<s>
Catastrophic	B-Algorithm
interference	I-Algorithm
,	O
also	O
known	O
as	O
catastrophic	B-Algorithm
forgetting	I-Algorithm
,	O
is	O
the	O
tendency	O
of	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
to	O
abruptly	O
and	O
drastically	O
forget	O
previously	O
learned	O
information	O
upon	O
learning	O
new	O
information	O
.	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
are	O
an	O
important	O
part	O
of	O
the	O
network	O
approach	O
and	O
connectionist	O
approach	O
to	O
cognitive	O
science	O
.	O
</s>
<s>
Catastrophic	B-Algorithm
interference	I-Algorithm
is	O
an	O
important	O
issue	O
to	O
consider	O
when	O
creating	O
connectionist	O
models	O
of	O
memory	O
.	O
</s>
<s>
Specifically	O
,	O
these	O
problems	O
refer	O
to	O
the	O
challenge	O
of	O
making	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
that	O
is	O
sensitive	O
to	O
,	O
but	O
not	O
disrupted	O
by	O
,	O
new	O
information	O
.	O
</s>
<s>
Lookup	B-Data_Structure
tables	I-Data_Structure
and	O
connectionist	O
networks	O
lie	O
on	O
the	O
opposite	O
sides	O
of	O
the	O
stability	O
plasticity	O
spectrum	O
.	O
</s>
<s>
On	O
the	O
other	O
hand	O
,	O
connectionist	O
networks	O
like	O
the	O
standard	B-Algorithm
backpropagation	I-Algorithm
network	I-Algorithm
can	O
generalize	O
to	O
unseen	O
inputs	O
,	O
but	O
they	O
are	O
very	O
sensitive	O
to	O
new	O
information	O
.	O
</s>
<s>
Backpropagation	B-Algorithm
models	O
can	O
be	O
analogized	O
to	O
human	O
memory	O
insofar	O
as	O
they	O
mirror	O
the	O
human	O
ability	O
to	O
generalize	O
but	O
these	O
networks	O
often	O
exhibit	O
less	O
stability	O
than	O
human	O
memory	O
.	O
</s>
<s>
Notably	O
,	O
these	O
backpropagation	B-Algorithm
networks	O
are	O
susceptible	O
to	O
catastrophic	B-Algorithm
interference	I-Algorithm
.	O
</s>
<s>
This	O
is	O
an	O
issue	O
when	O
modelling	O
human	O
memory	O
,	O
because	O
unlike	O
these	O
networks	O
,	O
humans	O
typically	O
do	O
not	O
show	O
catastrophic	B-Algorithm
forgetting	I-Algorithm
.	O
</s>
<s>
The	O
term	O
catastrophic	B-Algorithm
interference	I-Algorithm
was	O
originally	O
coined	O
by	O
McCloskey	O
and	O
Cohen	O
(	O
1989	O
)	O
but	O
was	O
also	O
brought	O
to	O
the	O
attention	O
of	O
the	O
scientific	O
community	O
by	O
research	O
from	O
Ratcliff	O
(	O
1990	O
)	O
.	O
</s>
<s>
McCloskey	O
and	O
Cohen	O
(	O
1989	O
)	O
noted	O
the	O
problem	O
of	O
catastrophic	B-Algorithm
interference	I-Algorithm
during	O
two	O
different	O
experiments	O
with	O
backpropagation	B-Algorithm
neural	B-Architecture
network	I-Architecture
modelling	O
.	O
</s>
<s>
In	O
their	O
first	O
experiment	O
they	O
trained	O
a	O
standard	O
backpropagation	B-Algorithm
neural	B-Architecture
network	I-Architecture
on	O
a	O
single	O
training	O
set	O
consisting	O
of	O
17	O
single-digit	O
ones	O
problems	O
(	O
i.e.	O
,	O
1	O
+	O
1	O
through	O
9	O
+	O
1	O
,	O
and	O
1	O
+	O
2	O
through	O
1	O
+	O
9	O
)	O
until	O
the	O
network	O
could	O
represent	O
and	O
respond	O
properly	O
to	O
all	O
of	O
them	O
.	O
</s>
<s>
They	O
found	O
that	O
the	O
amount	O
of	O
training	O
on	O
the	O
A-C	O
list	O
in	O
Barnes	O
and	O
Underwood	O
study	O
that	O
lead	O
to	O
50%	O
correct	O
responses	O
,	O
lead	O
to	O
nearly	O
0%	O
correct	O
responses	O
by	O
the	O
backpropagation	B-Algorithm
network	O
.	O
</s>
<s>
However	O
none	O
of	O
these	O
manipulations	O
satisfactorily	O
reduced	O
the	O
catastrophic	B-Algorithm
interference	I-Algorithm
exhibited	O
by	O
the	O
networks	O
.	O
</s>
<s>
Ratcliff	O
(	O
