Hacker News new | past | comments | ask | show | jobs | submit | indweller's comments login

   Location: Los Angeles, CA
   Remote: Office/Hybrid/Remote
   Willing to relocate: Yes, within CA - prefer Bay Area
   Technologies: Reinforcement Learning, Robotics - Python, Pytorch, Isaac Lab, MuJoCo, C++, ROS, HTML/CSS/JS
   Résumé/CV: https://indweller.github.io/assets/Prashanth-Ravichandar-Resume.pdf
   Email: rprash99[at]gmail[dot]com
   Website: https://indweller.github.io/
----------------------------

I hold an MS in CS from the USC. Before joining USC, I worked as a Software Developer at Morgan Stanley for two years. I graduated from the Indian Institute of Technology (IIT) Guwahati with a B.Tech in Engineering Physics and a minor in Computer Science and Engineering.

I’ve worked with humanoid locomotion, deep learning, and RL for robotics - paper submitted to ICRA on dynamic loco-manipulation (like playing soccer) - https://indweller.github.io/ogmplm

I've work authorization and am looking for a 6-month internship from January-June 2025.


What should be added/removed in this post given that it's been 5 years?


This is amazing! Can you extend this to a) allow choosing regions and b) provide some weekly digest kind of thing where the news items from the past week are summarized?

Also, what's your source of info btw?


I mention it in the about page, it’s based on a recent Indian news shorts app called Beatroot (“be true”).

I haven’t looked at other regions, but there’s already some good plain text news websites for US/UK - https://greycoder.com/a-list-of-text-only-new-sites/

I don’t wanna add a summary, or in general host older news - the more news there is to read, the more time you’ll spend. I don’t want that. It’s like a newspaper, if something important enough happened this week that you missed - the news will still be reporting on it.


Are all the assignments self-graded or do we have to be enrolled at MIT to submit somewhere and see if our code is correct?


Looks like the labs include a makefile with a "grade" target so you can run "make grade" and it will tell you if you pass.


Where do you keep the rest of the things that you need to do at some time? Like do you mainatin some other list of all the things that need/want to do, but not necessarily do today?


Where do you keep the rest of the things that you need to do at some time? Like do you mainatin some other list of all the things that need/want to do, but not necessarily do today?


There's always the team's backlog, but that's not something I personally committed to finish.

Realistically, there's no point in having a very long list of items you will never go through.


Can someone explain why trie is not faster than hash table?


What advantage would a trie's prefix searching give you in a word count task? Worse, you're following word.length() pointers, potentially all over your memory. You could collapse with Patricia or Merkle or whatever, but they'll all be worse than the one pointer lookup of a hash table (modulo collision chasing if you pick a bad hash/size) & all in the service of giving you a feature (prefix lookup) you don't need.


The source code of the trie program I wrote is here: https://github.com/benhoyt/countwords/blob/master/rust/optim...

The comments explain why I think it's slower. The TL;DR is that using a trie requires more memory accesses than a hash table (per byte of input), and those memory accesses slow the whole enterprise down.

But, not all memory accesses are created equal. So perhaps a more clever representation that exploits locality better would do better. "Obvious" techniques for shrinking memory usage made a big difference and brought the performance of the trie program closer to C: https://github.com/benhoyt/countwords/pull/2


Yeah, I get it now. Thanks!


Focus on the means rather than the end. This takes off the pressure and helps you work efficiently so that you achieve your goal, instead of just dreaming about it. It also helps overcome procrastination.


Your blog is a great resource. Especially the fact that you've linked other blogs and links is really helpful.


In the blog, he refers to test losses at an early stage, like in "add significant digits to your eval". Does he actually refer to the test data or is he referring to validation data? I was under the idea that we were supposed to touch the test data only once at the end of all training and validation. What is the right way to handle the test data?


By "eval" you can also mean the training subset. As I understood is at the code to evaluate the network at a given point with a given dataset. For instance, after epoch epoch, the model is evaluated for both training and validation (you see both losses)

As you said, the test subset should only be used at the very last.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: