Biases

Random thoughts during the internship in Korea, 2018.

I believe that each and every individual does this at some level, trying to get the bigger picture behind the scenes and finding out how things work. I am in Korea for an internship, the best thing about coming to a place with a completely different culture is that you realise that no matter how logical we claim ourselves to be, a majority of our daily chores are decided by where we were born and what sort of a place we grew up in. “The set you train on will never tell you how badly you overfitted.” I’ll keep on invoking silly analogies, I have been thinking about a lot of them lately. I can go on talking about analogies and how(I think) they work, but I don’t want to keep on extending new threads to confuse the reader here. Lately I was reading one of my posts from a few months back, it took me time to return back to the state in which I had written it and to actually understand the meaning behind it. I realised that it is very important to give enough context to your reader. Context- “I have been working at Language Understanding Lab(read ‘Machine Learning’ if you don’t get it) at Samsung Research for the past two months.” So coming back to analogies, I love to go on into analogies and see upto how far do they hold. I believe that the higher up a person is on spirituality, more balanced are the analogies he gives. This stems from the belief that learned men think a lot about what they speak. I’ll leave this to the reader to think upon. I started my internship by reading a paper by Kyunghun Cho, the man who introduced attention mechanism into the encoder-decoder framework in (Neural) Machine Translation. A few days later I was attending a talk by him, where he talked about an all-language translator. He talked about the universal encoding step, a representation which can be decoded further to any language. A common or universal representation is something which the Indian media would refer to as ‘The Language of God’, if it ever catches hold of it. Basically you can convert any language ‘A’ into the ‘Language of God’ and convert it back to any language ‘B’ . This would prove to be revolutionary for resource deficient languages. You don’t need tonnes of data for all of them, for it suffices to fine tune the model to these languages at later stages. The model can be trained on data-rich languages and then fine tuned to others(which requires much less data). Coming back to cultural differences, I wanted to get over the bias of the culture I grew up with, it would include understanding and accepting all the cultures and practices around the world, pretty much same as the translator would understand all the languages. I wanted to use this statistically different environment to retrain myself. The fact that I was a Vegetarian (largely) and knew very little of hangul(I could only read it) restricted the amount of data I could use, making Korea a resource deficient culture for me. Thanks to Cho, I hoped I could just fine tune to this place and reach the ‘Culture of God.’ (Extending the analogy) I believed that this ‘Culture of God’ could help me get closer to the ultimate truth. There are gazillion ways to reach it, cultural differences and human behaviour being just two of them. Bye Bye!