joraso / karn

A notebook of basic machine learning and deep neural network architectures and implementations, along with examples/experiments.

Karn

This repository is intended to be a collection of various machine learning model and deep neural network implementations, as well my experiements with them. This serves as a notebook for me as I teach myself the basics of artificial intellegence, as well as a resource for rapidly deploying similar models in the future.

Currently included tools:

Building Blocks

Compound-layer block object used to construct neural networks.

• DenseBlock: Simple block of dense layers.
• Conv2Dblock: A block consisting of a 2D convolution followed by pooling, for image-based archetectures.
• Deconv2Dblock: A de-convolution block consisting of Conv2DTranspose followed by batch normalization, for image-based archetectures (mainly autoencoders).

Network Archetectures

Sample constructions of different neural network archetechtures.

• SimpleClassifier: Simple DNN for supervised classification.
• SimpleAutoEncoder: Simple autoencoder archetecture.
• ImageClassifier: Basic structure for a convolutional image classification network.
• ImageAutoEncoder: Convolutional AutoEncoder for image processing (and other stuff, I suppose, as applicable).
• SimpleVariationalAutoEncoder: Simple archetecture for a variational autoencoder.

Notebooks / Experiments:

• SimpleClassifier_MINSTfashion: A simple network to classify images of articles of clothing from the MINST fashion data set.
• SimpleAutoEncoder_MINSTfashion: A test of autoencoder archetecture on the MINST fashion data set, testing it's reducibility to 2 dimensions.
• ImageClassifier_CIFAR10: A first pass at implementing/testing the ImageClassifier object using the CIFAR 10 data set.
• ImageAutoEncoder_MINSTfashion: A test of the convolutional autoencoder archetecture on the MINST fashion data set, evaluating it's quirks and loosely comparing it's performance to the SimpleAutoEncoder.
• ImageAutoEncoder_Denoising: Experiment on the use of a convolutional autoencoder archetecture to perform image denoising, cleaning Guassian noise from the MINST handwritten digits dataset.

Planned subjects to cover are:

• Training regimes and learning schedules
• Testing Bias vs. Variance
• Strategies for improving robustness
• Graph Convolutional networks
• Support vector machines, clustering algorithms and other classic ML methods
• Unsupervised / Deep Learning models
• Adversarial networks