The Operation of Memory Systems in the Brain - Computational Neuroscience: A Comprehensive Approach - page 514

and [82]). Thus manipulations that impair the long-term potentiation of synapses
(LTP)air the formation of new short-term memory states, but not the use of previ-
ously learned short-term memory states. [37] analyzed many studies of the effects
of blockade of LTP in the hippocampus on spatial working memory tasks, and found
evidence consistent with this prediction. Interestingly, it was found that if there was
a large change in the delay interval over which the spatial information had to be re-
membered, then the task became susceptible, during the transition to the new delay
interval, to the effects of blockade of LTP. The implication is that some new learning
is required when the rat must learn the strategy of retaining information for longer
periods when the retention interval is changed.

16.4 Invariant visual object recognition

[74] proposed a feature hierarchical model of ventral stream visual objecting from
the primary visual cortex (V1), via V2 and V4 to the inferior temporal visual cortex
which could learn to represent objects invariantly with respect to position on the
retina, scale, rotation and view. The theory uses a short-term ('trace') memory term
in an associative learning rule to help capture the fact that the natural statistics of the
visual world reflect the fact that the same object is likely to be present over short-
time periods, for example over 1 or 2 seconds during which an object is seen from
different views. A model of the operation of the system has been implemented in a
four-layer network, corresponding to brain areas V1, V2, V4 and inferior temporal
visual cortex (IT), with convergence to each part of a layer from a small region of
the preceding layer, and with local competition between the neurons within a layer
implemented by local lateral inhibition [20, 82, 83, 116] (see Figure 16.12). During
a learning phase each object is learned. This is done by training the connections
between modules using a trace learning rule with the general form

y i τ x τ j

w ij = α (16.15)

where x τ j is the j th input to the neuron at time step

, y i is the output of the i th

neuron, and w ij is the j th weight on the i th neuron.
The trace y i τ is updated according to

y i τ =( y i τ + η y i τ − 1 1 − η ) . (16.16)

controls the relative contributions to the trace y i τ The parameter η ∈ [ 0 , 1 ] from

th e instantaneous firing rate y i τ at time step and the trace at the previous time step

y i τ − 1 .