Feature, Configuration, History
A Bio-inspired Framework for Information Representation in Neural Networks
eric Alexandre, Maxime Carrere and Randa Kassab
Inria Bordeaux Sud-Ouest, 200 Avenue de la Vieille Tour, 33405 Talence, France
LaBRI, Universit
e de Bordeaux, Institut Polytechnique de Bordeaux, CNRS, UMR 5800, Talence, France
Institut des Maladies Neurod
eratives, Universit
e de Bordeaux, CNRS, UMR 5293, Bordeaux, France
Information Representation, Computational Neuroscience, Pavlovian Conditioning.
Artificial Neural Networks are very efficient adaptive models but one of their recognized weaknesses is about
information representation, often carried out in an input vector without a structure. Beyond the classical
elaboration of a hierarchical representation in a series of layers, we report here inspiration from neuroscience
and argue for the design of heterogenous neural networks, processing information at feature, configuration and
history levels of granularity, and interacting very efficiently for high-level and complex decision making. This
framework is built from known characteristics of the sensory cortex, the hippocampus and the prefrontal cortex
and is exemplified here in the case of pavlovian conditioning, but we propose that it can be advantageously
applied in a wider extent, to design flexible and versatile information processing with neuronal computation.
Artificial Neural Networks (ANNs) have proven to
be an excellent formalism for adaptive information
processing and have been successfully applied to an
impressive range of application domains for such
tasks as classification, control and prediction. One
of the reasons for the undebatable success of ANNs
is certainly its unique position between the fertile
fields of Machine Learning and Computational Neu-
roscience. On the one hand, Computational Neuro-
science irrigates ANNs with architectural and com-
putational principles inspired from observation of the
living brain. On the other hand, Machine Learning an-
chors in solid mathematical grounds the most promi-
nent features of ANNs including learning algorithms.
Cross-fertilization from both domains explains
most of the advances of ANNs in the past decades,
even if some weaknesses remain. Among them, one
of the most critical flaws of ANNs is certainly re-
lated to information representation. Basically, ANNs
are mainly dedicated to unstructured information pro-
cessing (their input is a ’flat’ vector of data with-
out structure), whereas most real-world applications
require the elaboration and exploitation of complex
structures for an adequate knowledge representation.
Several measures have already been proposed to
overcome this problem. Firstly, input vectors can be
built according to a structured representation of infor-
mation proposed a priori, but this cannot be adapted
by learning and cannot be applied to the inner rep-
resentation of ANNs. Secondly, in opposition to
the connectionist approach, the symbolic approach is
generally more efficient at defining and processing
elaborated knowledge representations and it has been
proposed accordingly to associate both paradigms in
the so-called Connectionist-Symbolic Integration ap-
proach (Sun and Alexandre, 1997) but here also, the
two kinds of modules are generally too independent.
Thirdly, some interesting attempts have been made to
propose fully connectionist solutions to this problem.
Connectionist solutions to structured information
processing are mainly centred on innovative archi-
tectures and propose to adapt existing learning algo-
rithms to exploit them. One of the early models in
this direction, the LRAAM network (Sperduti et al.,
1997), explains how a recurrent elaboration of a net-
work allows to encode graphs. More recently, deep
neural networks (Bengio et al., 2013) renew studies in
which complex representations can be extracted from
many-layered neural networks, as suggested by the ar-
chitecture of the visual system in the brain (Rousselet
et al., 2004).
These solutions to build and manipulate complex
representations in ANNs are diverse (recursive or iter-
ative), but both are hierarchical : Complex represen-
Alexandre F., Carrere M. and Kassab R..
Feature, Configuration, History - A Bio-inspired Framework for Information Representation in Neural Networks.
DOI: 10.5220/0005156003160321
In Proceedings of the International Conference on Neural Computation Theory and Applications (NCTA-2014), pages 316-321
ISBN: 978-989-758-054-3
2014 SCITEPRESS (Science and Technology Publications, Lda.)
tations are seen as more and more abstract represen-
tations. In this position paper, we argue that these ar-
chitectural solutions are not fully satisfactory because
they miss an important dimension, the granularity of
information representation. Taking inspiration from
the brain, it can be remarked that hierarchical integra-
tion is not the only way to elaborate structured infor-
mation representation : In the brain, specialized mod-
ules extract information in the data flow at various lev-
els of granularity and mix them in heterogenous (and
not only hierachical) systems for elaborated decision-
making processes. As far as sensory information pro-
cessing is concerned, we propose that these levels of
granularity are of three kinds:
1. Feature. At a conceptual level, features can be
extracted in the information flow and arranged in
a hierarchical way;
2. Configuration. At the individual level, the ag-
glomeration in a ’flat’ vector of all the perceived
sensory information allows for the representation
of specific cases;
3. History. At the statistical level, overall trends can
be extracted from the recent history of perception.
In the remaining of the paper, we first give some
elements from neuroscience for the existence of such
representations in the brain of mammals, then we pro-
pose to illustrate them and put them in situation in the
case of pavlovian conditioning, before discussing the
usefulness of such representations and of their associ-
ation in heterogenous systems.
Information representation in a hierarchical way lead-
ing to abstraction of concepts has been extensively
reported in the brain, for example in the visual cor-
tex. This hierarchical view of knowledge representa-
tion has already inspired ANNs like deep networks,
as evoked in section 1. From Hubel and Wiesel ini-
tial idea that neurons selective to orientation in the
primary visual cortex are built from the activity of
contrast-sensitive cells in a receptive field in the tha-
lamus (Hubel and Wiesel, 1962), further experiments
along the ventral visual pathway have identified neu-
ronal areas responding to a hierarchy of concepts
(Rousselet et al., 2004) captured in increasingly large
receptive fields, from simple shapes to complex pat-
terns like faces. In that respect, the ventral pathway of
the cortex is proposed as the locus of semantic mem-
ory, with a cascade of neurons extracting a hierarchy
of concepts. Here we will speak about feature repre-
Nevertheless, hierarchy is probably not the main
characteristic to be evoked when characterizing infor-
mation processing in the brain, but rather modular-
ity and heterogeneity. Indeed, the brain is organized
along complementary learning systems, as it has al-
ready been captured in early modeling approaches
(Doya, 1999; Alexandre, 2000).
Particularly, learning mode and information rep-
resentation in the hippocampus have been contrasted
with that of the cortex described above (McClelland
et al., 1995). Whereas the cortex learns slowly to
elaborate a semantic memory as a hierarchy of con-
cepts that will be exploited in generalization, the hip-
pocampus receives as sensory input a summary sketch
of current cortical activity and learns in one shot an
episodic memory of a specific event in its spatial and
temporal context. Contrarily to the cortical case, this
memory process is not prone to generalization; in-
stead, when faced to a similar episode, the recall pro-
cess will rebuild the original episode: the represen-
tation in the hippocampus is consequently unitary,
storing configural representations of specific events
(O’Reilly and Rudy, 2001). Here we will speak about
representation of configuration.
Though complementary, these memories are also
in cooperation : The hippocampus is fed with sensory
inputs encoded with cortical features and reciprocally
participates to the slow cortical learning through the
phenomenon of consolidation : Specific cases stored
in the hippocampus during some task are sent back to
the cortex for the extraction of new features, pertinent
to the task (O’Reilly and Rudy, 2001). Whereas the
cortex can work in generalization, the task is more
difficult for the hippocampus : Faced to an episode
similar to a stored one, the hippocampus can only
choose between pattern completion (supposing that
the episodes are equivalent, the stored one will be
recalled) and pattern separation (supposing that the
episodes must be distinguished one from the other,
the new one will be stored as a different pattern)
(O’Reilly and Rudy, 2001).
As far as task execution is concerned, the pre-
frontal cortex is certainly one of the most involved
cerebral structures (Fuster, 2001), since it supports
the temporal organization of behavior, i.e. selecting
at each moment, the best action (or decision) to trig-
ger, depending on the external (perceptive) and in-
ternal (emotional and motivational) sensory informa-
tion. Most of the time this decision is not purely re-
active (not depending directly on current stimuli) but
deliberative (depending on a more complex evalua-
tion, also based on past experience). To that end, the
prefrontal cortex builds task sets and maintains them
active in a working memory mechanism, for compar-
ison and selection (by a loop with subcortical struc-
tures, the basal ganglia) of the most adapted solution.
In short, task sets are collections of recent cases of
associations between actions (or decision) that have
been triggered in certain sensory and temporal con-
texts, together with their performance in reaching the
goal that was expected.
The prefrontal cortex is organized, from its pre-
motor to its anterior part, in a hierarchy of levels of
control (Koechlin et al., 2003) depending on the na-
ture of the sensory and temporal context (and on the
structure that sends this information). This cascade
of control is made according to the perceptual con-
text, the episodic context and the motivational and
emotional context (Kouneiher et al., 2009). At each
level of description, the decision to engage an action
is taken, based on its recent history of success in this
context. This statistical analysis makes us describe
the prefrontal cortex as a locus for the representation
of history.
