Learning by Teaching

I had heard before that the best way to learn anything is to try to teach it to others. If you can explain a topic of interest to different types of audience, from very detail to a high level concept, then you have understood the topic well. Recently I have experienced it so many times that I really wanted to share the value of this unique approach. I also made note of certain steps during the process, if followed properly, can lead to the best outcome for you and others in the company.

1. Choose a topic: Depending on your interests and skills you want to build keeping track of recent developments is a first step. Before that new methodology or tool become a common knowledge there is a small window of time that you can self-learn and practice it to a level where you know better than your peers.

2. Create your audience: People will only spend time listening to you if they find value in those technologies. In your workplace, identify opportunities where those new methods/tools can be applied to improve the performance. Its good to have a mapping or outline in your mind of how this new technique can be applied to their data, before you talk to them.  Offer to give a company wide demo or a lecture. There may be many online lectures available but they are very general. Since you know more about the problems/data your company is interested in, you can act as a bridge between those concepts and the possible solutions.  

3. Plan properly and execute smoothly: Asses your audience level and build accordingly. Depending on their background they may need different path to the concept. It is easier if you think of some concrete example from the domain they are working, they will understand fast, and you will practice applying the same knowledge to different domains.  

4. Deploy in the new domain/application: If you have time, work with some other team to apply those tools and asses the performance on their data. You will know about new failure cases or limitations when you start applying on very different kind of data than the initial publication/tech report showed.  

At PARC we are trying to build a competency around deep learning. A bunch of us started looking into the different libraries and toolbox available online. I investigated Caffe a deep learning toolbox [1], and gave a center wide tutorial/lecture for other groups. Now, not only I see deep learning being used by several groups at PARC, people are going beyond basic CNN (convolutional neural network) and trying combinations of various other techniques (LSTMs, bidirectional-RNNs and so on). Its really exciting to see that it has become one of the first things to try for any video and image analytics work.        

[1] Caffe: https://caffe.berkeleyvision.org/