Interview with Ian Goodfellow — GAN’s, DeepLearning Book
We’ve seen colour put into monochrome stills, silent movies being empowered with voice, and languages being translated with a single look. What paradigm-shifting innovations do you believe Deep Learning has in store for us all?
I hope that AI will help us to develop new medicinal techniques and green energy technologies.
Secondly, what do you think is underrated? These could be techniques that are not so well known or just ones that work well but aren’t popular/trendy.
Numerical stability. Very few people emphasize getting this right, but it has a large effect on the performance of deep learning algorithms. I decided to put a lot more about this in the lecture slides for the deep learning book than we were able to put in the book itself: http://www.deeplearningbook.org/slides/04_numerical.pdf
Which of the most popular deep learning libraries do you use most often? Why?
I use TensorFlow. It’s very good for large-scale problems, using multiple GPUs, reading data from disk using multiple threads, etc.
If you had to start learning machine learning again from scratch, where will you start? Would you recommend a theory-based or a project-based approach?
I would recommend reading www.deeplearningbook.org while also carrying out a project.
What is, from your perspective, a success factor when doing research?
Successful researchers are the ones who are very sceptical of their own work, making strong baselines, evaluating your new algorithm very carefully, etc. Usually, your evaluation will show you that your new idea does not work very well, but then you move on and eventually do something better. Less successful researchers spend a lot of time trying to promote ideas that don’t actually work very well.
What are some challenges a person new to Machine learning might face? How should they go about dealing with them?
Whenever you learn any new skill, it can be useful to know about the four stages of the learning cycle: https://en.wikipedia.org/wiki/Four_stages_of_competence
When you first start learning the skill, you’re very excited about it. Early on, you start to realize just how much there is that you don’t know, and you become discouraged or intimidated.
When most people get to this second stage, they think they need more information about how to do the skill better, but actually the bigger problem is motivation, not information. You need to find a way to make practising the skill be more enjoyable so that you stick with it.
A lot of the time I see newcomers to the field get discouraged when they are trying very hard to accomplish some very difficult research goal. At times like that, it can be helpful to spend some time on a more straightforward implementation project for a while so that you can feel some success and emotionally recharge before diving back into the difficult research.
Generative Adversarial Networks are picking up steam nowadays. Did you see this coming? How was your reaction to this sudden popularity?
In 2014, when my co-authors and I wrote the original paper, I actually did expect that GANs would be a big deal. By mid-2015, I had started to think maybe I was wrong, but then after DC-GANs came out it was clear that they would really take off.
Thank you so much for taking out the time to interact with us. Your unrelenting devotion to quality research is something which inspires our readers. Do you have any final words of wisdom for them?
Being truly creative requires free time to think. In machine learning research, you’re likely to get very busy. Be sure to leave some time for spontaneity.
Thank you for taking the time to read until the end. If you liked this post please leave a clap.
This interview is from November 2017 by nybles and republished after being lost for a while. The interview was conducted by Gautham Santhosh visit http://gauthamsanthosh.com/ or https://twitter.com/gauthamzzz
I am working on a customer feedback tracker visit https://www.featuremonkey.com/ which is a great alternative for canny, hellonext, uservoice which can be used for feature request tracking , internal feedback, public roadmap etc
