The APL notation would appear to be a clear match for convolutional neural networks, but traditional implementations of APL have lagged behind the performance of highly tuned, specialized frameworks designed to execute CNNs on the GPU. Moreover, most demonstrations of APL for neural networking have involved relatively small examples. We explore a more complex example in the U-net architecture and utilize a modern APL compiler with GPU support, Co-dfns, to compare the state of the art of APL against the current crop of specialized neural network frameworks in the form of PyTorch. We compare performance as well as the language design of APL for neural network programming and the clarity and transparency of the resulting code.

We found that the complete “from scratch” APL source was on par with the complexity of the PyTorch reference implementation, albeit more foreign, while being more concise and complete. We also found that when compiled with Co-dfns, despite the naïve implementation both of Co-dfns and our own code, performance on the GPU and the CPU were within a factor of 2.2 - 2.4 times that of the PyTorch implementation. We believe this suggests significant avenues of future exploration for machine learning language design, pedagogy, and implementation, both inside and outside of the APL community.