Applied AI 2019

When the participants of our popular DeepLearning.ai Study Group asked for a succeeding program where they could gain hands-on coding experience, we were quick to take applications for our Applied AI Study Group, in which we ran discussions by following IBM’s Applied AI course.

This study group proved to be as community-driven as it gets since it was led by one of our previous DeepLearning.ai Study Group and AI Projects participants: Ahmet Melek.

Even though the study group itself lasted only four weeks, it was an intensive course which offered a productive journey thanks to both the nature of the course itself and the ways Melek effectively led a focused discussion aligned with the needs of the participants to have a dynamic session that met everyone’s needs.

Not only the participants coming from different backgrounds learned about Deep Learning frameworks such as Keras, TensorFlow, PyTorch, DeepLearning4J and Apache SystemML; they also had the opportunity to use them for building various models. These models were used to experiment in solving real-life problems such as anomaly detection in sensor data and stock market forecasting.

In addition to the ones mentioned above, the participants learned about the data storage solutions, real-time data processing and distributed computing methods to deploy and scale their projects under the guidance of Melek.

Other than the content offered by the course, the participants completed a final assignment prepared by our guide, designed in a way that covers the topics of the study group. The group completed a total of eight mini projects and one assignment.

You can find the study group’s Github page here.

In addition to the aforementioned, you can see what we had been up to week by week below.

First Week:

We have worked on three problems:

• Image Classification with MNIST dataset on TensorFlow, using Fully-Connected Neural Networks.

• Image Classification with MNIST dataset on Keras, using Convolutional Neural Networks.

• Image Generation with aligned UTK dataset on Keras, using Deep Convolutional Generative Adversarial Networks (DCGAN Autoencoder).
  

We have worked on Google Colab for each example. 

Second Week:

We have worked on three problems:

• Anomaly Detection with Bearing Data Center Seeded Fault Test dataset on Keras, using LSTM autoencoders.

• Timeseries future prediction with Federal Reserve Economic Data Crude Oil Prices Chart on Keras, using LSTM Networks.

• NLP Embedding and Classification with a custom mini-dataset on Keras, using perceptrons, Fully-Connected Neural Networks, and Embedding layers.

For all examples of the Second Week, we have worked on Google Colab.

Third Week:

We have worked on two problems:

• Converting Keras models to DL4J models. Converting DL4J models to Apache Spark models via SystemML. After that, making classification with Iris dataset on Apache Spark using Fully-Connected Neural Networks. For this example, we worked on IBM Watson studio.

• Converting Keras models directly to Apache Spark models via SystemML.

• After concluding the above stated we worked on making image generation with an aligned UTK dataset on Keras, using multilayer perceptrons. 

Though we worked on Google Colab for this example also, we have failed due to the complications with setting up Google Colab's environment.

Fourth Week:

We have worked on the solution of the project assignment which we had assigned in Week2.
The problem was detecting joint coordinates of a hand (knuckle coordinates), using images taken with webcams. We have assumed that the coordinates of the hand itself are already detected, and we have tried to predict the knuckle coordinates.

We trained with the "Large-scale Multiview 3D Hand Pose Dataset" by Rovit.

We then used Convolutional Neural Networks on Keras.