Neural Semantic Pointers in Context

Alessio Plebe and Arianna Pavone

Department of Cognitive Science, University of Messina, v. Concezione 8, 98121, Messina, Italy

Keywords:

Context, Neural Model, Semantic Disambiguation, Brain Model.

Abstract:

Resolving linguistic ambiguities is a task frequently called for in human communication. In many cases,

such task cannot be solved without additional information about an associated context, which can be often

captured from the visual scene referred by the sentence. This type of inference is crucial in several aspects of

language, communication in the ﬁrst place, and in the grounding of language in perception. This paper focuses

on the contextual effects of visual scenes on semantics, investigated using neural computational simulation.

Speciﬁcally, here we address the problem of selecting the interpretation of sentences with an ambiguous

prepositional phrase, matching the context provided by visual perception. More formally, provided with a

sentence, admitting two or more candidate resolutions for a prepositional phrase attachment, and an image that

depicts the content of the sentence, it is required to choose the correct resolution depending on the image’s

content. From the neuro-computational point of view, our model is based on Nengo, the implementation of

Neural Engineering Framework (NEF), whose basic semantic component is the so-called Semantic Pointer

Architecture (SPA), a biologically plausible way of representing concepts by dynamic neural assemblies. We

evaluated the ability of our model in resolving linguistic ambiguities on the LAVA (Language and Vision

Ambiguities) dataset, a corpus of sentences with a wide range of ambiguities, associated with visual scenes.

1 INTRODUCTION

In recent years, a number of different disciplines have

begun to investigate the fundamental role that context

plays in different cognitive phenomena. The problem

of context spans from the abstract level of semantics

down to the level of neural representations. It has in-

creasingly been studied also for its role in inﬂuenc-

ing mental concepts and, more speciﬁcally, linguistic

communication has been the area of study that has

traditionally explored these issues. The term context

is not easy to deﬁne: it is something that cannot be

speciﬁed independently of a speciﬁc frame and it may

play quite different roles within alternative research

paradigms. In wider terms, as stated by Goodwin and

Duranti (1992), we can deﬁne the context as a ”frame

that surrounds the event and provides resources for its

appropriate interpretation”. However, in order to ob-

tain a more complete understanding of what context

stands for, it is necessary to investigate how it inter-

acts with cognitive phenomena at three different lev-

els: the linguistic, the cognitive and the neural level

(Plebe and De La Cruz, 2020).

There is a long tradition in linguistics and prag-

matics which invokes context to help account for as-

pects of meaning in language that go beyond the

scope of semantics. The main elements of context

has roots that dates back to the past, and regards the

degree to which truth-functional semantics depends

on context. Gottlob Frege raised the point in his un-

completed 1897 volume Logik, and though he was not

explicitly using the term context, he underlined how

for many expressions, ﬁxing their truth value requires

supplemental information, coming from the circum-

stances, the ”frames”, in which such expressions are

pronounced. The ﬁrst clear elucidation of the depen-

dence of language on context was proposed by Searle

(1978). He takes up Frege’s idea that a word has

meaning only if related to the meaning of the whole

sentence and if its meaning is perceived by both in-

terlocutors, speaker and listener. At a cognitive level

the issue regards concepts and the degree to which

they are dependent on context. Barsalou (1983) has

been one of the ﬁrst to underline how difﬁcult it is

to conceive them as stable subjective entities, while

it appears more appropriate to think of categories as

dynamically constructed and tailored to speciﬁc con-

texts, or as ad hoc categories. A recent review of stud-

ies of the cognitive perspective on the linguistic issue

of context can be found in (Airenti and Plebe, 2017).

Just as in the strictly linguistic domain, they ﬁnd in the

wider cognitive view a variety of positions, some that

Plebe, A. and Pavone, A.

Neural Semantic Pointers in Context.

DOI: 10.5220/0010145904470454

In Proceedings of the 12th International Joint Conference on Computational Intelligence (IJCCI 2020), pages 447-454

ISBN: 978-989-758-475-6

447

minimize the destabilizing effect context has on con-

cepts, such as that of Machery (2015), or others that

assume a more intermediate position such as. that of

Mazzone and Lalumera (2009), that while acknowl-

edging the fundamental. role context might play in

concepts, sustain that a characterizing stable nucleus

of mental concepts is also a part.

On the other hand, cognitive neuroscience is now

starting to consider in a systematic way how context

interacts with neural responses (Stark et al., 2018).

The way context drives language comprehension de-

pends on the effects of context on the conceptual scaf-

folding of the listener, which in turn, is the result of

his neural responses in combination to context.

