The above image is a Japanese rescue robot named Robokiyu that is sometimes confused with the unphotographed DARPA EATR robot. The latter being a military vehicle that could feed itself biomass. The EATR robot was decommissioned after fear of it consuming people became popular in the media. Right now, most algorithms exist only in digital form. It is a striking example of why machine humility might become increasingly important.
In respect to the definition of machine humility, this writing defines it as a state of being unassuming or a conditional state of reserve.
The problem with an increasing amount of algorithms is appropriately handling unexpected data. Here is an example, someone is given a task to sort cat and dog photos. Instinctively an algorithm is created that sorts only cat and dog photos. In the future, data of people is mistakenly input and the algorithm falsy classifies people as either a cat or a dog. This example task is unassuming but one could replace it with a task to sort people in ways that will impact their daily lives.
The simplest solution to handle unexpected data is to fix the input. In real situations, input data is always messy. One suggested way forward is to monitor data drift. Data drift is the degradation of accuracy over time due to changes in input data. Monitoring data drift makes sense as a reactionary measure. However, a baked-in state of reserve is more compelling.
For vision AI, the next simplest solution is to use control categories. The goal is to expand the number of outputs beyond the original need. The additional output(s) increase the reliability of the desired output. Although not obvious from the start, this method should be incorporated into today's production algorithms. It's common to overlook the need for this, however, it does increase the overall accuracy by adjusting the data alone. For Text, pre-training text with a classified humility score might be the way to bake-in a form of humility. It will require pretraining with a sentiment analysis algorithm.
Independent control outputs are better than no control categories as they add an extra dimension to compare results against. However, targeted control categories can predict outlying future inputs before they happen. With the example of cats and dogs, adding a third category with sufficient junk data and sufficiently trained lower layers to understand the junk data will detect unexpected input instead of misidentifying it. The additional category in the example may contain images of solid colors, people, structures, unrelated animals, etc. Side note, I suspect independent control outputs might be useful in a large ensemble where nuance and generalization can be improved by combining sparsely related data. More testing is needed though.
One thing to note, is that control categories don't work and have the opposite effect when the input data is homogeneous and only subtle changes are being detected. Control categories seem to work best when the input data is diverse.
Combining this strategy with passive methods of reserve like managing low probability predictions is also needed. There might also be a better way to handle a state of reserve directly in the model architecture. One possible solution is to check for low observation regions using a humility layer. This concept would assign weights to data sent to the humility layer to influence the final prediction. Similar to how current attention mechanisms work in deep learning algorithms to improve output results. I think it's possible to approximate latent non-linear sparsity with Jacobian + dimensional reduction + k nearest + mean for each dimension. I need more math to be sure.
The final aspect is to determine what to do if a humility score is low. One could log the results, override the output, or return an error. In the case of potentially damaging robots, hopefully, they are not just saving logs of meals.