1990	O
)	O
used	O
multiple	O
sets	O
of	O
backpropagation	B-Algorithm
models	O
applied	O
to	O
standard	O
recognition	O
memory	O
procedures	O
,	O
in	O
which	O
the	O
items	O
were	O
sequentially	O
learned	O
.	O
</s>
<s>
Well-learned	O
information	O
was	O
catastrophically	O
forgotten	O
as	O
new	O
information	O
was	O
learned	O
in	O
both	O
small	O
and	O
large	O
backpropagation	B-Algorithm
networks	O
.	O
</s>
<s>
The	O
main	O
cause	O
of	O
catastrophic	B-Algorithm
interference	I-Algorithm
seems	O
to	O
be	O
overlap	O
in	O
the	O
representations	O
at	O
the	O
hidden	O
layer	O
of	O
distributed	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
Catastrophic	B-Algorithm
forgetting	I-Algorithm
occurs	O
because	O
when	O
many	O
of	O
the	O
weights	O
where	O
"	O
knowledge	O
is	O
stored	O
"	O
are	O
changed	O
,	O
it	O
is	O
unlikely	O
for	O
prior	O
knowledge	O
to	O
be	O
kept	O
intact	O
.	O
</s>
<s>
Below	O
are	O
a	O
number	O
of	O
techniques	O
which	O
have	O
empirical	O
support	O
in	O
successfully	O
reducing	O
catastrophic	B-Algorithm
interference	I-Algorithm
in	O
backpropagation	B-Algorithm
neural	B-Architecture
networks	I-Architecture
:	O
</s>
<s>
According	O
to	O
French	O
(	O
1991	O
)	O
,	O
catastrophic	B-Algorithm
interference	I-Algorithm
arises	O
in	O
feedforward	B-Algorithm
backpropagation	B-Algorithm
networks	O
due	O
to	O
the	O
interaction	O
of	O
node	O
activations	O
,	O
or	O
activation	O
overlap	O
,	O
that	O
occurs	O
in	O
distributed	O
representations	O
at	O
the	O
hidden	O
layer	O
.	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
that	O
employ	O
very	O
localized	O
representations	O
do	O
not	O
show	O
catastrophic	B-Algorithm
interference	I-Algorithm
because	O
of	O
the	O
lack	O
of	O
overlap	O
at	O
the	O
hidden	O
layer	O
.	O
</s>
<s>
French	O
therefore	O
suggested	O
that	O
reducing	O
the	O
value	O
of	O
activation	O
overlap	O
at	O
the	O
hidden	O
layer	O
would	O
reduce	O
catastrophic	B-Algorithm
interference	I-Algorithm
in	O
distributed	O
networks	O
.	O
</s>
<s>
French	O
recommended	O
that	O
this	O
could	O
be	O
done	O
through	O
'	O
activation	O
sharpening	O
 '	O
,	O
a	O
technique	O
which	O
slightly	O
increases	O
the	O
activation	O
of	O
a	O
certain	O
number	O
of	O
the	O
most	O
active	O
nodes	O
in	O
the	O
hidden	O
layer	O
,	O
slightly	O
reduces	O
the	O
activation	O
of	O
all	O
the	O
other	O
units	O
and	O
then	O
changes	O
the	O
input-to-hidden	O
layer	O
weights	O
to	O
reflect	O
these	O
activation	O
changes	O
(	O
similar	O
to	O
error	O
backpropagation	B-Algorithm
)	O
.	O
</s>
<s>
Kortge	O
(	O
1990	O
)	O
proposed	O
a	O
learning	O
rule	O
for	O
training	O
neural	B-Architecture
networks	I-Architecture
,	O
called	O
the	O
'	O
novelty	O
rule	O
 '	O
,	O
to	O
help	O
alleviate	O
catastrophic	B-Algorithm
interference	I-Algorithm
.	O
</s>
<s>
As	O
its	O
name	O
suggests	O
,	O
this	O
rule	O
helps	O
the	O
neural	B-Architecture
network	I-Architecture
to	O
learn	O
only	O
the	O
components	O
of	O
a	O
new	O
input	O
that	O
differ	O
from	O
an	O
old	O
input	O
.	O
</s>
<s>
When	O
the	O
novelty	O
rule	O
is	O
used	O
in	O
a	O
standard	B-Algorithm
backpropagation	I-Algorithm
network	I-Algorithm
there	O
is	O
no	O
,	O
or	O
lessened	O
,	O
forgetting	O
of	O
old	O
items	O
when	O
new	O
items	O
are	O
presented	O
sequentially	O
.	O
</s>
<s>
McRae	O
and	O
Hetherington	O
(	O
1993	O
)	O
argued	O
that	O
humans	O
,	O
unlike	O
most	O
neural	B-Architecture
networks	I-Architecture
,	O
do	O
not	O
take	O
on	O
new	O
learning	O