The kind of ambiguity addressed in this paper is

the canonical case of structural ambiguity, technically

known as Prepositional Phrase Attachment, where a

sentence includes a prepositional phrase that can be

attached to more than one higher level phrases (Hin-

dle and Rooth, 1993). The attachment resolution is

context dependent, we deal speciﬁcally with the case

when depends on the visual context.

Speciﬁcally, provided with a sentence, admitting

two or more candidate interpretations, and an image

that depicts its content, it is required to choose the

correct interpretation of the sentence depending on

the image’s content. Thus we address the problem of

selecting the interpretation of an ambiguous sentence

matching the content of a given image.

This type of inference is frequently called for in

human communication occurring in a visual environ-

ment, and is crucial for language acquisition, when

much of the linguistic content refers to the visual sur-

roundings of the child (Bates et al., 1995; Bornstein

and R.Cote, 2004).

This kind of task is also fundamental to the prob-

lem of grounding vision in language, by focusing on

phenomena of linguistic ambiguity, which are preva-

lent in language, but typically overlooked when using

language as a medium for expressing understanding

of visual content. Due to such ambiguities, a super-

ﬁcially appropriate description of a visual scene may

in fact not be sufﬁcient for demonstrating a correct

understanding of the relevant visual content.

From the neurocomputational point of view, our

model is based on Nengo (https://www.nengo.ai),

the implementation of Eliasmith’s Neural Engineer-

ing Framework (NEF) (Eliasmith, 2013). The basic

semantic component within NEF is the so-called Se-

mantic Pointer Architecture (SPA) (Thagard, 2011),

which determines how the concepts are represented

as dymanic neural assemblies. The model works by

extracting the three involved entities from the input

sentence and identifying the categories involved.

Early experimental results show that the presented

computational model achieves a reliable ability to dis-

ambiguate sentences.

1.1 A Framework for Neural Semantics

The two main requirements we seek in the identiﬁ-

cation of a suitable neural framework to be adopted

all along this work is the biological plausibility and

the possibility of modeling at a level enough abstract

to deal with full images and with words in sentences.

The two requirements are clearly in stark contrast.

Today the legacy of connectionism has been taken

up by the family of algorithms collected under the

name deep learning. Unlike the former artiﬁcial neu-

ral networks, deep learning models succeeds in highly

complex cognitive tasks, reaching even human-like

performances in some visual tasks (VanRullen, 2017).

However, the level of biological plausibility of deep

learning algorithms is in general even lower than in

connectionism, these models were developed with en-

gineering goals in mind, and exploring cognition is

not in the agenda of this research community (Plebe

and Grasso, 2019). In our model we will also include

a very simple deep learning component, but only for

the low-level analysis of the images. This choice

makes the model simpler, by exploiting the ease of

deep learning model in processing visual stimuli. It

would have been easy to solve also the crucial part

of our problem, the semantic disambiguation, through

deep learning, but this would have been of little value

as a cognitive model.

Currently, the neural framework that can simu-

late the widest range of cognitive tasks, by adopting a

uniﬁed methodology with a reasonable degree of bi-

ological plausibility, is Nengo (Neural ENGineering

Objects) (Eliasmith, 2013). The idea behind Nengo

dates back to 2003, thanks to the former NEF (Neu-

ral Engineering Framework) (Eliasmith and Ander-

son, 2003), which deﬁnes a general methodology for

the construction of large cognitive models, informed

by a number of key neuroscientiﬁc concepts. In brief,

the three main such concepts are the following:

• The Representation Principle: neural representa-

tions are deﬁned by the combination of nonlinear

encoding of spikes over a population of neurons,

and weighted decoding over the same populations

of neurons and over time;

• The Transformation Principle: transformations of

neural representations are functions of the vari-

ables represented by neural populations. Trans-

formations are determined using an alternately

weighted decoding;

NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications

448

• The Dynamic Principle: neural dynamics are

characterized by considering neural representa-

tions as state variables of dynamic systems. Thus,

the dynamics of neuro-biological systems can be

analyzed using control theory.

According to the listed principles, the basic com-

putational object in Nengo is a population of neurons

that collectively can represent a multidimensional en-

tity. The meaningful entity is retrieved from the neu-

ral activation by the equation (Eliasmith, 2013):

~x =

∑

i=1

∑

j=1

−

t−t

i, j

(1)

where N is the number of neurons in the population,

and M is the number of spikes that happen in the time

windows of the computation; t

i, j

is the time when the

i-th neuron in the population has ﬁred for the j-th

time;

is the i-th row of the N ×D decoding matrix

with D the dimension of the entity to be represented;

τ is the time constant of decay of the postsynaptic ac-

tivation. The activity of the neurons in a population

depends from the encoding of their input that can be

multidimensional with a dimension different from D.

