Web developers don’t need a math degree to get started with machine learning
What are notable entry hurdles for software and web developers who want to get started with machine learning? Google AI researchers Carrie J. Cai, Senior Research Scientist, Google Research and Philip J. Guo, Assistant Professor, UC San Diego, decided to find out. They analyzed the results of a survey among 645 TensorFlow.js users and published their results in a research paper as well as on the Google AI Blog.
In the TensorFlow.js survey, most participants were software or web developers, and were not using ML as part of their primary job. Among the motivations for learning ML was that they found the idea intellectually fascinating (26%) or believed that ML is the future (17%). A specific job task or use case was not reported very frequently, as only 11% of respondents gave this answer.
Let’s take a look at the respondents’ challenges and expectations in machine learning.
No need for imposter syndrome
According to the survey, developers struggle most with their “own lack of conceptual understanding of ML.” This shows in the initial stages of choosing when to apply ML as well as in creating the architecture of a neural net and carrying out the model training.
The respondents often felt they faced these challenges due to lack of experience in advanced mathematics, leading to imposter syndrome. As the Google AI researchers point out, though, this isn’t the case. While an advanced math degree is useful, it is not necessary—despite mathematical terminology in API documentation that may suggest otherwise.
Expectations and wishes
The survey also aimed to find out what software developers expect from ML frameworks and what features they would like to see implemented. It turned out that the respondents were often interested in pre-made ML models that should be customizable for a specific use case, and thus provide explicit support for modification.
Tips for best practices were also high up on the list of common wishes, and just-in-time strategic pointers such as diagnostic checks could improve the developer experience as well. Some respondents voiced their desire for learning-by-doing tutorials in ML frameworks.
As the study concludes:
Software developers are now treating modern ML frame-works as lightweight vehicles for learning and tinkering. Our work provides evidence that, even with the existence of such APIs, developers still face substantial hurdles due to a perceived lack of conceptual and mathematical understanding. In the future, ML frameworks could help by de-mystifying theoretical concepts and synthesizing ML best practices into just-in-time, practical tips.
See the Google AI Blog for more information.