Geology Reference
In-Depth Information
Fig. 4.8 Simple Hop
eld
network
to neural network-based research. However, these networks require a lot of training
to become an ef
eld
networks in practical time series application is limited due to theoretical restrictions
of the network structure. One simple Hop
cient and useful model. The relevance of the different Hop
eld network structure is shown in
Fig.
4.8
.
4.3.6 Long Short Term Memory Networks
A schematic view of a Simple Long Short Term Memory (LSTM) arti
cial neural
network is given in Fig.
4.9
. This is a type of recurrent neural network
rst pub-
lished in 1997 by Hochreiter and Schmidhuber [
34
]. This architecture contains
LSTM blocks in addition to regular networks. This
is able to
remember a particular value for an arbitrary length of time. This kind of model is
used in time series modeling, speech recognition, robot controls, etc. The practical
applications of conventional recurrent networks were limited in situations where
there are long lags between relevant events. LSTM models can solve this issue as,
when error values are back-propagated from the output, the error becomes mem-
orized in the memory portion of the
'
smart block
'
'
smart block,
'
and thus rapid decay of back-
propagation error is avoided.
