Scarcity of Quality Data Sample Clauses

Scarcity of Quality Data. The third reason for ATR brittleness is a chronic data-shortage problem. While the above limitations may all be resolved with a xxxxxxxx xxxx of training and test data, in many cases these are unique to each mission scenario.299 For example, training samples need to include appropriate data not just on each target category, but also on mission- specific variations like different backgrounds and climatic conditions.300 Because of the potentially transient nature of such variables, this problem is often difficult to resolve in practice. Online learning may seem to offer a potential solution, but in the near-term the results are arguably too unpredictable for a safety-critical mission.301 293 For example, for recreational sport, lawful self-defence, or simply for cultural reasons, such as celebratory gunfire, which is not uncommon in parts of the Middle East. 294 Xxxxxxxx (n 28). 295 Xxxxxxxxx (n 250), 245. 296 Xxxxxx (n 50), 1392-93 (discussing a hypothetical counter-insurgency operation, where a confluence of factors leads to ambiguity. As soldiers approach a house where they suspect an insurgent may be hiding, child inhabitants playing with a ball kick it towards the gate and run after it. Adult male inhabitants carrying a kirpan dagger for religious reasons scream at the children to stay away from the gate, and begin running towards them. Such a scenario may be easy to interpret for a human soldier, but an ATR may perceive the rapid approach of multiple males carrying a dagger to be a sign of impending attack). 297 Ibid.; Xxxxxxxxx (n 250). See also Chapter 4. 298 Ibid.; Xxxxxxx (n 13), 146. 299 Boulanin and Verbruggen (n 19). 300 Ibid. Consistent with Xxxxxx’x kirpan scenario, Xxxxxxxxx (n 250) adds, at 244, that learning local cultural traits is also vital if an ATR is to target persons (even if only in a reactive or self-defence capacity), to avoid misconstruing innocuous behaviours. 301 Boulanin and Verbruggen (n 19). Moreover, rare and unexpected combinations, as above, are – by definition – not amenable to machine learning.