This section contains labs that consist of a notebook and corresponding presentation ("content bundle") for community participation in machine learning research. The target audience for these hands-on labs are ML researchers, ML engineers, data scientists and those who are already familiar with deep learning.
The Community Lab - Intro to Composable (Common).pptx presentation is common across all the content bundles.
A deep understanding of AutoML, a new approach based on new design patterns, latest research -- turning AutoML from a blackbox to one where the data scientist can bring their own custom macro/micro architectures and self-guide the search space.
| Lab | Description |
|---|---|
AutoEncoder for CNN |
Research using lower dimensional encoded inputs to CNN (vs original image) |
Regularization |
Research using regularization to counter overfitting |
Ensemble |
Research using intra-model ensemnble methods |
Objective
To replace a traditional "stem convolution group" of higher input dimensionality with lower dimensionality encoding, learned from first training the dataset on an autoencoder. Goal is that by using a lower dimensionality encoding, one can substantially increase training time of a model.
Question: Can one achieve the same accuracy as using the original input image?
Question: How fast can we speed up training?
Objective
To explore methods of regularization and learning rates to prevent the training data from "fitting" to the weights in a compact model -- without use of historical methods such as dropout or data augmentation.
Question: Can we generalize a compact model without image augmentation?
Question: How is training time effected?
Question: How small can a compact model be made and maintain accuracy on the validation/test data?
Objective
To replace a traditional "inter-model" ensemble of models of high complexity with an "intra-model" ensemble of lower complexity, while retaining the performance benefits.
Question: Can one achieve the same performance with intra-model bagging vs. traditional inter-model ensemble?
Question: Can one achieve the same performance with intra-model stacking vs. traditional inter-model ensemble?