News for hackers. Stuff that matters.
Re:NumPy another Iverson Ghost (2018) by imglorp on Tuesday August 18, 2020 @09:03AM (#24198067)
I'm curious why APL hasn't met all the AI/ML work with Tensors. It seems like a well suited marriage at first glance.
Re:NumPy another Iverson Ghost (2018) by jpf0 on Tuesday August 18, 2020 @11:32AM (#24199903)
I lead a company that has a large portfolio of AI/ML/optimization algorithms in J (Monument). Happy to answer questions.
Re:NumPy another Iverson Ghost (2018) by tosh on Tuesday August 18, 2020 @12:30PM (#24200725)
does J have support for using GPUs? or can you share a bit about how you use J? I guess there are also some ways where GPU support is not as important (different architectures or leaning more on transfer learning?).

can you share a bit about how you came to use J?

Re:NumPy another Iverson Ghost (2018) by jpf0 on Tuesday August 18, 2020 @01:57PM (#24201908)
We use J in all ML and data handling functions, everything except UI and viz.

We primarily focus on algorithms that require rapid sequential iteration rather than parallelism. However: 1) we run an enhanced version of J that is implicitly parallel on CPUs, and 2) it is possible to call GPUs (both CUDA and openCL) through J (e.g. j-fire). We do not call GPUs or other accelerators currently, though we may do so in the future.

Our goal with Monument is to create an Excel-like substitute for ML / AI. In comparing across languages, we selected J as "the right level of abstraction" for ML / AI because we wanted to write fast-running code, quickly.

J runs on virtually all CPUs and OSs, and it is highly tuned for minimized cache misses and mis-predicted branches. It is well-documented, well-supported, and the community is highly supportive.

Like Python, J makes it possible to run highly-optimized code without a great deal of boilerplate in an interactive development environment. Unlike Python, J is composed of reusable functions that readily express math. Our dev productivity is extremely high, and often we're able to engage with mathematicians who have little or no programming background.

Our team comprises polyglot programmers, and I strongly avoid "the language wars," but overall we have found J to be well-suited to our needs.

Re:NumPy another Iverson Ghost (2018) by skruger on Tuesday August 18, 2020 @09:19AM (#24198193)
Several papers available on using APL for ML, e.g. https://dl.acm.org/doi/pdf/10.1145/3315454.3329960
Re:NumPy another Iverson Ghost (2018) by vmchale on Tuesday August 18, 2020 @09:53AM (#24198541)
So far the GPU backends are not all that.