Who wins the APL Forge?

“My first experiment in this area was trap, an implementation of transformers (e.g., GPT) in APL. I was surprised by how concise the code was compared to the Python reference, even though the latter wasn’t written from scratch, instead utilising the deep learning package PyTorch. To take trap a step further, I started to implement more traditional machine learning models such as linear regression and support vector machines in APL, forming the basis of APLearn. Transformers are conceptually straightforward since they are, in terms of concrete implementation, little more than a sequence of matrix multiplications. On the other hand, machine learning algorithms are often trickier, relying on more than just basic linear algebra. For example, random forests depend heavily on trees, and Lasso regression is solved using an iterative, non-parallel approach called co-ordinate descent. This created difficulties at times because it wasn’t always obvious how to translate, say, ifs or for loops into APL, and I had to actively force myself to avoid falling back on a more standard, imperative mindset. During this time, I maintained a correspondence with Aaron Hsu, the developer behind the Co-dfns APL compiler, concerning performance improvements, and he encouraged me to submit my projects to APL Forge.

“In the future, I plan to study array programming in greater depth, with a focus on machine/deep learning. I believe APL is a viable tool for quickly and succinctly solving many computational problems, especially in research, and I hope I’ll have the opportunity to take advantage of the many benefits it offers in my work.”