@joelkuiper I see that architecture could work, but I disagree that Clojure is too slow for this stuff. We have people doing GPU based work in Clojure, for example. And core.matrix vectorz-clj is competitive with C / Java CPU based libraries even without GPU support, which I have found in practice to be good enough for most ML tasks

Would love to see more people helping develop the Clojure ecosystem for data science, rather than adding system complexity by delegating out to other environments / systems

@mikera: true, it’s not just speed though. It’s also that the eco system is much less developed, like you say. Libraries like scikit-learn really make a huge difference, and porting that to Clojure (including all the nice documentation) would be a monumental task. Personally I’ve always liked a polyglot approach (between and with any language) I find it actually reduces complexity, since it saves you from rewriting/reimplementing often well established and peer reviewed libraries

Bayesian Analytics stuff for example, I’d rather use one of the wrappers for JAGS from Python or R, than to try (and likely fail) to reimplement 15 years of MCMC optimization in Clojure or Java. But I guess it depends on the task at hand and the constraints in which you have to work

it’s a shame the JVM-native interop story is such a pain still, I often find myself tempted to wrap native libraries, but then end up doing that from a non-JVM language or via RPC

since it’s such a hassle to build the shared libs and make it do anything that is not extremely frail