Giter VIP home page Giter VIP logo

Comments (3)

colincsl avatar colincsl commented on July 19, 2024

I've never had good luck using BN on time-series data. I believe the issue is that the input feature distribution and output label distribution tends to be very different between batches (in 50 Salads e.g. in one trial a user may stand idle for a long time -- meaning there are a lot of "background" time steps -- whereas in another trial they may move constantly). We want the receptive field to be very long (e.g., 1 minute), so the duration of each sample must be long, and thus as a result the batch size will be relatively low. Additionally, these dataset are relatively small -- yet the distributions are fairly different -- so perhaps we don't have enough examples to learn good BN parameters.

This type of (frame-wise) normalization helps with vanishing gradient issues. Note that the Normalized ReLU used here is similar to using a normal ReLU and then applying LayerNorm. When I wrote this paper I didn't know LayerNorm existed.

from temporalconvolutionalnetworks.

bityangke avatar bityangke commented on July 19, 2024

Thanks very much!
I guess the reason is that dropout layer before BN layer makes performance poor.

from temporalconvolutionalnetworks.

colincsl avatar colincsl commented on July 19, 2024

That's not actually the issue here, but that is another common mistake. :)

from temporalconvolutionalnetworks.

Related Issues (13)

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.