search

JanRocketMan / regression-prior-networks

An official PyTorch implementation of "Regression Prior Networks" for effective runtime uncertainty estimation.
MIT License
35 stars 2 forks source link

Not understand why normal distributions can be used for classification #2

Open LinyeLi60 opened 2 years ago

LinyeLi60 commented 2 years ago

As in the case for classification, we can consider an ensemble of networks which parameterize multivariate normal distributions (MVN) We usually use categorical distribution for classfication, I don't known how to use normal distribution for classification, can you give some examples?

JanRocketMan commented 2 years ago

Hi, let me resolve some ambiguity of this sentence - it means that for performing uncertainty estimation in regression, we could, similarly to classification case, train an ensemble of networks which parameterize MVNs (compared to Categorical distributions in classification). The emphasis is on ensemble here, not MVNs.

At the same time, you can indeed use normal distribution for classification, e.g. by treating the outputs of your network as p(x|y) instead of p(y|x), and then performing Bayesian inference (similarly to LDA). E.g. see impostor networks.

LinyeLi60 commented 1 year ago

Thank you for your explanation.

------------------ Original ------------------ From: Sergey @.> Date: Tue,Apr 19,2022 7:38 PM To: JanRocketMan/regression-prior-networks @.> Cc: Linye Li @.>, Author @.> Subject: Re: [JanRocketMan/regression-prior-networks] Not understand whynormal distributions can be used for classification (Issue #2)