But watch what happens when you leave it all up to the model to decide.

但是，如果完全由模特來決定，會發生什麼呢？

So first things first, what is bias?

首先，什麼是偏見？

Bias is often unconscious tendency to see, think, or feel about certain things in a certain way.

偏見通常是指以某種方式看待、思考或感受某些事物的無意識傾向。

Biases are somewhat hardwired into our brains to help us navigate the world more efficiently.

偏見在某種程度上是我們大腦的硬傷，幫助我們更有效地駕馭世界。

The problem is biases often lead to stereotypes.

問題在於，偏見往往會導致刻板印象。

And you'd think that this is a uniquely human problem, but surprise, it's not.

你會認為這是人類獨有的問題，但令人驚訝的是，事實並非如此。

It is a known issue.

這是一個已知的問題。

These models tend to default to certain stereotypical representations.

這些模式往往默認了某些刻板的表述。

Deepti is a staff research scientist at Runway, and she led a critical research effort in understanding and correcting stereotypical biases in generative image models.

Deepti 是 Runway 的一名研究科學家，她上司了一項重要的研究工作，旨在瞭解和糾正生成影像模型中的刻板偏見。

Now I think it's the best time to fix it because generative content is everywhere.

我認為現在是解決這個問題的最佳時機，因為生成性內容無處不在。

We don't want to amplify any existing like social biases.

我們不想放大任何現有的社會偏見。

There are mainly two ways to approach this problem, algorithm and data.

解決這一問題主要有兩種方法，即算法和數據。

Today, we're going to focus on data.

今天，我們將重點討論數據。

These models are trained on mountains and mountains of it.

這些模型是在堆積如山的數據中訓練出來的。

And because the data comes from us humans, here and there, our biases start to show up.

由於數據來自於我們人類，在這裡和那裡，我們的偏見開始顯現。

But just like we can uncover and prove our own biases, so too can AI models.

但是，就像我們可以發現並證明自己的偏見一樣，人工智能模型也可以。

And this process is crucial to ensure fair and equitable use of AI technologies.

而這一過程對於確保公平公正地使用人工智能技術至關重要。

The defaults that the model tends to produce are geared towards like younger population, very attractive looking women or men with like really sharp like jawline, one form of beauty that is pushed onto us by the society.

這種模式傾向於產生的默認設置是面向年輕群體、長相非常迷人的女性或下巴線條非常銳利的男性，這是社會強加給我們的一種美。

Within these models, there are a lot of repetition of certain types of data, over-indexing and sometimes a general lack of representation altogether.

在這些模型中，某些類型的數據存在大量重複、過度索引的現象，有時甚至完全缺乏代表性。

We noticed if the profession tends to be like of power, like CEO or doctors, it does tend to default to lighter skin tone people and like most likely perceived male as opposed to professions of not very high income do tend to default to like darker skin tone and females.

我們注意到，如果職業傾向於權力，如首席執行官或醫生，則默認膚色較淺的人和最有可能被認為是男性，而收入不高的職業則默認膚色較深的人和女性。

And this is not a true representation of the current state of the world.

而這並不是世界現狀的真實寫照。

This is a big problem we're starting to create solutions for.

我們正在著手解決這個大問題。

We call it diversity fine tuning or DFT.

我們稱之為多樣性微調或 DFT。

You might have heard of fine tuning before.

您可能聽說過微調。

It's widely used across models to hone styles and aesthetics.

它被廣泛應用於各種車型，以磨練風格和美感。

The way it works is by putting more emphasis on specific subsets of data that represent the outcomes you're looking for.

它的工作方式是更加重視代表您所尋求的結果的特定數據子集。

So if you want things to look like anime, you would fine tune with images like these.

是以，如果你想讓東西看起來像動漫，就可以用這樣的圖片進行微調。

And this actually works incredibly well.

而這實際上效果非常好。

Even with a very small subset of data, the model can learn to generalize from it.

即使只有很小的數據子集，模型也能從中學習歸納。

And this is what diversity fine tuning sets out to do with bias.

這就是多樣性微調的目的所在。

We generated a lot of text prompts, which are pictures of like female doctor, female doctor who belongs to a particular ethnicity and used a text to image model to generate synthetic images using these prompts.

我們生成了大量文本提示，其中包括女醫生、屬於特定種族的女醫生等圖片，並使用文本到影像模型，利用這些提示生成合成影像。

Deepthi and her team used 170 different professions and 57 ethnicities, and they generated close to 90,000 synthetic images to create a rich and diverse data set to diversity fine tune our model.

Deepthi 和她的團隊使用了 170 種不同的職業和 57 個種族，生成了近 90,000 張合成圖片，創建了豐富多樣的數據集，以便對我們的模型進行微調。

It was very exciting to see what we thought was like a simple solution of like augmenting the data and just retraining the model that helped in significantly like fixing the biases.

令人興奮的是，我們看到了一個簡單的解決方案，即擴充數據並重新訓練模型，這對修復偏差有很大幫助。

Diversity fine tuning is already proving to be an effective way to make text to image models that are safer and more representative of the world we live in.

多樣性微調已經被證明是一種有效的方法，可以使文字到影像的模型更安全，更能代表我們生活的世界。

I'm being optimistic that we will get to a place where the models are more inclusive.

我很樂觀地認為，我們一定能找到一種更具包容性的模式。