The Data Science Nigeria Bootcamp/Hackathon is made up of the top 150 users from both the Deep Learning and Machine Learning streams of the qualifying Kaggle competitions organized by Data Science Nigeria.
To make it into the Deep Learning stream of this Bootcamp, I came up with a recommender system that used a Neural Network to predict comedy ratings for each user based on the user’s rating for other comedy shows and it was developed with Keras.
Although the plan of the organizers was to incorporate some technical sessions on Day #1 of the Bootcamp, it didn’t happen due to some issues beyond their control. So, I spent the first day of the Bootcamp meeting and interacting with different AI and Machine Learning enthusiasts from different fields both in the industry and academia.
After a very short round of sleep, it was time to go out and explore the unique opportunities and challenges that were about to be unfolded.
Day #2 of the Bootcamp started with a session from Dr. Jacques Ludik, the CEO of Cortex Logic, South Africa who introduced the concepts and theories of deep learning. In his explanations, he showed how deep learning had gained a lot of popularity due to the exponential growth in computing power as well as the availability of tones of data. Furthermore, He introduced some of the popular deep learning algorithms such as Convolutional Neural Networks(CNNs), Recurrent Neural Network(RNNs), and Reinforcement Learning(RL) and their applications in the real world. Ultimately, It became clear that CNNs are revolutionalizing the field of Computer Vision and the reward/penalty system of RL keeps unlocking opportunities that were previously unimaginable.
After a short one hour break, It was time to learn the roles deep learning plays in financial inclusion. This session was handled by Matt Grasser, the Director, Inclusive Fintech, Bankable Frontiers Associate, USA. It was really an eye-opening session as we literally had to go through the entire chain of business thoughts required to ship a production-ready machine learning model. These thoughts center around asking the right questions, taking the right actions, making the necessary assumptions and hypothesis, understanding the process and necessary tools, creating the right model and getting the expected results.
The talk by Matt was immediately followed by a video call from Prof. Thomas G. Dietterich, an Emeritus Professor of computer science at Oregon State University. He introduced us to the concept of Anomaly Detection and showed us how they can be applied to Fraud Detection and Cyber Security. Also, I learned that Anomaly Detection can be difficult and one of the reasons for this is that their distributions are usually very similar to those of a normal distribution. One very good way of detecting anomaly is to use the isolation forest method. The overall efficiency of detecting true anomalies can greatly be improved if an analyst feedback is incorporated.
At this point, with so much to chew already and two more sessions left, I have almost begun to overfit. However, upon remembering the contents of the last two sessions, I felt a sense of regularization and regained some energy for the last lap.
The next session was on Convolutional Neural Networks(CNNs). The trainer for this session was Dr. Emmanuel Doro, Principal Data Scientist, Jet.com, USA. He started by building the underlying concepts such as Backpropagation, Gradient Descent, Artificial Neural Networks(ANN), Activation Functions, Loss Functions etc. He further went on to introduce the popular MNIST dataset which was then used as the dataset for a classification example in Keras. The Keras API provides a very good abstraction that gives programmers the flexibility needed to construct the NNs or CNNs with ease.
Abiodun Modupe, a Ph.D. researcher at the University of Wits, took the last session. He introduced deep learning in Natural Language Processing and identified an array of use cases where these kinds of algorithms find special use. The fact that we have a lot of languages in Nigeria makes these kinds of algorithms very useful in this part of the world.
Ultimately, after a very long and rigorous day, the sessions came to an end. Overall, the experience has been good and my expectations for the Bootcamp have so far been met.
Today is another long day that I look forward to even though I only have about 3hrs sleep to catch.
Source: Deep Learning on Medium