I have a decent background in Mathematics and Computer Science .I started learning AI from Andrew Ng's course from one month back. I understand logic and intuition behind everything taught but if someone asks me to write or derive mathematical formulas related to back propagation I will fail to do so. I need to complete object recognition project within 4 months. Am I on right path?

  • $\begingroup$ @ManuelRodriguez I did not get you .I mean i know multivariable calculus , linear algebra .Can you explain what you're trying to say ? $\endgroup$ – user27556 Jul 31 '19 at 7:38
  • $\begingroup$ You're on a right path. Go ahead. $\endgroup$ – Manuel Rodriguez Jul 31 '19 at 7:42
  • $\begingroup$ I think that you're on the right path. When it comes to backprop only understating the concept and how it works is enough. Most of the times when developing a model/project we are using an out of the box optimizer which does the job for us. What I would suggest you is to get familiar with the mathematical notations of CNN, and afterwards dive deeper into more complex aspects. $\endgroup$ – razvanc92 Jul 31 '19 at 8:22
  • $\begingroup$ Welcome to SE:AI! (This is unlikely to be a suitable question, but I'm leaving open pending sufficient close votes.) $\endgroup$ – DukeZhou Aug 1 '19 at 19:46

I think the key part of your question is "as a beginner". For all intents and purposes you can create a state of the art (SoTA) model in various fields with no knowledge of the mathematics what so ever.

This means you do not need to understand back-propagation, gradient descent, or even mathematically how each layer works. Respectively you could just know there exists an optimizer and that different layers generally do different things (convolutions are good at picking up local dependencies, fully connected layers are good at picking up connections among your neurons in an expensive manner when you hold no previous assumptions), etc.. Follow some common intuitions and architectures built upon in the field and your ability to model will follow (thanks to the amazing work on opensource ML frameworks -- Looking at you Google and Facebook)! But this is only a stop-gap.

A Newton quote that I'm about to butcher: "If I have seen further it's because I'm standing on the shoulders of giants". In other words he saw further because he didn't just use what people before him did, he utilized it to expand even further. So yes, I think you can finish your object detection project on time with little understanding of the math (look at the Google object detection API, it does wonders and you don't even need to know anything about ML to use it, you just need to have data. But, and this is a big but, if you ever want to extend into a realm that isn't particularly touched upon or push the envelope in any meaningful way, you will probably have to learn the math, learn the basics, learn the foundations.

  • 1
    $\begingroup$ Also, those others of us who are not Newton are sometimes in a position to see a little further because we drag ourselves to the top of a pile of dead dwarves. $\endgroup$ – DrMcCleod Aug 15 '19 at 17:19
  • $\begingroup$ @DrMcCleod calling the hoards of papers being produced in the field, "a pile of dead dwarves" is fitting. $\endgroup$ – mshlis Aug 15 '19 at 18:04
  • Not only is it 100% ok, it's the process.

You may be surprised to know that even mathematicians struggle with mathematics, both the proofs they are working on, and the proofs of their colleagues. Some thinkers are so far ahead of the curve, very few understand what they're stating until generations later.

The main thing is to keep with it.


If you want to bee engineer who work with models as black boxes it could be OK. If you want to be researcher, as the job position or for better understanding of the subject it's not OK. Backporpagation is just basic multivariate calculus. If you straggling with it things like Hessians, regularizers, stochastic processes etc. would cause even more problems. If you want to go research track it could be good idea to take some math courses and prioritize them.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy