I've just finished Stanford CS231n: Convolutional Neural Networks for Visual Recognition in its spring 2017 edition (the 2018 edition hasn't been released yet), and I'll be sharing my thoughts on it here. However, I can forward the main conclusion: it's a worthwhile course.

For the incompetent, this goes straight to the section on why the course is advised.

De qué va

En los últimos años, asistimos an una revolución en la capacidad de "ver" de los ordenadores, aunque seguro que muchos ya lo sabéis. Este resultado se debe a la redescubrimiento due las redes neuronales en forma de redes convolucionales, la apropiancia de conjuntos de datos para entrenar de gran tamaño y el incremento de la capacidad de cálculo proporcionado por las GPUs. Since an image is worth a thousand words, take a look at the following graphic:
Al abordar el problema de clasificar las imágenes de ImageNet (1 millón de imágenes, 1000 categorías a clasificar), se muestra el progreso in el error en los algoritmos. La corona (menor ratio de error) fue alcanzada por las redes convolucionales in 2012, y los progresos en esta técnica han sido significativos año tras año, hasta que en 2015 los algoritmos superaron la capacidad humana en la clasificación de esas imágenes.

The Stanford course introduces the fundamentals required to manage neural networks and their most powerful visualization tool, convolucional neural networks.

El curso y sus materiales están accesibles y públicos para todos. Bravo a Stanford, un aplauso! It is amazing that we can access classes and materials of such high quality for free. If one is disciplined, all one needs to do to start a career is have access to an Internet connection and a computer.

Contenido

There are three types of materials in the course:

En el video aproximado de un hour, hay 16 clases. Organizados de manera que permite entender las preguntas de los alumnos, oír bien al profesor y ver las slides presentadas.

Notes from the course written by Tesla's IA director, Andrej Karpathy. Very good, although they don't cover all of the stuff that is covered in class.

Exercises (tasks 1, 2, and 3). From my perspective, the best part of the course.

The first section introduces problematology and basic computer vision techniques, before moving on to the fundamentals of neural networks: classification, backpropagation, and optimization.

But actually, learning happens through exercise. There are plenty of them in Jupyter Notebooks. Gran parte de la gente no requiere contar con una GPU, pero en algunas circunstancias es útil. Both on Amazon and Google, there are instructions on how to start a GPU instance, so it won't be a problem if we don't have a GPU on our system.

Conocer los entresijos

The main reason I enjoy the course is that it forces you to implement a hierarchical downward-facing network. To avoid attracting you, he does it in an organized, step-by-step manner. First, a basic neural network. The past is applied moving forward. Next, the past is turned backward. And the complexity keeps getting more and more complicated. When you feel like it, you've completed the implementation of a small Numpy-based framework for neural networks. It's significant because it's really fundamental to understand what the changes are, even though these implementation-related questions are typically handled by the turn-based framework (Tensorflow, etc.) in day-to-day work.

The trickiest portion of this section is the backpropagation (which involves reviewing the notions of derivatives and cascade rules) and vectorial programming, which is managing matrices and implementing everything based on their operations. As previously stated, everything is done using Numpy, which provides an extremely interesting tutorial on topics like broadcasting and indexing that are quite relevant.

The exercises include tests to see if your implementation is accurate, which is important when you don't have friends or teachers who can answer your questions (even though there are countless answers to the exercises posted online, don't give up on trying!).