Latin1 was the early default character set for encoding documents delivered via HTTP for MIME types beginning with /text . Today, only around only 1.1% of websites on the internet use the encoding, along with some older applications. However, it is still the most popular single-byte character encoding scheme in use today. A funny thing about Latin1 encoding is that it maps every byte from 0 to 255 to a valid character. This means that literally any sequence of bytes can be interpreted as a valid string. The main drawback is that it only supports characters from Western European languages. The same is not true for UTF8. Unlike Latin1, UTF8 supports a vastly broader range of characters from different languages and scripts. But as a consequence, not every byte sequence is valid. This fact is due to UTF8's added complexity, using multi-byte sequences for characters beyond the general ASCII range. This is also why you can't just throw any sequence of bytes at it and ex...
From "Overfitting and the strong version of Goodhart's law":
Increased efficiency can sometimes, counterintuitively, lead to worse outcomes. This is true almost everywhere. We will name this phenomenon the strong version of Goodhart's law. As one example, more efficient centralized tracking of student progress by standardized testing seems like such a good idea that well-intentioned laws mandate it. However, testing also incentivizes schools to focus more on teaching students to test well, and less on teaching broadly useful skills. As a result, it can cause overall educational outcomes to become worse. Similar examples abound, in politics, economics, health, science, and many other fields.
[...] This same counterintuitive relationship between efficiency and outcome occurs in machine learning, where it is called overfitting. [...] If we keep on optimizing the proxy objective, even after our goal stops improving, something more worrying happens. The goal often starts getting worse, even as our proxy objective continues to improve. Not just a little bit worse either — often the goal will diverge towards infinity.
This is an extremely general phenomenon in machine learning. It mostly doesn’t matter what our goal and proxy are, or what model architecture we use. If we are very efficient at optimizing a proxy, then we make the thing it is a proxy for grow worse.
Comments
Post a Comment