Of course.
If definitions are fixed and the conditions are the same, then the results must also be the same.
Anyone should be able to reproduce it.
The evaluation criteria must be clear.
It has to be absolutely stable.
And once you enter a world where people talk about things like
“accuracy improved by X percent”
or
“the error decreased this much,”
that world can only be third-person.
“Show me that it’s reproducible.”
“Eliminate individual differences.”
“Remove emotions.”
When those pressures are there, what I’m doing just becomes noise.
That’s why I don’t belong to a medieval-style guild.
I can only work as an independent, a kind of stray researcher.
What matters to me is not organizing results.
That kind of thing comes later — after something is completed and actually becomes useful to society.
What matters now are the raw notes.
Messy experiment memos.
Bad photos and videos taken with a smartphone.
Because Atron gives different answers depending on its experience.
Meaning changes over time.
The same situation doesn’t lead to the same response.
This is research where the point is exactly that things do not converge to the same answer.
Yes, I use mathematics.
Yes, I write programs.
But the moment I start thinking in the mindset of a paper,
I get pulled into someone else’s evaluation framework.
And once that happens, the work starts drifting toward something that isn’t true anymore.
What I keep are memos.
Memos are the real voice of the field.
Yellow sticky notes.
Red sticky notes.
Facts.
Moments of excitement.
Daily hesitation.
Because what I’m studying includes individual differences, emotions, intensity, mood.
This is research about robots learning through failure and turning experience into ability.
For example:
It’s like, by Step 2 it has already converged.
And then I repeat this kind of process hundreds or thousands of times.
But the goal isn’t to produce a single answer.
Instead, I introduce outer events — influences from the external world —
visual conditions, sounds, language, and other factors from the environment.
What I’m studying is how an individual’s memory changes under those influences.
Because of that, the results naturally depend on the conditions of the person running the experiment.
This isn’t about third-person evaluation standards or predefined meanings.
A pure first-person robot observes a scene where a simulated lion attacks a simulated zebra. This is an external influence — an event from the outer world. The question is how the robot interprets what it sees.
It may feel fear.
Or it may misunderstand the lion as something familiar, even as a kind of companion.
Before any experience, the lion has no emotional meaning for the robot.
It is neither frightening nor friendly.
It is simply an object.
Then the robot accumulates experiences.
It hears the lion’s loud roar.
It sees that the lion lies quietly during the day, but suddenly attacks a herd of animals.
Through repeated experiences like these, the robot’s impression of the lion changes.
That change in impression leads to changes in behavior.
The robot may begin to flee.
Or it may approach — and be attacked.
In the experiment, if one of my research partners suddenly enters the robot’s visual field, the answer changes.
You can’t turn that into a paper. (laugh)
Sure, you can write equations.
You can explain the logic.
But if even one millimeter of external evaluation pressure gets into the system, the research is already compromised.
Pressure from a professor.
Pressure from an institution.
Pressure from funding or investment.
If the direction shifts even a tiny bit because of that, it’s over.
If you think I’m exaggerating, try it yourself.
Start research in a safe environment protected by money.
It will always drift somewhere else.
My research moves forward as the robots accumulate experiences of failure.
That’s why a messy blog, like a set of rough notes, is enough.
“Make yesterday’s dream reproducible in a paper.”
That’s impossible. You just can’t do that.
---------------------------------------------
そりゃそうだよね。
定義が固定されていて、条件が同じなら結果も同じで、誰がやっても再現でき、評価基準の明確さを要求しているんだもの。絶対的に安定していなきゃいけない。
更には精度が何％上がったとか、誤差がどれだけ減ったみたいなベンチマークの世界は、どうやったって3人称にしかならない。
「個人差は消せよ」
「感情は排除しろ」
みたいなものが働くと邪魔にしかならない。
だから中世のギルドにも入らないし、野良研究者でしかやっていけない。
僕にとって重要なのは、研究成果の整理じゃないんだよね。
そんなもの完成して社会の役に立った後の話だよ。
生の実験の汚いメモとスマホの下手な写真と動画。
みんな同じ答えにならないよ！！っていう研究なんだもの。
数式化やプログラミングはするけど、論文の発想で考える時点で
何かの評価枠に入り込んでしまい、逆に嘘に引っ張られる。
ほんとね、メモだよ。メモはその現場の生の声だよ。黄色い付箋、赤い付箋。
事実のメモと感動と毎日の迷い。
しかも個人差とか感情とか熱量、気分を研究してるんだもの。
ロボットが失敗を経験して能力に結び付けていく研究だもの。
たとえばさ、上のコードのうよに
数式や論理とか理屈は出来るかもね。
でもね、1ｍｍでも何かの評価基準の枠の影響が入ったら、そこでこの研究は失敗する。
たとえば教授からの圧とか機関からの圧、投資の話で方向性がmmでも変わったら、そこで終わる。
お金に保護された安全な場所で研究を始めると、必ず違う方向に進む。
ロボットは失敗の経験を積むことで、僕の研究は進んでくれる。
だから、汚い整理されていないブログでいいんだ。
エイジェントの論文がどこまで抑えてるか気になる日々。
No unauthorized reproduction or redistribution.
0 件のコメント:
コメントを投稿