2018/1/11 Tech feeds highlight:
From KDNuggets:
tensorflow customization: Introduce PowerSign/AddSign as better choice of optimizer than Adam when using on recurrent neural network and show how to customize a tensorflow optimizer:
basic optimizer consisting of two major methods: compute_gradients and apply_gradients
customization is done with apply_gradients method by adding a new Python class with the following methods (very similar to extending Theano):
_create_slots(): create memory slots for variables, taking a var_list as input
_apply_dense(): define how to update variables
_apply_sparse(): same as _apply_dense() but a sparse version
_prepare(): convert python variables to tensor finally using Rosenbrock function to test on optimizer’s performance
lasso & ridge explanation: Have a nice picture illustrating the difference of lasso & ridge
docker primer: Provides the reasons for using docker: reproductivity, portability and deployment (can release as a data App). Its content covers the basic use of docker from installation to building own docker image (mainly use ubuntu platform). Another docker tutorial
From Google research:
introduce the image compression challenge in CVPR 2018: compressed images works for mobile-scale training but lose information during compression. To maintain a best quality in the compressed images requires machine-learning skill (mainly a neural network).
jobs & other non-technical topcis:
study on data scientist profiles: doesn't have that much insight
what’s like being a data scientist: Part 1, Part 2, Part 3
TED talk about AI (most of them are social ethical issues)
Next action: reding "Neural Optimizer Search with Reinforcement Learning”
