4 ms·

Wonderful analogy. I'm keeping it. That said, this is not totally unlike many human explanations. In fact, increasingly, if a decision can be explained entirel

Wonderful analogy. I'm keeping it. That said, this is not totally unlike many human explanations. In fact, increasingly, if a decision can be explained entirely as a product of decision making policies, it's likely to have been automated already. Eg. If I ask you why you chose to give this analogy, you will probably be able to provide a plausible explanation. But, that explanation is probably constructed after the fact. What we decide to say or write "comes to us" and we don't really have access to how that works. This is also true of judge's rulings, for example. They refer to rules (laws, precedents and legal concepts), and rulings must be justifiable in these terms... but the giraffe metaphor still plays. I suppose this is equivocation in a sense, but I'm not arguing that human and AI decision making are equivalents whether or not they share this core feature.

TeMPOraL 5y ago
This looks very similar to the argument that there's no problem with black-box DNN-based algorithms making important decisions about people, because humans are also black boxes and their explanations are just after-the-fact rationalizations. To which I counter: yes, humans are black boxes too. But they're all similar black boxes. We all share the same brain architecture, the same firmware. We have specialized hardware in our brains to simulate each other, predict other people's reactions. This is what allows us to work in groups and form societies. Additionally, as societies, we've learned to force predictability on each other. In the past, high-variance individuals - ones that are too unpredictable - were shunned or outright killed. Today, we isolate most severe cases from society entirely, by putting them in prisons or mental hospitals, and isolate less severe cases from high-impact responsibilities - e.g. by gating pilot licenses, driver licenses and gun permits behind various kinds of physical and mental health evaluations. For the day-to-day needs, humans are very predictable indeed, and their explanations are good enough. We know how to compensate for unreliable introspection. That's not the case with modern machine learning - the way DNN models "think" is completely alien to us. We can't work with this. We also don't have to, those algorithms are our creations. We can make them explainable. And we need to.
- dalbasal 5y ago
  As I said, while I'm technically equivocating (you too), I am not really saying that there are no reasons to be skeptical of black box algorithms. We agree on this, I think. That said, I'm not sure that we can make algorithms explainable. We may get better at interrogating them, and the analogy to human "black boxes" may get spookier, but I suspect that they're unexplainable in a "hard" sense. An explainable algorithm would have made different decisions, and probably won't be as good at making them.

TeMPOraL 5y ago

This looks very similar to the argument that there's no problem with black-box DNN-based algorithms making important decisions about people, because humans are also black boxes and their explanations are just after-the-fact rationalizations. To which I counter: yes, humans are black boxes too. But they're all similar black boxes. We all share the same brain architecture, the same firmware. We have specialized hardware in our brains to simulate each other, predict other people's reactions. This is what allows us to work in groups and form societies. Additionally, as societies, we've learned to force predictability on each other. In the past, high-variance individuals - ones that are too unpredictable - were shunned or outright killed. Today, we isolate most severe cases from society entirely, by putting them in prisons or mental hospitals, and isolate less severe cases from high-impact responsibilities - e.g. by gating pilot licenses, driver licenses and gun permits behind various kinds of physical and mental health evaluations. For the day-to-day needs, humans are very predictable indeed, and their explanations are good enough. We know how to compensate for unreliable introspection. That's not the case with modern machine learning - the way DNN models "think" is completely alien to us. We can't work with this. We also don't have to, those algorithms are our creations. We can make them explainable. And we need to.

dalbasal 5y ago
As I said, while I'm technically equivocating (you too), I am not really saying that there are no reasons to be skeptical of black box algorithms. We agree on this, I think. That said, I'm not sure that we can make algorithms explainable. We may get better at interrogating them, and the analogy to human "black boxes" may get spookier, but I suspect that they're unexplainable in a "hard" sense. An explainable algorithm would have made different decisions, and probably won't be as good at making them.