To better understand the way these representations
are elaborated in a concrete case, their association and
their impact on cerebral processes, we propose an il-
lustration in the case of pavlovian conditioning.
Pavlovian or respondent conditioning is the learning
process by which an animal is able to associate an Un-
conditional Stimulus (US = biologically significant
stimulus announcing pain or pleasure, e.g. an electric
shock) to a conditional stimulus (CS, a neutral event
like a tone, that predicts the US). This learning has
been extensively studied in so-called fear condition-
ing experiments and is reported to be acquired quickly
if the electric shock is given subsequently to the tone
(Herry et al., 2008). Later on, the animal exhibits fear
behavior (ex: freezing) when solely exposed to the au-
ditory CS. This can be explained as an anticipation or
a preparation to the forthcoming pain. This response
will disappear if the electric shock is no longer given
(extinction process) and get re-activated by a new as-
sociation (renewal process). As such, this simple de-
scription could assimilate pavlovian conditioning to
an elementary associative learning where, for exam-
ple, an hebbian learning rule could increase and de-
crease a weight between two neurons standing for the
CS and the US. But reality is much more complex,
as shown by many behavioral studies that required to
refine our understanding on pavlovian conditioning.
One of the most famous paradigms is the blocking
paradigm : In an early step, an association CS1-US is
acquired. Then CS1 is paired with CS2 to announce
the US. Here it could be said that CS2 becomes also
a faithful predictor of US and, from a purely asso-
ciative learning point of view, it should acquire also
a predictive power. This is not confirmed by exper-
iments: CS2 alone triggers no conditioned response.
This paradigm has be interpreted as a parsimonious
learning (CS1 is sufficient to predict US; no need to
perform a new learning about CS2) and led to pro-
pose a non purely associative learning rule in the early
70’s, the competitive Rescorla-Wagner learning rule
(Rescorla and Wagner, 1972). This rule states that
the modification of the strength of association (the
weight, considering a neuronal implementation with
an input layer of CS connected to an output layer of
US) between CSi and US is proportional to the er-
ror of prediction, i.e. the difference between what ac-
tually happens (the US) and what was expected (the
sum of the predictive values of all present CSi). Two
other multiplicative terms in the rule are defined as
the associability (saliency) of the CS and the effec-
tiveness (behavioral importance) of the US and are
often defined as constant (but cf. below).
Now comes the question of the nature of the CS
and here also behavioral experiments are confusing:
CS often corresponds to a salient stimulus (a tone, a
flashing light), but sometimes as in the case of extinc-
tion and renewal, the pertinent element is the context
in which conditioning occurs. This makes often mod-
elers add a specific neuron in the CS layer standing
for ’the context’, even if this is not fully satisfactory,
concerning representation of information. Another
problem is about configural learning : In a prototypi-
cal case, CS1 or CS2 alone predicts US, whereas the
occurrence of both predicts no US. This is difficult to
explain with representation of single CSs because in
this case CS1 and CS2 will acquire a strong predictive
strength and the conjunction of both should add these
strengths and predict the US. A practical solution to
this classical problem is to create a new representa-
tion for CS1+CS2, seen as a new configural stimulus
(Schmajuk and DiCarlo, 1992) but this raises again
the question of building by hand the representation of
sensory information during pavlovian conditioning.
All the models of pavlovian learning evoked so
far are said US-processing models because their main
goal is to predict the US from current CSs and con-
sequently the same rule of predictive strength mod-
ification is applied to each CS, whatever its history,
which is not realistic, as other behavioral experiments
indicate (Le Pelley, 2004), hence the need for CS-
processing models.
On the one hand, the Pearce-Hall model (Pearce
and Hall, 1980) tries to capture the fact that a CS of-
ten associated with US learns less quickly than a not
well known CS and proposes that the term of associa-
bilility of the CS could change in time and depend on
the level of surprise (absolute value of error of predic-
tion) and also on its own current value.
On the other hand, the Mackintosh model (Mack-
intosh, 1975) maintains a high associability for reli-
able CS, to explain the fact that assigning a CS to a
new US is quicker after an overtaining of the CS with
the previous US. Both cases underline the need to take
into account the recent history of conditioning. They
also raise new problems since it is obvious that the
Pearce-Hall and Mackintosh models are somewhat
contradictory, respectively proposing to decrease and
increase the associability of a frequent CS.
