2. Only partial information is available to
cognitive agents, and the distribution of
available information varies across contexts.
Only some units in a neural network are fixed
(their values are given), and the distribution of the
fixed units varies across contexts.
examples
recommendation
predict the user’s evaluation of a product
unknown to her when her evaluations of some
other products are given
sentence recognition
infer the whole sentence when you see/hear only
parts of it
Partiality of Information
2
3. Intelligence is a combination of cycles
to keep creating value (suppressing entropy) by
hypothesis-test cycles (feedback loops) realizing
constraints (persisting generative models).
Cyclic neural networks could perform better
than feedforward ones (finite-iteration
approximation of cyclic ones), as programs
involving loops are much more powerful than
loop-free ones.
better accuracy, as demonstrated by dual learning
fewer layers & links
faster learning & less overlearning
Cyclic networks can combine parallelly with
each other, whereas feedforward networks
can be cascaded only.
Ubiquity of Cycles
3
4. execution & learning
Some units are fixed (their
values are given).
Sets of the fixed units
may differ for different
execution/learning
sessions.
The other units (their values)
are updated.
learning
Weights of links are updated.
convergence guaranteed?
probably yes, given
consistent data
Cyclic Network for Constraint Satisfaction
B
A
links from units in
A to units in B
layer (set of units)
4
5. autoencoder
One visible layer is fixed.
dual learning
One visible layer is fixed in
execution.
Two visible layers are fixed in
learning.
restricted Boltzmann machine
one visible and one hidden layer
Link weights are symmetric.
Different parts of the visible layer
are fixed in different sessions.
Examples of CNN
5
hidden
visible
visible
…
…
visible
visible
6. Cyclic networks can share units to affect each other,
implementing compound constraints.
applications
Languages use phonetics, phonology, lexicon, syntax, semantics,
pragmatics, common sense, etc.
Their integration could improve speech recognition in noisy
environment, etc.
Robots use mechanics, vision, language, common sense, etc.
Parallel Combination of Cyclic Networks
6
some units are shared