A fundamental extension of the general neural

population, ruled by equation (1), is the the Semantic

Pointer Architecture (SPA), used when representing

entities at higher cognitive level, i.e. conceptual and

linguistic. In addition to the encoding and the decod-

ing features, SPA structures allow a number of high

level operations, that may correspond to conceptual

manipulation, with some degree of biological plausi-

bility. The foundation of these conceptual operations

is in the mathematics of holographic representations,

as theorized by Plate (2003). One of the basic opera-

tions is the binding of two SPA, computed by circular

convolution (Eliasmith, 2013):

~x ~~y = F

−1

(F (~x) · F (~y)) (2)

where F is the discrete Fourier transform and F

−1

is its inverse, and · is the element-wise multiplication.

Informally, SPA together with its associated opera-

tions, can be thought as a neural process that com-

presses information in other neural processes to which

it points and into which it can be expanded when

needed, providing shallow meanings through symbol-

like relations to the world and other representations,

and expanding to provide deeper meanings with rela-

tions to perceptual, motor, and emotional information,

support complex syntactic operations. They also help

to control the ﬂow of information through a cognitive

system to accomplish its goals. Thus semantic point-

ers have semantic, syntactic, and pragmatic functions,

just like the symbols in a rule-based system, but with

a highly distributed, probabilistic operation.

2 THE AMBIGUITY TESTBED

The testbed of our study is the LAVA (Language and

Vision Ambiguities) corpus, recently introduced by

Berzak et al. (2015).

Such corpus contains sentences with linguistic

ambiguities that can only be resolved using exter-

nal information. The sentences are paired with short

videos that visualize different interpretations of each

sentence. Such sentences encompass a wide range of

syntactic, semantic and discourse ambiguities, includ-

ing ambiguous prepositional and verb phrase attach-

ments, conjunctions, logical forms, anaphora and el-

lipsis.

Overall, the corpus contains 237 sentences, with

2 to 3 interpretations per sentence, and an average of

3.37 videos that depict visual variations of each sen-

tence interpretation, corresponding to a total of 1679

videos. Each sentence involves two or more entities

in one among four categories (person, bag, telescope

and chair).

In their paper Berzak et al. (2015) also addressed

the problem of selecting the interpretation of an am-

biguous sentence that matches the content of a given

video. In our case also, the road to solve the ambigu-

ities is in pairing sentences with images that visualize

the corresponding scene. In order to simplify the task,

we limited on sentences where exactly three distinct

entities belonging to three distinct categories are in-

volved (among person, bag, telescope and chair).

For instance, given the sentence ”Sam approached

the chair with a bag”, three different categories in-

volved: person (Sam), chair and bag. In addition two

different linguistic interpretations are plausible: the

ﬁrst interpretation assumes that Sam has the bag while

approaching the chair, while the second one assumes

that the bag is on the chair while Sam is approaching.

In addition, in this preliminary work, the ambigu-

ous phrases examined by the system are limited to the

preposition “with”: the system is able to solve the am-

biguity thanks to the given image and therefore to un-

derstand who the proposition refers to.

For example in the following sentence

“Dany approached the chair with a yellow bag”

the system is able to recognize to whom it refers

“with” and speciﬁcally if Dany brings a yellow bag

while approaching the chair or if the bag is on the

chair while Dany is approaching them.

Our corpus contains 81 sentences, with 2 to 3 in-

terpretations per sentence.

Neural Semantic Pointers in Context

449

3 METHOD

Without restricting the general case we can assume

that the world is populated by objects that are grouped

into categories C =

{

,· ··

}

. The small world

of LAVA is populated with a limited number of cate-

gories, whose instances can appear in images, corre-

sponding to scenes viewed by the agent. Images are

matrices I of pixels in two dimensions. Inside an im-

age I it is possible to cut submatrices B

(C)

x,y

with centers

at coordinates x, y, and size suitable to contain objects

of category C. The submatrices B

(C)

x,y

bear a resem-

blance with foveal images captured during saccadic

movements gazing over different portion of the visual

scene, with the purpose of recognizing each single ob-

ject. Unlike natural saccades, in our model the cen-

ters of the detailed views are not driven by top-down

mechanisms, are instead sampled at ﬁxed regular in-

tervals, scanning the whole image:

X =

+ ∆

,· ·· , x

+ N∆

, (3)

Y =



+ ∆

,· · · , y

+ M∆



. (4)

