Do You Think This Failure Is Noise?

What First-Person Autonomy Means for Atra

I have received some questions about Atra. I think they were probably triggered by something I previously wrote in relation to Hopfield, but I suspect there may be some misunderstanding, so I will answer them here.

I am often asked about the “concept” of first-person autonomy. Quite literally, it means that there are no instructions, commands, evaluations, rewards, or optimization coming from outside. Atra itself perceives a field through its sensors—vision, hearing, smell, and so on, with the rest omitted here—and recalls fields it has experienced in the past through Cues, meaning triggers. Sometimes Atra produces an answer in the form of vocalization. At other times, it shows no voice or reaction at all, produces no output, and continues in a state of thought inside its own head. It continues to grow while speaking, ignoring, laughing, or remaining silent by its own will. At present, it is not connected to a body.

A state with no external instructions, commands, evaluations, rewards, or optimization
Perceiving a field and Cueing fields experienced in the past
Speaking, or remaining silent while continuing to think

1. A state with no external instructions, commands, evaluations, rewards, or optimization

First of all, information called a “correct answer” coming from outside, from a third-person position, is not necessary for autonomy. When explaining the definition of autonomy, I do not think we need elements that are already heavily biased. Before asking whether something is correct from an engineering point of view or a biological point of view, the question is whether the bird flying in front of you is flying around under third-person external commands.

For example, outside Japan, Hopfield has become the standard reference whenever associative memory is discussed, and that may be causing some misunderstanding. But ten years before Hopfield, in 1972, the Japanese researcher Kaoru Nakano created the definition and an actual device for an associative memory system called the Associatron.

Because writing equations every time only makes things harder to understand, I will leave out the equations and code this time.

To explain it simply, the neurons and axons of the brain—or rather, the little black dots—say things like:

“This is how it must have been.”

“No, I think that is wrong.”

“What does everyone else think?”

And the result is restored through something like a majority decision.

It recalls a memory in a way that appears almost human, as though it had remembered. Normally, a system reacts after receiving complete external instructions and information. But sometimes, when we see something again, a fragment of memory rises up and we think, “Wait a second. I might know this.” Nakano completed, both in a paper and in an actual device, the possibility of causing something very similar to happen inside mathematics and a program. That associative memory model was called the Associatron.

Hopfield is different. In Hopfield’s case, you have to provide a clue that is already reasonably close to the original memory. Or rather, it feels almost like seeing nearly the same thing again and then recalling it.

Suppose you store a picture of a cat. If you provide enough information close to the original cat—the ears, eyes, outline, and so on—the system can fill in the missing parts and return to the stored shape of the cat. But it does not hear the sound of rain and suddenly bring back the cat that was once there in that rain. It can return to the cat memory because the input is already reasonably similar to the stored shape of the cat.

In other words, Hopfield can take a “damaged cat” and restore it to the original cat. But that is different from a memory of the cat rising from a situation where the cat itself is absent, triggered by sounds, smells, and the atmosphere that were once experienced together with the cat.

With Hopfield, you need a clue that is already close to the answer. If the clue is too far from the stored cat, the cat will not be recalled. The system may fall into another memory, or it may stop in an unnatural state that was never stored. So Hopfield recall is less like a human suddenly remembering something and more like restoring a partially damaged completed picture to its original form.

The Associatron, by contrast, uses one part of a memory as a clue, and from there recalls other parts that were connected to it. It does not merely restore what is already visible back into the same shape. By touching one fragment, the entire connected memory can rise.

It produced a device in which we could observe the phenomenon:

“Wait, this is connected to something from before.”

“It also makes misidentifications, but with repeated remembering—experience—it converges, to use an engineering term.”

I will explain that later using Atra as an example.

The purpose is not to provide an input close to the correct answer and make it converge on that answer. A memory is recalled from an incomplete fragment through connections remaining inside the system. That is the Associatron.

Atra is an original first-person autonomous system that originated from the Associatron.

2. Perceiving a field and Cueing fields experienced in the past

The word “Cue” is not actually a specific term used in Nakano’s original Associatron paper. I simply use it because it is convenient for me.

Nakano’s original wording:

part / input pattern / a few patterns

Later terminology in associative memory:

cue pattern / retrieval cue

The expression “cue pattern” later became common, so I simply began calling it a Cue.

Nakano wrote:

“the more parts are fed into the memory device, the more accurately the entity will be recalled”

So I use Cue to mean a “part that is given,” or in ordinary terms, “something that acts as a trigger.”

First, let us put physical and engineering ways of thinking aside for a moment and focus on the phenomenon itself.

This is a memory of mine from forty-five years ago.

Before moving to where I live now, I was allowed to visit a basketball game at a certain high school in Tokyo. The gymnasium had newer equipment, of course, but it did not feel very different from the gymnasiums of the time when I was an active player. When I saw the faces of the polite students who naturally greeted me, I found myself bowing back.

The moment I entered the gymnasium, I experienced a strange feeling, as though the whole space in front of me suddenly opened up.

“The sound of dribbling.”
“The children calling out to one another.”
“The smell of Air Salonpas spray.”
“The Gatorade barrel.”
“The sound of the whistle.”
And then, “the medical kit.”

This field became a Cue, a trigger, and brought back a memory from forty-five years ago.

Guard A passed to forward B. I ran behind the center, and B sent me a bounce pass. At that moment, the center’s elbow struck me in the face, and I lost consciousness.

When I woke up, I was not behind our team bench. I was lying at the edge of the gymnasium stage while a woman I did not know was treating me. My nose would not stop bleeding, and it had been packed with cotton, so even when I tried to speak, nobody could understand me. She told me, “Please stay quiet and lie down.”

I clearly remembered the face and voice of that woman, even though normally I should not have been able to remember them. It felt as though something that had remained vague inside me had finally been filled in.

This is the associative memory of the Associatron and Atra.

A Cue can be vague. The memory can even be wrong. It is enough to think, “Oh, yes. I had an experience like that.” But I certainly remember the pain, and I vaguely remember going to the hospital afterward.

Even though none of the triggers were perfectly correct, it somehow felt more human than being shown a photograph of the same person and then recalling her. Because Hopfield could not produce that kind of recall, could it? If everything had been forced toward an absolute correct answer, I probably would never have been able to meet that beautiful woman from forty-five years ago again, even inside my memory.

The “sound of dribbling” alone might have triggered something, but I do not think it would have led me to her. Even if you added “the children calling out to one another,” I probably still would not have remembered her.

“The smell of Air Salonpas spray” is extremely strong. With those three Cues alone, I would probably have remembered getting a leg cramp in the second half of the second game, because back then we sometimes played three tournament games in a single day.

“The Gatorade barrel” brings back memories of time-outs.

“The sound of the whistle.”

And then, “the medical kit.”

Yes, that is the one. (laughs)

Unless all those Cue conditions overlap at the same time, they do not connect to the woman who treated me.

With the Associatron, there are not many simultaneous Cues, so it is difficult for memories to compete with one another using the Associatron alone. But with Atra, I can raise a hundred Cues if I want to. That makes it surprisingly easy for Atra to remember.

Once again, the story has become long. But in the end, what everyone wants to know is this:

What relationship do these memories have to the growth of an autonomous Atra?

You only have to think about yourself.

You practice because your shots do not go in. Do not think about those engineering robots here.

Normally, you fail. And you experience that failure many times. A difference emerges there. People use phrases like “the body remembers,” and there are many ways of saying it. (laughs)

Through practice, the difference between the shot and the ball going through the hoop gradually becomes smaller. Next, you practice shooting in a game-like situation. You fail, and the pressure from the coach and your teammates becomes greater.