Thread by @jangiacomelli

Don't know where and how to start with ML?

#100daysOfMLCode #100daysOfCode

🧵This is my journey so far👇

1️⃣ I started with ML during my master's degree.

I decided to try to build a model of my guitar amplifier.

I started with this course:

udemy.com/course/machine…

2️⃣ After I finished it I started playing around with Keras and scikit-learn to build a model of my guitar amplifier.

It took 72 hours of recording, 250+ hours of training, and 6 months of experimenting to conclude - ML algorithms like NN won't produce anything useful.

3️⃣ I've learned a lot about ML from these sources:

A. L. Samuel - Some studies in machine learning using game of checkers

K. Gurney - An introduction to neural networks

J. N. Nilsson - Introduction To Machine Learning

J. Patterson in A. Gibson - Deep Learning

4️⃣ And from these ones:

T. Kam Ho - Random decision forests

T. Hastie, R. Tibshirani in J. Friedman - The Elements of Statistical Learning.
Springer

Keras documentation

scikit-learn documentation

5️⃣ We started our startup in 2017 with a project for some industrial optimization.

I've learned GA (genetic algorithms) on the fly to solve the problem.

After that, we started @typlessAPI - I used Keras to build initial data extraction models for invoices.

6️⃣ Results weren't very promising. I experimented with algorithms.

Using scikit-learn results were much better.

But we wanted more.

7️⃣ During that time I used these resources a lot:

* https://machinelearningmastery. com
* https://www.pyimagesearch. com
* https://towardsdatascience. com

8️⃣ Soon after that @galgiacomelli and me - build the first custom ML library for metadata extraction from invoices (invoice number, total amount, ...).

It was changed and redesigned numerous times.

Still not good enough.

9️⃣ About 1 year ago we built the initial release of a library for data extraction from any document - including line items/tables.

It becomes stable after 6 months of hard work.

Now I mostly deploy and monitor models.

1️⃣0️⃣ During that time we used mostly:

* scientific papers
* whiteboard & pen

1️⃣1️⃣ Takeaways:

* just start (Udemy, FreeCodeCamp, ...)
* find a problem to solve (classify feelings of your girly/boy, a recommendation system for your friend's online store, classify your comments on Reddit, ...)
* be creative
* NNs are not a good fit for all problems