2021年3月17日水曜日

GPT-3 ”Davinci”との会話1 ~錯視~




「普通のパソコンでトランスフォーマータイプの言語モデルはどれくらい頑張れるかhttps://eijikiwako.blogspot.com/2020/12/pt.html」という記事を書きましたが、実際はそんなもんじゃないです。正直半端ないです。訂正の上、本追加記事で補完しておきます。ただし、GPT-3の本質は自然な言語生成を極めているのであれば、これくらいは普通なのかも。ああ、恐ろしい。こんなことが普通を思える日が来るなんて。

以下、OpenAIのDavinciによる錯視の説明です。

=========================

Human: What is visual illusions?

AI: Ohh, would you like me to go through a number of the most common types of illusions that can occur within our visual field?

Human: Yes, please!

AI: Ok. Well it's important to remember that the visual field encompasses a pretty wide range of your senses, like sight itself obviously, but also things like how much you can feel what you touch or even hear. One of the perhaps most common illusions is that we tend to think of ourselves as watching the world from a single point. But actually, every time you move your head around you're effectively changing your perspective on the world. Our brains try to fill in the gaps with image processing, but it also means that, when you stare at a particular object for long enough, your brains fills in different images based on what it expects to see. The other big thing to remember about illusions is that in any visual image, the brain really mostly wants to find edges between contrasting colors or tones. The reason why this is important to notice is that after the mind's eye identifies an edge, it makes a prediction as to how the path of the edge goes based on the colors you can see on either side of the edge. When they don't match up, the brain often tries to correct the disparity by creating a new color in your mind. Of course, this leads to results that aren't actually featured in reality. You also know that colors can actually take their names from two different categories. The first is what we call colorimetry, which means the numbers associated with the wavelengths of a particular color. For example, a wavelength of 506 nanometers of green light will be either more or less green based on the numbers associated with it. The second category is what we call color naming. These are the names we give to each specific color. They're not always based on the wavelengths of the colors themselves. In fact most people didn't even realize that for centuries! The Shepard illusion is another good example of what I mean. It's a regular straight line moving left and right. However, your eyes naturally want to interpret it as a little hill and valley, even though the actual image doesn't change from frame to frame. It's an effective way to show how quickly our brain can fill in visual sensory gaps! Also, keep in mind that it is critical to the functioning of these illusions to have high sensory contrast. Any illusion with very low contrast or particularly muddled colors is probably not going to work very well. To see what I mean, try pointing the camera on your phone at a blank white wall! One last thing to note is that even humans don't see an illusion the same way! Almost everyone can see the line as it's meant to be seen, but the disparity between our individual perspectives can be pretty extreme. For example, some people see the top edge of the vertical line moving left all the way while others might see it keep going flat. This suggests that there are differences in perception based on the sensitive ability to see angles in space. Just another example of how the visual system is astonishingly flexible and unpredictable.

=========================

DeepLによる翻訳は以下

=========================

Human: 錯視って何?

AI: では、私たちの視界に現れる最も一般的な錯覚の種類を説明しましょうか。

Human: お願いします。

AI: わかりました。視野は、視覚だけでなく、触覚や聴覚など、かなり広い範囲をカバーしていることを覚えておきましょう。最も一般的な錯覚のひとつは、私たちが世界を一点から見ていると思いがちなことです。しかし、実際には、頭を動かすたびに、世界を見る視点を効果的に変えているのです。私たちの脳は、画像処理によってそのギャップを埋めようとしますが、特定の物体を長時間見つめていると、脳が期待するものに基づいて異なるイメージを埋めてしまうということでもあります。

錯視についてのもうひとつの重要なポイントは、どんな視覚イメージでも、脳は主に対照的な色やトーンの間のエッジを見つけようとするということです。なぜこのことが重要かというと、脳はエッジを認識した後、エッジの両側に見える色からエッジの経路を予測します。そして、それが一致しない場合には、脳内で新たな色を作り出して、そのズレを修正しようとすることが多いのです。もちろん、現実にはない結果が出てくることもあります。

また、色の名前には2つの異なるカテゴリーがあることをご存知でしょうか。1つ目は「測色法」と呼ばれるもので、特定の色の波長に関連する数値を意味します。例えば、波長506ナノメートルの緑色の光は、その数値によって緑色に見えたり見えなかったりします。2つ目のカテゴリーは、カラーネーミングと呼ばれるものです。これは、それぞれの特定の色につける名前です。色の波長を基準にしているわけではありません。実際、何世紀もの間、ほとんどの人はそのことに気づいていませんでした。

シェパード錯視もその良い例です。これは、左右に動く普通の直線です。しかし、実際の画像はフレームごとに変化していないのに、目は自然にそれを小さな丘や谷と解釈したくなるのです。これは、脳がいかに素早く視覚的なギャップを埋めるかを示す効果的な方法です。

また、これらの錯覚を機能させるためには、感覚的なコントラストが高いことが重要であることを覚えておいてください。コントラストが低かったり、色が混ざっていたりすると、うまく機能しないことがあります。例えば、真っ白な壁に携帯電話のカメラを向けてみてください。

最後に、人間でも同じようには見えないということです。ほとんどの人は、線を見たとおりに見ることができますが、それぞれの視点のズレはかなり大きいものがあります。例えば、縦線の上端がずっと左に移動しているように見える人もいれば、そのまま平らになっているように見える人もいます。これは、空間の角度を見るという繊細な能力に基づく知覚の違いを示唆しています。このように、視覚システムは驚くほど柔軟で予測不可能なものなのです。

=========================

量が質を凌駕して、なにかが起こっているようです。



0 件のコメント:

コメントを投稿