Anjishnu
Anjishnu
CS PhD student. Working on Fairness and Interpretability methods in NLP.

Week One

Week One

I got your backs… just, you know, from the front.

A new FPS game from Riot Games was released a few days back and the gaming communities are going gaga over Valorant. The quote above is from one of the fan-favorite agents in the game, named Jett (she is also the one in the blog card for this post). I have spent some time playing this new game and it has been decently enjoyable to say the least. To be exact this game is the combination of CS 1.6 and OverWatch, with all the agent abilities and the high movement inaccuracy with randomized spray patterns. The game is still sort of in a developing phase, I feel, because the characters and maps could definitely improve in some areas like graphics and also tactical sturdiness. But, it is fun to play nonetheless.

Now coming to the work done this week. Let’s start of with where I left last time. The UnPool layer. Consider that you have a 4x4 feature map, over which you apply a 2x2 maxpooling operation. The order of operations that will be executed during pooling in mlpack can be seen from the gif below.

Once these operations are complete, we will have a 2x2 reduced feature map and also 4 values of pooling indices. Both of these are required by the UnPooling layer to calculate its output. The steps can be visualized as below.

Currently, in mlpack there is no way to access the pooling Indices and so I opened an issue to facilitate discussion on this matter. But we haven’t reached any solid conclusions so far. I have completed a mock demonstration of the code for UnPool locally, but there is no meaning of pushing it till the issue above is resolved.

Another thing that I worked on during this week was the implementation of the Soft Margin Loss function.

The implementation of this function was pretty straightforward and writing the backward method was only a very simple differentiation away.

Forward method

Differentiation for Backward method (calculation of dL/dx)

Another feature I was working on was introducing reduction facility for Loss functions in mlpack. However, to do this I first needed to verify the loss functions first and see if they were correctly implemented. While doing that I ran into an unexpected issue. Some of the loss functions are not correctly implemented or don’t work as expected. I have opened an issue for listing all these errors in one place, while simultaneously working out the correct implementations locally. The work with these portions will be part of a major change in the loss functions folder which I hope to complete and push in the coming weeks.

Coming to music recommendations for the week, I have been listening to a mix of songs, all of which are really good. But, here is one that has really stuck to me for the last couple of days or so. Maybe its because the mood of that song resonates with the general theme of protests going on everywhere. #BLM

Slide by M.I.M.E, Drama B and BULGANG

See you next week!
XOXO