Confronting abstract models to behavioral obser-
vations is also an incitation to integrate more realistic
constraints in the models. This is for example the case
with uncertainty : Our word is stochastic (real CSs
do not always reliably predict US) and non station-
ary (a CS-US association rule valid at one time can be
obsolete the time after). Classical models of pavlo-
vian conditioning generally address poorly this kind
of constraints, which is a strong incentive to develop
more realistic models, for example including bayesian
approaches (Yu and Dayan, 2005; Deco et al., 2008).
In summary, this brief overview of pavlovian con-
ditioning shows that, more than one hundred years af-
ter its identification, pavlovian conditioning is still not
fully understood. We have mentionned several obser-
vations non consistent with the current understanding
of this learning mechanism and, for each of them, a
model yielding the corresponding behavior, but at the
moment, there is no complete model integrating all
these mechanisms. We think that going deeper in the
understanding of the underlying cerebral substrate is
an excellent way to design such a complete model.
The amygdala is the core cerebral structure for pavlo-
vian conditioning (Holland and Gallagher, 1999) but
it cannot be considered in isolation for two reasons.
On the one hand, the amygdala receives from other
cerebral structures, sensory information about the CS
and the US, for their association, but also other kinds
of information needed at various steps of the condi-
tioning. On the other hand, pavlovian conditioning re-
sults in a series of effects, that are transmitted to other
regions in the brain. These effects are of course the
pavlovian response itself, but also other attentional
and representational effects.
Amygdala is in fact a complex and heterogenous
structure (LeDoux, 2007), composed of several sub-
divisions and it can be convenient to distinguish three
regions with specific functions and connectivities :
1. The lateral amygdala (LA) receives sensory infor-
mation about the US and the CS from the thala-
mus and the cortex. This region is responsible for
learning the CS-US association.
2. The central nucleus of the Amygdala (CeA) is in
charge of the three aspects of the pavlovian re-
sponse: the motor aspect (e.g. in the case of
fear, freezing), the autonomic aspect (e.g. phys-
iological changes like heart beat or temperature
increase) and the hormonal aspect (release of
stress hormones and neuromodulators like acetyl-
choline, norepinephrine, dopamine, etc.).
3. The basal amygdala (BA) has a major represen-
tational role. It receives from LA information
about CS-US association, modulates this informa-
tion with contextual inputs coming from the hip-
pocampus and prefrontal, to elaborate a represen-
tation about the sensory nature of the US (Cardi-
nal et al., 2002). This information is sent to the
CeA to produce the pavlovian response but also
to the sensory cortex (attentional affect) and to the
prefrontal cortex-basal ganglia system (emotional
evaluation of stimuli has also strong effect in de-
cision making and operant conditioning).
This very rough description of the amygdalar
circuitry allows to put into context all the mecha-
nisms and observations about pavlovian conditioning
evoked in section 3. We posit a general computational
principle with this circuitry : Sensory thalamic and
cortical inputs to LA (generic features) should be suf-
ficient to reliably predict US but other sensory infor-
mation coming from the hippocampus to BA will pos-
sibly propose CS-US association rules corresponding
to more specific cases. Later on, these specific cases
could be ’compiled’ from the hippocampus, in the
creation of new features in the cortex.
The LA and BA regions project in CeA to trig-
ger the good pavlovian response. The choice between
the corresponding CS-US rules will be learned from
a monitoring of their performance in predicting US,
evaluated in the orbitofrontal cortex (OFC). This ante-
rior region of the prefrontal cortex has very dense re-
lations with all regions of the amygdala and builds an
history of errors of US prediction, depending on the
sensory context (in the cortex and in the hippocam-
pus) in which they occurred (Pauli et al., 2011). These
errors of prediction are of two kinds.
A positive error of prediction corresponds to the
case where an US is received whereas it was not pre-
dicted by LA neither by BA : in this case BA will send
a signal to the hippocampus to make this structure
learn by heart the current episode (Paz et al., 2009).
On the next occurrence of this episode the hippocam-
pus will be able to recall the corresponding US and
send a configural representation to BA. If this is often
repeated, this will result in consolidation in the cortex
and in the emergence of a feature representation in the
cortex, leading to a reliable rule in LA.
