It's unfortunate that this needs to be said.
Models should be as accurate as possible. The place to insert your values is in the decision processes that make use of them.
-
-
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
Tämä twiitti ei ole saatavilla.
-
Some measure of closeness to reality? (Serious point: if your loss function includes subjective values, you are probably biasing it away from the truth)
Keskustelun loppu
-
-
-
What if your labels and/or the data is confounded? And your model ends up learning something quite different from what you thought it was learning.
-
E.g. say you have 2 sets of images, people wearing uniforms and people not wearing uniforms. You train a classifier and it shows remarkable performance even on a held out test set. Unfortunately, the uniformed images all had similar backgrounds.
- Näytä vastaukset
Uusi keskustelu -
-
-
Defining the training objective is the hardest past. There are tons of models that are highly accurate and very wrong.
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
To be fair, checking that the model errors are as random as possible can be a legitimate part of the model training process. It can go a long way towards addressing (real) concerns about model biases.
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
I understand and agree with what you're saying, but what distribution of inputs are you going trying to get your accuracy score on? Equiprobable classes or real-world distribution or relevant subset of real world? Those answers are not obvious...
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
There ~is a way to reduce bias in a model that some entity wants reduced: Create a dataset, which can grow dynamically of challenge/response pairs for "good" and "bad" results. Filter or adapt the model to get to a desired good/(bad+good) threshold. At least it's then explicit.
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
Deployed model should be as useful as possible. In a 1% positive data,the most accurate model is all negative which is totally useless. Here you reduce accuracy to improve usefulness. Luckily usefulness here can be mathematically described(F1 score). Reality is harsher...
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
Lataaminen näyttää kestävän hetken.
Twitter saattaa olla ruuhkautunut tai ongelma on muuten hetkellinen. Yritä uudelleen tai käy Twitterin tilasivulla saadaksesi lisätietoja.