Presentation is loading. Please wait.

Presentation is loading. Please wait.

Neural Networks in RStudio with the H2O module

Similar presentations


Presentation on theme: "Neural Networks in RStudio with the H2O module"— Presentation transcript:

1 Neural Networks in RStudio with the H2O module
Raymond Fraizer

2 Goal To learn how to implement a neural network using RStudio and H2O.
To learn how to use the H2O web interface. Get experience with sensor data.

3 Dataset Background Found PUC-rio dataset at UCI repository.
The paper “Wearable Computing: Accelerometers’ Data Classification of Body Postures and Movements” by Ugulino et al.[1] They used C4.5 with AdaBoost using 10 boosting iterations and 10-fold cross validation. Their overall accuracy was 99.4%.

4 What is H2O A machine learning platform with interfaces for many languages such as R, Python, Scala, Java, etc. It also has a web interface for interactive control. It has many common models: GLM, GBM, DRF, PCA, K-Means, “Deep Learning”, and “Naive Bayes”.[2]

5 What is H2O The platform is setup for easy model building by automating common pre- and post-processing tasks and providing good defaults for most settings. You don’t need to worry about missing values, categorical features, or scaling your data. H2O deals with all of that for you.[3]

6 Method Using RStudio and the H2O module I trained a few individual neural networks of different configurations to compare results. With so many parameters I tried using random grid search to choose the activation unit, the number of nodes, and in some runs l1 and l2.

7 Results

8 Conclusion Small networks did poorly, if you can call 92% or better accuracy poor. The node networks all did better than 98% accuracy, but with some over fitting. The large 400+ node networks that did not use dropout exceeded 99% accuracy, but with a lot of over fitting. Even without using an ensemble the neural network came close to the accuracy of the original work’s 99.4% for a 10 times AdaBoosted 10-fold validated decision tree.

9 References Ugulino, W.; Cardador, D.; Vega, K.; Velloso, E.; Milidiu, R.; Fuks, H. "Wearable Computing: Accelerometers' Data Classification of Body Postures and Movements.", Proceedings of 21st Brazilian Symposium on Artificial Intelligence. Advances in Artificial Intelligence - SBIA In: Lecture Notes in Computer Science. pp Curitiba, PR: Springer Berlin / Heidelberg, 2012. Aboyoun, P., Aiello, S, Eckstrand, E., Fu, A., Landry, M., and Lanford, J. (Sep. 2016). "Machine Learning with R and H2O.", (accessed Apr 11,2017). Arora, A., Candel, A., Lanford, J., LeDell, E., and Parmar, V. (Sep. 2016). "Deep Learning with H2O.", (accessed Apr 11,2017).


Download ppt "Neural Networks in RStudio with the H2O module"

Similar presentations


Ads by Google