3 ms·
Right now I'm mostly focused on building NeuralObjects, an open source "Machine Learning As A Service" platform, and the associated hosted service offering to g
by mindcrime 10y ago
Right now I'm mostly focused on building NeuralObjects, an open source "Machine Learning As A Service" platform, and the associated hosted service offering to go with it.
The idea is to simplify the provisioning of compute clusters and the installation and configuration of various F/OSS platforms and toolkits for Machine Learning and Artificial Intelligence, and then provide convenient APIs for common operations. For the first release (coming Real Soon Now, really) we'll have engines based on DL4J, probably SystemML, and a very simple engine based on Apache Commons Math (for doing simple stuff on a small scale). We also bundle Apache Zeppelin as part of the stack, for interactive analysis in a workbook format.
We'll have a REST API for operations like - defining models, training models, working with model versions, and making predictions with models. We'll probably also offer an RPC based API for predictions, likely based on gRPC.
Long-term, the vision is to serve as a one-stop-shop for all of the popular (and maybe some not-so-popular) F/OSS tools in the ML/AI space. Some things are are on the roadmap include Tensorflow, Caffe, Keras, Mxnet, SK-learn, and MLLib.
The first cut is using clusters based on Hadoop/Spark, but we'll probably also add the ability to provision MPI clusters. We'll also be exploring how to possibly take advantage of the Amazon FPGA instances down the line.
And since the whole thing will be OSS, anybody who wishes to install it locally will be able to do that as well. To that end, we plan to eventually add support for provisioning with OpenStack. We'll also add support for other cloud environments besides AWS at some point.