Thinking about the future of machine learning programming frameworks, I recently reread computer scientist Fred Brooks’ classic essay, “No Silver Bullet: Essence and Accidents of Software Engineering.” Three decades after its initial publication, it still holds important lessons for software engineers building ML tools.
Despite progress from typewriters to text editors, why is writing still hard to do? Because text editors don’t address the most difficult part: thinking through what you want to say.
Programming tools have the same limitation. I’m glad to be coding in Python rather than Fortran. But as Brooks points out, most advances in programming tools have not reduced the essential complexity of software engineering. This complexity lies in designing a program and specifying how it should solve a given problem, rather than in expressing that design in a programming language.
Deep learning is revolutionary because it reduces the essential complexity of building, say, a computer vision system. Instead of writing esoteric, multi-step software pipelines comprising feature extractors, geometric transformations, and so on, we get data and train a neural network. Deep learning hasn’t just made it easier to express a given design; it has completely changed what we design.
As we work on ML programming frameworks, we should think about how to further reduce the essential complexity of building ML systems. This involves not just specifying an NN architecture (which is indeed waaay easier to do in TensorFlow or PyTorch than C++), but also deciding what is the problem to be solved and designing all the steps from data acquisition to model training to deployment.
I don’t know what will be the key ideas for reducing this essential complexity, but I suspect they will include software reuse, ML model reuse (such as libraries of pretrained models) and tools not just for code versioning and reuse (like github) but also for data versioning and reuse. Breakthroughs in unsupervised and other forms of learning could also play a huge role.
Even as I occasionally struggle to get an ML system to work (it’s not easy for me either), I am excited to see how our community is pioneering this discipline.
P.S. My best learning creation so far, seven month-old Nova, just said her first words! 🙂