tasks	O
with	O
a	O
random	O
set	O
of	O
weights	O
.	O
</s>
<s>
Robins	O
(	O
1995	O
)	O
described	O
that	O
catastrophic	B-Algorithm
forgetting	I-Algorithm
can	O
be	O
prevented	O
by	O
rehearsal	O
mechanisms	O
.	O
</s>
<s>
This	O
means	O
that	O
when	O
new	O
information	O
is	O
added	O
,	O
the	O
neural	B-Architecture
network	I-Architecture
is	O
retrained	O
on	O
some	O
of	O
the	O
previously	O
learned	O
information	O
.	O
</s>
<s>
French	O
(	O
1997	O
)	O
proposed	O
a	O
pseudo-recurrent	O
backpropagation	B-Algorithm
network	O
(	O
see	O
Figure	O
2	O
)	O
.	O
</s>
<s>
Inspired	O
by	O
and	O
independently	O
of	O
Ans	O
and	O
Rousset	O
(	O
1997	O
)	O
also	O
proposed	O
a	O
two-network	O
artificial	O
neural	O
architecture	O
with	O
memory	O
self-refreshing	O
that	O
overcomes	O
catastrophic	B-Algorithm
interference	I-Algorithm
when	O
sequential	O
learning	O
tasks	O
are	O
carried	O
out	O
in	O
distributed	O
networks	O
trained	O
by	O
backpropagation	B-Algorithm
.	O
</s>
<s>
What	O
mainly	O
distinguishes	O
this	O
model	O
from	O
those	O
that	O
use	O
classical	O
pseudorehearsal	O
in	O
feedforward	B-Algorithm
multilayer	O
networks	O
is	O
a	O
reverberating	O
process	O
that	O
is	O
used	O
for	O
generating	O
pseudopatterns	O
.	O
</s>
<s>
After	O
a	O
number	O
of	O
activity	O
re-injections	O
from	O
a	O
single	O
random	O
seed	O
,	O
this	O
process	O
tends	O
to	O
go	O
up	O
to	O
nonlinear	O
network	O
attractors	O
that	O
are	O
more	O
suitable	O
for	O
capturing	O
optimally	O
the	O
deep	O
structure	O
of	O
knowledge	O
distributed	O
within	O
connection	O
weights	O
than	O
the	O
single	O
feedforward	B-Algorithm
pass	O
of	O
activity	O
used	O
in	O
pseudo-rehearsal	O
.	O
</s>
<s>
The	O
memory	O
self-refreshing	O
procedure	O
turned	O
out	O
to	O
be	O
very	O
efficient	O
in	O
transfer	O
processes	O
and	O
in	O
serial	O
learning	O
of	O
temporal	O
sequences	O
of	O
patterns	O
without	O
catastrophic	B-Algorithm
forgetting	I-Algorithm
.	O
</s>
<s>
Such	O
generative	O
replay	O
can	O
effectively	O
prevent	O
catastrophic	B-Algorithm
forgetting	I-Algorithm
,	O
especially	O
when	O
the	O
replay	O
is	O
performed	O
in	O
the	O
hidden	O
layers	O
rather	O
than	O
at	O
the	O
input	O
level	O
.	O
</s>
<s>
Latent	O
Learning	O
is	O
a	O
technique	O
used	O
by	O
Gutstein	O
&	O
Stump	O
(	O
2015	O
)	O
to	O
mitigate	O
catastrophic	B-Algorithm
interference	I-Algorithm
by	O
taking	O
advantage	O
of	O
transfer	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
In	O
effect	O
,	O
this	O
technique	O
uses	O
transfer	B-General_Concept
learning	I-General_Concept
to	O
avoid	O
catastrophic	B-Algorithm
interference	I-Algorithm
,	O
by	O
making	O
a	O
network	O
's	O
responses	O
to	O
new	O
classes	O
as	O
consistent	O
as	O
possible	O
with	O
existing	O
responses	O
to	O
classes	O
already	O
learned	O
.	O
</s>
<s>
(	O
2017	O
)	O
proposed	O
elastic	O
weight	O
consolidation	O
(	O
EWC	O
)	O
,	O
a	O
method	O
to	O
sequentially	O
train	O
a	O
single	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
on	O
multiple	O
tasks	O
.	O
</s>
<s>
This	O
technique	O
supposes	O
that	O
some	O
weights	O
of	O
the	O
trained	O
neural	B-Architecture
network	I-Architecture
are	O
more	O
important	O
for	O
previously	O
learned	O
tasks	O
than	O
others	O
.	O
</s>
<s>
During	O
training	O
of	O
the	O
neural	B-Architecture
network	I-Architecture
on	O
a	O
new	O
task	O
,	O
changes	O
to	O
the	O
weights	O
of	O
the	O
network	O
are	O
made	O
less	O
likely	O
the	O
greater	O
their	O
importance	O
.	O
</s>
