[4eyes] Summary of ML discussion meeting

[Peter Zhe Fu]

Hi all,

     Just would like to summarize a few links we mentioned during our ML
discussion meeting last Friday. Feel free to add links that I missed :-D

     1) at the beginning of our meeting, we talked about that often we have
to wait 10+ hours before getting feedback on whether our change to
algorithms work. While we discussed about various tips, earlier this week,
Andrej Karparhy posted 'A recipe for training neural networks' on his blog.
He shared many useful tips for training neural networks. It can be found
here: http://karpathy.github.io/2019/04/25/recipe/

     2) the theme of our meeting was Jedrzej's idea on random networks.
Earlier this month, Kaimimng He posted on ArXiv 'Exploring Randomly Wired
Neural Networks for Image Recognition'. The authors obtained competitive
accuracy on ImageNet Benchmark with randomly wired graph networks. It can
be found here: https://arxiv.org/abs/1904.01569

     Also, maybe some of you could not make it due to time conflicts. So
far we have been holding this ML discussion meeting on Friday 10 am.
Usually one of us give a 5 min introduction to our research problem and
challenges. Then all of us would pitch in to brainstorm. It has been
tremendously helpful in my point view. So you are welcome to join or let me
know if you prefer a better timing so we can conduct a Doodle poll.

    Have a great Sunday ahead :-D

Thanks and best regards,
Zhe Fu (Peter)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cs.ucsb.edu/pipermail/ilab-users/attachments/20190428/2b02ec83/attachment.html>