Melanie Mitchell summarized a 2018 Santa Fe Institute workshop called Artificial Intelligence and the Barrier of Meaning published in AI Magazine (2020).
« The terms understanding and meaning are ill-defined. Marvin Minsky called such mental terms “suitcase words”, ones that are packed to the breaking point with different meanings. »
« Rather than proposing specific definitions, the workshop participants collectively listed a set of “correlates” of understanding in humans and other living systems, and discussed how these correlates contrast with today’s predominant AI systems. »
- « Core knowledge…
- Abstraction and generativity…
- Active perception, learning, and inference…
- Object-based, causal models…
- Metacognition…
- Embodiment…
- Evolutionary considerations…
- Sufficiency of the “information processing metaphor”… »
« Supervised machine learning is often framed in terms of distributions over the training and test data. In fact, the theoretical basis for much of machine learning requires that training and test examples are “independently and identically distributed” (IID). In contrast, human learning—and teaching—is active, sensitive to context, driven by top-down expectations, and transferable among highly diverse tasks, whose instances may be far from IID. »
« Going further, modern AI systems often focus on the optimization of a “cost function”. It’s unclear what should be “optimized” to achieve the kinds of correlates of understanding described in the previous section, or even if optimization itself is the right framework to be using. »
« ML systems are typically trained on narrow problems, using highly restricted datasets that are not necessarily “ecologically relevant” for developing understanding. »
« AI pioneer Marvin Minsky [said] “Though prescientific idea germs like ‘believe,’ ‘know,’ and ‘mean’ are useful in daily life, they seem technically too coarse to support powerful theories… Real as ‘self’ or ‘understand’ may seem to us today… they are only first steps towards better concepts.” »
Melanie Mitchell is Professor of Computer Science at Portland State University and the Davis Professor at the Santa Fe Institute. She originated the Santa Fe Institute’s Complexity Explorer project, which offers free online courses related to complex systems. She is author of Artificial Intelligence: A Guide for Thinking Humans (2019).
Articles of Interest 