Blending Research and Development
There is pure research and there is applied research. Pure research seeks to learn new things for the sole purpose of learning. Applied research seeks to learn new things for the purpose of achieving a practical objective, such as enabling a capability or feature in a product.
As an example of applied research, SpaceX’s highly advanced Raptor rocket engine contains a “pre-burner” that contains an oxygen-rich fuel mixture. It operates at very high pressure and temperature, and SpaceX had to invent a new metal alloy that could withstand those conditions. The metallurgical research was done for the practical purpose of enabling the engine to operate at a high pressure and temperature, which increased its efficiency.
Pure research is almost always decoupled from product development. In fact, few companies do pure research. The ones that do are usually very large and well funded. IBM and Bell Labs were well known for their pure research during the 1960s through the 1980s. Pure research is less common within companies today, and is usually done in collaboration with universities in order to access government funding to supplement the cost. The reason is that it is not possible to write a business case for pure research, because—by definition—pure research has no expected outcome. Companies like IBM justify it by having a very large research portfolio and very broad and deep markets, so that they can show a business case based on the expected outcome at a research portfolio level.
Applied research is often done as a separate step, disconnected from product development. For example, an applied research project might produce an improved method for doing something; a separate follow-on product initiative might seek to use that new method in a product.
A different approach is to perform the applied research as needed, in the course of product development. In the SpaceX example above, the new alloy was developed in the course of engine development, rather than as a separate precursor effort. The advantage of this is that there is no delay in incorporating the applied research result into the product. In fact, SpaceX is known to discover new technologies and have them fully incorporated into their product with weeks.
Machine Learning Model Development is Actually Applied Research
Machine learning development falls into two categories: (1) commodity models from provides like Azure and AWS, and (2) custom models, created by machine learning experts. All machine learning development is sometimes viewed as development, but the development of a custom machine learning model is actually applied research. Machine learning experts are mathematicians, and when they develop a machine learning model, they are experimenting all of the time. They are also theorizing. And they cannot generally tell you how long it will take to develop a successful model, especially if it is using datasets that they have not worked with before.
While they work, they produce refined models. The latest model is immediately introduced into the product that is being developed in parallel. And from time to time they might throw out a model and start over.
ll research is exploratory. There is theory, model building, and experimentation. These are not like development: there is no process. And different researchers work differently: do not try to make the all work the same. There is a well-known dichotomy between theorists and experimentalists, and we need both.
Creating a flow between applied research and development:
Have one or a few people who have a good understanding of both, and use them as a liaison between applied research teams and development teams.
Invest in training the development teams about the research and how it is done, and vice-versa: establish discussion between the two groups.
Don’t make promises for the research teams: they can never promise.
Always have fall-back plans for if research does not produce results when expected.
Share development challenges with researchers when there is a chance that research might produce a solution.