Page 291 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

284 CHAPTER 14 Meaning Versus Information, Prediction Versus Memory

bulbs? Suppose the joystick is linked to a camera external to the room and the direc-
tion of gaze of the camera follows the joystick control. In the external visual envi-

ronment, assume there is a single long line that is oriented 45 degrees , and that the
camera is pointed toward one segment of the line. In this setup, the second light bulb
(representing 45 degrees) will be turned on. If the joystick is moved in a direction
other than 45 and 225 degrees, the lights will go off (note: if there were other lines
in the environment, a different light will turn on). However, when the joystick is
moved in these two directions (45 and 225 degrees), the second light bulb will be
kept turned on (i.e., it will remain invariant). In this case, the property of the second
light bulb and the property of the movement that keeps the light invariant are exactly
aligned. Through this kind of sensorimotor exploration, the property of the internal
representation can be recovered, from within the system (without direct perceptual
access to the external environment), thus the meaning can remain intrinsic to the
system. In our lab, we explored these ideas in a reinforcement learning setting (learn
a policy p that maps from state S [orientation] to action A [gaze direction]), where we
showed that the internal state invariance criterion (the reward) can be used for motor
grounding of internal sensory representation in a simple visuomotor agent. See
Ref. [1] and subsequent works for more details.
To sum up, meaning is central to brain science and artificial intelligence, and to
provide meaning to information, it is critical to consider the sensorimotor aspect of
the information system, whether natural or artificial.

3. PREDICTION VERSUS MEMORY
Many questions in brain and neuroscience focus on the concept of plasticity, how the
brain changes and adapts with experience, and this leads to the question of memory.
Connections between neurons adapt over time (synaptic plasticity: long term, short
term, etc.), and ongoing neural dynamic of the brain can also be altered by the im-
mediate input stimulus. On a higher level, plasticity is usually considered in relation
to various forms of memory: long-term memory, short-term memory, working mem-
ory, episodic memory, implicit memory, explicit memory, etc. Also, in a common
sense way, people ask how the brain remembers, and what constitutes memory in
the brain. In artificial intelligence, the same is true: How information should be
represented, stored, and retrieved? How connection weights should be adapted in
artificial neural networks to store knowledge? How neural networks can be used
to utilize external memory? etc.
What is memory, and how is it related to prediction, and why should we think
more about prediction than memory? Memory is backward-looking, directed toward
the past, while prediction is forward-looking, and is directed toward the future.
Memory enables prediction, since without memory, the system will be purely reac-
tive, living in the eternal present. So, again, why should we direct our attention
toward prediction? In terms of brain function and artifacts that try to mimic it, pre-
diction is of prime importance. In our everyday life, moment-to-moment prediction