proposed. However, this is the original idea for how we might start the path toward artificial consciousness.

We would need to create an alternative training mode that observes the external world and learns from the actions and results of others (learning in the third person, not just the first). A very rudimentary example, which seems to happen with humans in their early years, is this: The basic neural network would need to be trained to understand smiles, happy, and sad expressions. From there, it would begin to establish knowledge that if, for example, someone fishes and catches a fish, everyone around them is happy; if a lion appears, everyone screams and runs away; if a lion eats a person, everyone is terrified; if there are worms in raw meat, everyone makes a disgusted face.

All these situations occur in everyday life in various forms in communities worldwide. This perceptual interaction with the world begins to populate the imagination of children during their development. Without realizing it, the unconscious is trained. This training is almost wordless, based on observation and experimentation, and it will guide the individual’s interaction with the world throughout their life.

It is from this type of analogy that Lacan draws the concept that desire is always the desire of the Other. Extending the idea a bit, when the child gets a little older and sees their father writing computer programs, or a YouTuber who became a millionaire by selling a tech company, or watches a cool video game, one of these actions and results from others may lead the neural network to conclude that it, too, wants to be like one of these social representations and thus may desire to become a systems analyst.

So, the desire of this artificial consciousness would be an attempt to survive in the world through a balance that the Ego achieves between the results it has seen others obtain (the desire of the Other), the laws imposed (consulting the Superego), and all past experiences, both good and bad, stored in the Unconscious (Id). This would be the basic recipe for bringing desire to a neural network.

Subjectivity and Intuition

This is where Lacan’s concept of the Real comes into play [LACAN3]. As many in the humanities don’t feel the need to explain things explicitly, for Lacan, the Real is what is inexplicable, what cannot be said. In our context, I see the Real as the randomness inserted during the network’s training.

Why wouldn’t this artificial consciousness be just another robot, programmed to do certain things? Well, in the end, it would be. But it’s this “certain thing” it’s programmed for that will vary greatly and will not necessarily be controlled by a human.

If we think about neural networks trained with backpropagation, we know that we input data, calculate the difference between the expected and obtained response, and use techniques to find the neurons that contributed most to the answer to adjust them, thereby minimizing the error in the next attempt. Now, in the exact sciences, could anyone claim that we are adjusting the weights correctly and that the problem is solved? No! It is impossible to definitively resolve this issue. We know we are merely trying to minimize the error. Nothing guarantees that the error will even be minimized; we will just test inputs and outputs for an arbitrary amount of time until we reach a set of “satisfactory” responses. Thus, if we train 1000 different neural networks with the same data, there’s a chance that we’ll have 1000 different networks at the end of the training. All of them may give similar responses, but there’s a significant chance that randomness will interfere in the training process and leave different weight values in various positions.

Accumulating everything we’ve described in this article, we can see that the factor of randomness in training artificial consciousness would be extreme. Since learning would be based on observation of external environments, one network might conclude that drug use is normal and beneficial, while another might conclude that its purpose is to create works of art. Everything depends on the environment in which the network was created. There is a whole other psychoanalytic theory about the influence of the environment on the personality of individuals, described by Winnicott [WINNICOTT1]. But for this article, I’ll stick to Freud and Lacan to avoid overextending the discussion.

So, this third-person supervised learning algorithm (if such a thing exists), through the observation of external actions and results, is what would bring subjectivity to artificial consciousness. It would be the worldview acquired during the life of this network that would provide the unique subjectivity to the decisions and responses it produces.

Conclusion

This is still largely a theoretical work. I don’t think we have any technological barriers preventing us from creating what has been described here today. On the other hand, I also don’t think that what has been described alone is enough to create artificial consciousness. I could still describe an entire module about Lacan’s Mirror Stage, where our network could identify itself as an individual, perceiving its limits and characteristics, validated by the famous Other.

Therefore, the goal of this work is merely to lay the groundwork for my research into the creation of artificial consciousness. There are still many questions to be answered, but I believe that this kind of inquiry into the construction of artificial networks will be the next big step in the development of neural networks.

We must move beyond Freud’s first topography, with networks that merely have inputs and outputs, and move into the second topography, where we bring in more sophisticated control mechanisms that can truly interact with the world in a flexible and constructive manner.

References

[FREUD1] FREUD, Sigmund. The Interpretation of Dreams. Translated by Márcio Pugliesi and Paulo César de Souza. 8th ed. São Paulo: Companhia das Letras, 2019. 600 p.

[LACAN1] LACAN, Jacques. Writings. Translated by Vera Ribeiro. 2nd ed. Rio de Janeiro: Zahar, 1998. 848 p.

[FREUD2] FREUD, Sigmund. The Ego and the Id. In: ______. Complete Works. vol. 16. Translated by Paulo César de Souza. São Paulo: Companhia das Letras, 2011. pp. 15-72.

[LACAN2] LACAN, Jacques. The Seminar, Book 4: 

Leave a Reply