ImageNet

“Yeah. And the worst part is, it all comes down to a problem of logistics, not science. I’m as sure as ever that ImageNet is exactly what computer vision needs, if we could just finish the damn thing.” “Well, Fei-Fei…” he began, choosing his words carefully. “Everyone agrees that data has a role to play, of course, but…”

“He paused for a moment, then continued. “Frankly, I think you’ve taken this idea way too far.” I took a shallow breath. “The trick to science is to grow with your field. Not to leap so far ahead of it.”

反对自动标注的论据不是技术性的，而是哲学性的。我们意识到，即使是一个微小的算法捷径，都会违背ImageNet的使命。我们的目标是在每张图片中嵌入非合金（纯粹）的人类感知，计算机视觉模型可以基于整个充满智能火花的数据集被训练。

ImageNet是一个假设，一个赌注。受到我们自己生物起源的启发，激发真正的机器智能的第一步应该是（使机器）沉浸在完全的视觉世界。这种塑造我们进化的混沌和复杂经历也许对算法来说有相似的作用。

ImageNet 与 神经网络的结合

ImageNet是一个数据集，是机器睁开眼睛时看到的混沌世界。就像小婴儿睁开看到这个世界时一样，想象周围乱七八糟的点线面在发出奇奇怪怪的动作和声响。算法（algorithm）就是让这些混沌在机器的视界中逐渐清晰的过程，进而产生理解。如果类比机器学习到生物性，算法就像突触，或者缠绕大脑的线路。毕竟，生物智能并不是像设计算法，它是进化而来的。

不要检索，write

[322] "His idea was that our first glance at something is all it takes to understand it, at least on some level, right?" "Right. That includes the basic stuff like objects, of course. We are good at scanning "things" very quickly, but we re great at noticing how they are placed and arranged.The relationships between the things"

[328] The gist. We are not merely witness, but story-tellers. It feels like it was time for algorithm to learn how to do the same.

[333] "One to encode visual information and pair it with words, and the other to generate language. We will train our model on pairs of images and human writing descriptions>. “Well, there are definitely some unknowns to iron out, but I’m thinking the RNN generates each new word in the description conditionally, based on the words already in the sentence. That way, we’re describing the contents of the image while following whatever grammatical patterns have been inferred from the training data. The result, at least in theory, should be a completely novel description, in more or less natural language.”