After failing to learn deep learning three times, I’ve finally found the right approach. I’ve started FastAI’s Practical Deep Learning for Coders course for the second time. I’m through lesson one and feel more confident after reading Meta Learning: How To Learn Deep Learning And Thrive In The Digital World by Radek Osmulski.
I’m taking my approach from his book. Watch the lecture, recreate it in a Jupyter notebook, then build something similar with another dataset. I’ve wasted too much time before attempting to create or acquire the right data. Fortunately, FastAI recommends many external datasets.
I don’t just want to build; I want to learn the fundamentals. This includes debugging and making the code fast and efficient. Based on this article, I can use %debug to get an interactive debugger where the exception occurred, allowing me to look around! The %%timeit debugger also seems useful when creating datasets and data loaders. Every model I train should have at least 95% accuracy.
My goal is to recreate at least one model per week and write about any models I recreate by the end of the week. After, I’ll try to use the drivetrain approach to create my project, but I may just recreate with a new dataset. I’ll be using the Smart Brevity format in my writing while likely structuring using the Minto pyramid.
I’m writing to benefit my future self, and maybe others. Also, I hope to find a mentor and others near me in Atlanta (or online) interested in learning together. These are the benefits of learning in public.
If you have similar interests or just want to chat, message me on X/Twitter or LinkedIn:
LinkedIn: https://www.linkedin.com/in/%F0%9F%93%88-nick-swekosky-4090155b/
X/Twitter: https://x.com/nickswekosky