Page 292 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

3. Prediction Versus Memory 285

and long-term prediction play a critical role. Simple tasks such as walking,
navigating, and many daily activities involving interaction with the environment
and with other humans require prediction. Long-term predictions of phenomena
such as seasonal changes enable planning and improved productivity. So, in a sense,
prediction is an important brain function, and it is increasingly being recognized as a
central function of the brain, as well as a key ingredient in intelligent machines (for
an overview of related ideas, see Andy Clark’s paper [2], and various papers on the
use of predicted future states in reinforcement learning).
In this section, I will talk about how such predictive function could have emerged
in the brain; how it is related to synaptic plasticity mechanisms (memory); how it is
relevant to the study of neural networks; and how predictive properties in the brain
can be linked to higher level phenomena such as consciousness.
First, consider delay in the nervous system. Neurons send their signals to their
receiving neurons via elongated wires called axons. Transmission through these
axons can take few milliseconds (ms); the duration depending on various factors
such as the length, diameter, and whether the axon is insulated with myelin or
not. When you add up the delay, it comes to a pretty significant amount of time:
about 180e260 ms from stimulus presentation to behavioral reaction [3]. This
kind of delay may be considered bad for the system, since it can be a matter of
life and death, especially for fast moving animals. Also, in engineering systems,
delay is considered a great hindrance. However, delay can be useful in two ways:
(1) in a reactive system such as a feedforward neural network, addition of delay
in the input can effectively add memory, and (2) mechanisms evolved to counteract
the adverse effects of delay can naturally lead to predictive capabilities.
In Ref. [4], we showed that addition of delay in feedforward neural network
controller can solve a 2D pole-balancing problem that does not include velocity
input. Also, in a series of works we showed that certain forms of synaptic plasticity
(dynamic synapses) can be considered as a delay compensation mechanism, and
how it relates to curious perceptual phenomena such as the flash lag effect (see
Ref. [5] for an overview). In flash lag effect, a moving object is perceived as being
ahead of a statically flashed object that is spatially aligned. One explanation for this
phenomenon is that the brain is compensating for the delay in its system, by gener-
ating an illusion that is aligned, in real time, with the current state of the external
environment. For example, image of the two aligned objects hit the retina. The
information takes several milliseconds to reach the visual area in the brain. In the
meanwhile, one of the objects moves ahead; so by the time the two objects are
perceived (when the information arrives in the visual area), in the environment,
they are misaligned because the moving object has moved on. The argument is
that flash lag effect allows the brain to perceive this as misaligned objects, which
is more in line with the actual environmental state at the time of perception. We
showed that facilitating neural dynamics, based on dynamic synapses (the facili-
tating kind, not the depressing one: see Henry Markram and colleagues’ work
cited in Ref. [5]), can replicate this phenomenon, and furthermore, the use of