From Ballerina to AI Researcher: Part VI
What makes you most alive; how to build a simple RNN and my sixth week as an OpenAI scholar
You are welcome to join my three-month journey in the OpenAI scholarship program where I am developing my skills in machine learning. Every week, I share my thoughts about what I consider interesting and worth sharing with you and my progress in nailing AI.
What Makes You Most Alive?
This week, I took a sequence of impro dance classes. I love impro dancing — it lets me move as freely as I want, as my body wants. Dancing is truly healing for my soul: I open up what I’ve kept inside me. Through dancing, I turn into a state of flow, not feeling the difference between myself and my surroundings; and in that state I come up with the most interesting and elegant ideas.
More importantly, only when I dance I feel alive and free. I personally don’t believe in free will. We are already conditioned by DNA, the environment, and typical human limitations on the physical and mental level. This is one of the reasons why I’m so passionate about AI — the technology will augment us and help us overcome some of our limitations, comprehend universal truths and the world of physics that are not currently available for our understanding. Without AI, we would never be able to “see” those things.
Only when I dance I’m free from human conditioning; at least it gives me that feeling…
My Sixth Week as an OpenAI scholar
This week, I’ve built a simple RNN from scratch to get more of an understanding of basic stuff before moving to more complex architectures like LSTMs. I’m sharing below the details of the model (trained on DBpedia dataset):
Downloading libraries to work with, defining hyperparameters.
Data preprocessing.
Building a simple RNN from scratch.
Running the model.
Estimating the accuracy.
As you see, the model is overfitted. So the work with hyperparameters should be done. But also, I assume it happened because its a simple RNN and the more sophisticated architectures should be applied here.
This week, I’m going to play around with hyperparameters to compare the models’ performance and work on the implementation of word2vec and then move to building LSTM models.
If you have any questions or comments, feel free to ping me. You can learn more about me at Twitter.