When a negative error of prediction occurs (from
the current sensory context, an US has been predicted
and doesn’t occur), several causes can be evoked :
1. The rule is valid but the specific context corre-
sponds to an exception which must be stored in
the hippocampus as associated to no US (case of
an extinction). If this happens again in this con-
text, the rule in BA based on this hippocampal in-
put will be favored in this specific case with regard
to the general rule coming from LA.
2. The rule is stochastic and not fully reliable. In
this case the level of stochasticity of the rule has
to be updated in OFC. This case and the previ-
ous case also result in an increase of the level of
acetylcholine (Yu and Dayan, 2005), which favors
learning in BA, in the hippocampus and in the cor-
tex, with the aim of making more precise the rules
currently under consideration.
3. The rule is no longer valid because the world is
non stationary but it could be valid again in the
future. The rule is consequently conserved with-
out modification but inhibited by OFC that trig-
gers a release of norepinephrine (Yu and Dayan,
2005). This neuromodulator acts on the thalamo-
cortical inputs of LA, to look for a new rule (John-
son et al., 2011).
These causes have to be distinguished from the his-
tory of performance in US prediction stored in OFC.
A corresponding algorithm has been proposed by (Yu
and Dayan, 2005), to evaluate the decision threshold
between stochasticity and non stationarity : When the
rule is highly stochastic, many examples are needed
before deciding for a new rule; conversely when the
rule is very reliable, this change will be made more
easily. This alternative is also useful to propose a way
to release the seamingly contradictory effects between
the Pearce-Hall and Mackintosh models : The Mack-
intosh model could apply only when the rule is judged
valid and stable (no error of prediction), to evaluate
its level of reliability, whereas the Pearce-Hall model
could apply only in the early stages of selection of the
rule, to explore and detect candidate CS.
In this paper, we have laid emphasis on the fact that
different levels of granularity of information are rep-
resented and processed in different brain regions and
can be manipulated in ANNs as well : Multi-layered
ANNs can extract a set of features, as a hierarchy
of concepts, with some analogy to the sensory cor-
tex. Associative memories like the Hopfield model
have been shown to learn by heart configuration of
features representing specific cases in a recurrent ar-
chitecture, as it is the case in the hippocampus. In ad-
dition to these cases, history of the statistical tenden-
cies of certain information flows is believed to be cap-
tured in the prefrontal cortex. Though not yet mature
in ANNs, early models implementing this latter kind
of processing are proposed (Kouneiher et al., 2009).
In this paper, we have also shown, through the ex-
ample of pavlovian conditioning, why mixing these
levels of granularity is so powerful. On the one
hand, the combination of associative rules involv-
ing features and configurations is of central impor-
tance in this learning and is more generally reminis-
cent of the principle of mixing generic rules and spe-
cific cases. On the other hand, it seems possible to
reconcile different cases where pavlovian condition-
ing works differently by performing a statistical anal-
ysis that allows to distinguish different functioning
modes. Selecting strategies and switching between
procedures seem to be a major role of the prefrontal
cortex (Fuster, 2001; Koechlin et al., 2003), fed by
contextual and episodic information from the sensory
cortex and the hippocampus respectively, and oper-
ated through its control on many cerebral structures,
sometimes by the release of neuromodulators.
Pavlovian conditioning was chosen here to exem-
plify these principles but many characteristics of the
cerebral system support the fact that computing with
these different levels of representation is more gen-
eral than this only case. In addition, we are deeply
convinced that computing with these levels of rep-
resentation might be advantageously exported to the
domain of ANNs. It could provide this domain with
an efficient way to combine generic rules and spe-
cific cases and to make decisions based on the evalu-
ation of performances of such heterogenous modules,
which seems to underlie an important part of our cere-
bral system and its emergent cognitive capabilities.
This work was partly supported by the Keops ANR
Alexandre, F. (2000). Biological Inspiration for Multiple
Memories Implementation and Cooperation. In In
V. Kvasnicka P. Sincak, J. Vascak and R. Mesiar, ed-
itors, International Conference on Computational In-
Bengio, Y., Courville, A. C., and Vincent, P. (2013). Rep-
resentation learning: A review and new perspectives.
IEEE Trans. Pattern Anal. Mach. Intell., 35(8):1798–
Cardinal, R. N., Parkinson, J. A., Hall, J., and Everitt, B. J.
(2002). Emotion and motivation: the role of the amyg-
dala, ventral striatum, and prefrontal cortex. Neuro-
science & Biobehavioral Reviews, 26(3):321–352.
