Can you guide a teenager?


(Abhijith Chandran) #1

Hi fellow students and teachers,
My name is Abhijith and am a student from India. I turned 15 a couple of months back.
Background:
I started by completing the entire Python specialization on Coursera by University of Michigan. The I began on Machine Learning by Andrew Ng and deeplearning.ai by Andrew Ng. I was finding it very difficult for finding practical implications of it and to keep up with the theory intensive and highly mathematics oriented course.
Fast.ai has truly helped me a lot, and am shortly going to start my second run on the first course and have also convinced my parents to let me pay for the GPU(I come from a slightly disadvantaged background).

Questions:

  1. After doing both of the deep learning courses can I participate and excel in Kaggle competitions?
    2)Will a GPU+ on Paperspace suffice the need for this curse and also for similar projects in the future?

I would also like some General advice by the highly experienced people at fast.ai.
Also I have some ideas for the implications of this and would like to talk about it to anyone who is ready for a great chat on anything machine learning.
Thank you,
Abhijith


(Ramesh Sampath) #2

Hi Abhijith,
You are doing great and the Part 1 and forums should help you with your goals. The more time you put in trying out different things, the more you will get out of the course.

I wouldn’t let Kaggle determine whether you are successful in using them. I have not won anything in Kaggle and that doesn’t bother me. Kaggle is very competitive and some folks in the competitions may not care how much time a model trains just to get little better results. So, I would suggest you take a look at problems in Kaggle playground like Plant Species (https://www.kaggle.com/c/plant-seedlings-classification) where you can apply what you learn and see the results.

Jeremy will walk through or provide suggestions on possible options. Last year, some of used Crestle for a while. So I would suggest to wait till the course starts before you spend time setting up a paperspace machine or anything else to get ready for the course. Always double check to make sure any cloud machine you start is turned OFF when you are done.

The fastai forum is generally a safe place to have the conversations on ML and will get very active after the start of the course. That way lots of people can contribute and also learn from the conversations that happen :slight_smile:


(Abhijith Chandran) #3

Thanks Ramesh Sir,
Your response is deeply appreciated:)
Also, by not doing the theory part of it am I losing out on a lot(I plan on taking the theory and maths part of it once I am in a University).

So with the recent fake news drama is there a way to use NLP for identifying this? Can someone link any scholarly articles on Data Augmentation in Natural Language Processing?
Thanks,


(Maria) #4

Do you know there will be international fellowship for the pt1 soon?
For the math you can look stuff on khan academy, it will have you covered.

You can use AI to generate fake news and to detect them too: https://towardsdatascience.com/i-trained-fake-news-detection-ai-with-95-accuracy-and-almost-went-crazy-d10589aa57c


(Abhijith Chandran) #5

Hi Blanche
Sorry, but I did not fully understand your first point, would be very happy if you explained:)

Also, isn’t 95% slightly bad? If supposdly everyday 100,000 articles are uploaded on the internet then won’t it classify 5,000 articles wrongly. Is there any other way?

Thank you for your time,
Abhijith


(Maria) #6
  1. Check out this thread: Fast.ai Live - the new version of the international fellowship
  2. Don’t really think 95% accuracy can be labeled as bad, there is always space to improve, but 95% seems decent imo, maybe if you learn some deep learning you can do a better model.

(Abhijith Chandran) #7

Thanks Blanche


#8

In a ‘real world’ setting, 95% can be quite valuable. Say that the 100,000 articles in your example are split 50/50 real vs nonsense. It certainly seems that way sometimes. If the 95% error is evenly split, a filter would result in 47,500 accurate articles to 2,500 false articles, a big improvement over 50,000 real and 50,000 nonsense.

On the other hand, if only 10% of articles are false, then 95% accuracy would block a lot more accurate articles than false ones, in which case the value of the model is reduced. Perhaps you can tune the model so you only block the articles where you get the strongest nonsense score. The 95% accuracy would drop, but the purity of the rejects would improve.

It depends on the goal. I work in fraud. For my first fraud model, the company had a lot of fraud and wanted to stop all of it. They figured the good customers would try again if blocked, so my model was less than 95%, with the errors on the false positive side. Where I currently work, fraud is a tiny percentage of millions of transactions, so the goal is to stop as much as possible while trying not to annoy the customers doing legit transactions.


(Abhijith Chandran) #9

Thanks Ralph for shedding light into an entire new perspective. Any suggestion for deep learning projects?


#10

I’d say try to find something that relates to a passion of yours, or is a need in your local world. Find something that you can be the expert in.