Page 295 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

288 CHAPTER 14 Meaning Versus Information, Prediction Versus Memory

question-asking has been employed in interactive problem-solving in robotics [13]
and vision problems [14]. However, these are done within a strict task framework;
so open-ended questions or questions that question the validity of existing questions
cannot be generated (see Ref. [15] for a bit more open-ended approach called inverse
Visual Question Answering). For an intelligent agent, this latter form of questioning
(or problem-posing) will become increasingly important, as the current learning
algorithms cannot easily go beyond its defined task context. Some ideas we dis-
cussed in Ref. [10] for problem-posing include: (1) recognizing events in the envi-
ronment that can potentially become a problem to be solved, (2) checking if existing
problems are ill-posed, and (3) given an overarching goal, come up with smaller
problems that may be of different kind than the original goal (if they are of the
same kind, straight-forward, divide-and-conquer algorithms can be used).
How can the idea of question versus answer be relevant to brain science? I think
it is relevant since the topic has not received attention that it deserves, and question
asking (or problem-posing) is an important function of the brain. There are many
papers on decision making, but not much on how the brain asks questions that
requires subsequent decision making. Understanding the brain mechanism of
question-asking can lead to new discoveries regarding core brain function, and in
turn the insight can help us build better intelligent artifacts.
In sum, question-asking needs more attention from brain science and artificial
intelligence, in order for us to gain a deeper understanding of brain function and
to build more intelligent artifacts.

5. DISCUSSION

In this chapter, I talked about several dichotomies of concepts that are important to
brain science and artificial intelligence: meaning versus information, prediction
versus memory, and question versus answer, with an emphasis on the first concept
in each pair. Below, I will discuss additional related works that have relevance to
the topics I discussed in the preceding sections.
In terms of meaning, deep neural network research approaches the issue from a
different angle than the one presented in this chapterdthat of embedding, for
example, word and sentence embedding [16]. The main idea of embedding is to
map words or sentences (or other raw input) into a vector space where concepts
can be manipulated algebraically based on their meaning. For example, when vector
representations of “Germany” and “capital” are added, the resulting vector repre-
sents “Berlin” (example from Ref. [16]). The main idea in this case is to train a
neural network to map the input word to a series of words that appear before or after
the input word in a sentence. The hidden layer representation then tends to have this
desired semantic property. This is one step toward meaningful information. However,
whether the meaning in this case is intrinsic to the system, that is, interpretable from
within the system, is an open question.