Deco, G., Jirsa, V. K., Robinson, P. A., Breakspear, M., and
Friston, K. (2008). The dynamic brain: From spiking
neurons to neural masses and cortical fields. PLoS
Comput. Biol., 4(8).
Doya, K. (1999). What are the computations of the cerebel-
lum, the basal ganglia and the cerebral cortex? Neural
Networks, 12:961–974.
Fuster, J. M. (2001). The Prefrontal CortexAn Update :
Time Is of the Essence. Neuron, 30(2):319–333.
Herry, C., Ciocchi, S., Senn, V., Demmou, L., Muller, C.,
and Luthi, A. (2008). Switching on and off fear by dis-
tinct neuronal circuits. Nature, 454(7204):600–606.
Holland, P. C. and Gallagher, M. (1999). Amygdala cir-
cuitry in attentional and representational processes.
Trends in cognitive sciences, 3(2):65–73.
Hubel, D. and Wiesel, T. (1962). Receptive fields, binocular
interaction and functional architecture in the cat visual
cortex. J Physiol, 160:106–154.
Johnson, L. R., Hou, M., Prager, E. M., and LeDoux, J. E.
(2011). Regulation of the Fear Network by Media-
tors of Stress: Norepinephrine Alters the Balance be-
tween Cortical and Subcortical Afferent Excitation of
the Lateral Amygdala. Frontiers in Behavioral Neu-
roscience, 5.
Koechlin, E., Ody, C., and Kouneiher, F. (2003). The Archi-
tecture of Cognitive Control in the Human Prefrontal
Cortex. Science, 302(5648):1181–1185.
Kouneiher, F., Charron, S., and Koechlin, E. (2009). Mo-
tivation and cognitive control in the human prefrontal
cortex. Nature Neuroscience, 12(7):939–945.
Le Pelley, M. E. (2004). The role of associative history in
models of associative learning: a selective review and
a hybrid model. The Quarterly Journal of Experimen-
tal Psychology, 57(3):193–243.
LeDoux, J. (2007). The amygdala. Current Biology,
Mackintosh, N. J. (1975). A theory of attention: Variations
in the associability of stimuli with reinforcement. Psy-
chological Review, 82(4):276–298.
McClelland, J. L., McNaughton, B. L., and O’Reilly, R. C.
(1995). Why there are complementary learning sys-
tems in the hippocampus and neocortex: insights
from the successes and failures of connectionist mod-
els of learning and memory. Psychological review,
O’Reilly, R. C. and Rudy, J. W. (2001). Conjunctive Rep-
resentations in Learning and Memory: Principles of
Cortical and Hippocampal Function. Psychological
Review, 108(2):311–345.
Pauli, W. M., Hazy, T. E., and O’Reilly, R. C. (2011).
Expectancy, Ambiguity, and Behavioral Flexibility:
Separable and Complementary Roles of the Orbital
Frontal Cortex and Amygdala in Processing Reward
Expectancies. Journal of Cognitive Neuroscience,
Paz, R., Bauer, E. P., and Par
e, D. (2009). Measuring cor-
relations and interactions among four simultaneously
recorded brain regions during learning. Journal of
neurophysiology, 101(5):2507–2515.
Pearce, J. M. and Hall, G. (1980). A model for Pavlovian
learning: variations in the effectiveness of conditioned
but not of unconditioned stimuli. Psychological re-
view, 87(6).
Rescorla, R. and Wagner, A. (1972). A theory of pavlovian
conditioning: Variations in the effectiveness of rein-
forcement and nonreinforcement. In Classical Con-
ditioning II: Current Research and Theory, pages 64–
99. Appleton Century Crofts.
Rousselet, G., Thorpe, S., and Fabre-Thorpe, M. (2004).
How parallel is visual processing in the ventral path?
TRENDS in Cognitive Sciences, 8(8):363–370.
Schmajuk, N. and DiCarlo, J. (1992). Stimulus configura-
tion, classical conditioning and the hippocampus. Psy-
chological Review, 99:268–305.
Sperduti, A., Sperduti, R., and Starita, A. (1997). Super-
vised neural networks for the classification of struc-
tures. IEEE Transactions on Neural Networks, 8:714–
Sun, R. and Alexandre, F., editors (1997). Connectionist
- Symbolic Integration: from Unified to Hybrid Ap-
proaches. Lawrence Erlbaum Associates.
Yu, A. J. and Dayan, P. (2005). Uncertainty, neuromodula-
tion, and attention. Neuron, 46(4):681–692.