Bitter Lesson Response


A couple of days ago, RL founding father Rich Sutton posted a blog post. My reading is as follows:

Although I tend to agree with the overall idea that general methods are the way forward for AI, I feel that some nuance is in order. Specifically, I feel that task-specific methods fulfill a crucial role by showing what can be accomplished on a narrow task and thus serves like a telescope into the future. Furthermore, I believe that the role of data is not appreciated by Prof. Sutton. The availability data useful to some task may have natural limits.

Data has been a huge enabler of recent advances in AI. Prof. Sutton only mentions compute as the main driver for success in AI but I think availability of high-quality data sets may be equally or even more important. Curated data is necessary for all approaches in AI. Even unsupervised methods require curated data sets for evaluation and hyper-parameter tuning. We currently see this in GANs, which despite being purely unsupervised are still mostly used on data we can easily inspect ourselves (images and audio) or where labels are available in large quantities. Developing these methods in a systematic and reliable, e.g. scientific way requires a data set representative to some realistic problem setting. One could argue that, as AI techniques become more powerful and general, the datasets to evaluate them by can become equally general. This may be true in the long run, so I guess that from a decades-long POV the argument holds.

In the meantime, however, we want our research to be useful for society at large. In my view, the current triumphs of AI are the result of the fields' focus on practical solutions since the mid 90s. Without AI focusing on very practical applications, such as e.g. the Data Mining and Semantic Web subfields, we would have neither the datasets nor the support from public nor private funding we currently enjoy. These domain-specific approaches used e.g. manually crafted features to show how powerful data-driven computer vision techniques could be and thus served as a ‘telescope’ into the future. The main point to me is thus not about what to work on (‘general’ v.s. ‘task-specific’) but what to work on now.

In summary, my vision on the matter is a bit more humble than Prof. Sutton’s extremist stance: in determining what to work on, AI researchers should take into account the practical applicability now while considering implications to general approaches on the long run.