The visual component of the model is made by a set

of deep convolutional neural networks, tuned to rec-

ognize one of the categories C ∈ C . Each network

is a function f

(C)

(·) estimating the probability of an

object of category C to be inside a submatrix B

x,y

(C)

() : R × R → [0..1] (5)

By applying the deep convolutional neural network

(·) to all the submatrices of an image, a vector ~p

(C)

of probabilities to ﬁnd an object of category C in the

discrete horizontal positions X is constructed. An el-

ement p

(C)

of ~p

(C)

is computed as:

(C)

= max

y∈Y

(C)



(C)

o

(6)

The rationale behind equation (6) is that in an interior

environment the displacement of objects – therefore

their spatial relationship – appears mainly in the hori-

zontal dimension of the retinal projection of the scene.

It is therefore possibly to capture the probabilistic lo-

cations of objects as vectorial representations corre-

sponding to scanning the scene horizontally along X .

We can now compose equation (6) into a family of

functions φ

(C)

(·) that, given and image I, return prob-

ability vectors ~p

(C)

() : R × R → [0..1]

(7)

Let us move on the linguistic part of LAVA and

of the stimuli to the model. The full set of sen-

tences in LAVA use words from a closed lexicon L ,

and within this lexicon there are two subsets rele-

vant for our model. One is the lexicon of words

W ⊂ L =

{

,· · ·

}

used to name objects of the

categories in C . In the case of LAVA we can assume

a deterministic reference function:

c(W ) : W → C (8)

associating every word W to a category C. There is

then a smaller lexicon of prepositions, the grammat-

ical category responsible for the contextual ambigui-

ties: P ⊂ L =

{

with,· · ·

}

A sentence in the LAVA is an ordered set S , with

elements S

∈ L , from which a simple preprocessing

extracts three key words:

S →











noun under the head of S

ﬁrst noun possible head of S

second noun possible head of S

(9)

The noun W

is easily found by searching in S

the ﬁrst element S

∈ P , and then searching the ﬁrst

element S

with j > i such that S

∈ W . The other two

nouns W

and W

are the only two possible elements

l,k

∈ W with l 6= j, k 6= j. Let us call W

∈ {W

}

the correct head of the prepositional phrase.

The three key words W

ﬁnd a correspon-

dence in the model in terms of three Nengo SPA

items:

and

. The processed sentence S is

linked with an image I in which the objects of cate-

gories referred by W

, W

and W

are searched:

= φ

(c(W

))

(I), (10)

= φ

(

))

(I), (11)

= φ

(

))

(I). (12)

These vectors, expressing probabilities of locations

of the three categories along the horizontal view of

the agent, are bind to the corresponding Nengo SPA

items, using NEF ~ operator, introduced in equation

(2). We can express the binding in our case as a func-

tion b(·) : R

→ R





~ ~p

, (13)





~ ~p

, (14)





~ ~p

, (15)

and the disambiguate item

∗

is selected as following:

∗

= arg min

V ∈{

}







o

(16)

where ζ





is a measure of similarity between

the two Nengo SPA items

and

. Therefore, the

predicted head of the prepositional phrase W

∗

is the

lexical item associated with

∗

The combined deep convolutional and Nengo SPA

neural processes are sketched in Fig. 1.

NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications

450

Danny approached the chair with a yellow bag

chair

NP PP

NPNPNP

with bag

...

... ...

TensorFlow−Keras component

Nengo−SPA component

Figure 1: Sketch of the neural model.

4 RESULTS

We evaluated the performance of our neural model,

described in the previous section, on the LAVA

dataset. In this section we report our main prelim-

inary results. Each sentence and its associated pic-

ture in the dataset was processed, predicting the lexi-

cal item that most likely is the head of the ambiguous

propositional phrase. We recall that in this prelim-

inary work, the ambiguous phrases examined by the

system are limited to the preposition “with”. The code

we used for constructing the neural model and com-

puting our experimental results is available for down-

load at http:www.github.com/alex-me/nencog.

Here we will present ﬁrst a set of qualitative re-

sults useful for illustrating the neural processes per-

formed by the model, and then the quantitative results

of the disambiguation task.

Fig. 2 illustrates few qualitative results of

the intermediate stage of the process, when the

vectors ~p

(C)

of probabilities to ﬁnd an object of

category C in the discrete horizontal positions X

are computed. Vectors are generated by applying

the deep convolutional neural network f

(·) to

all the submatrices of a given image according to

equation (6). Fig. 2 include the following 6 examples:

LAVA ID code sentence

22-9570-9660 Danny approached the chair with a yellow bag

22-18590-18700 Danny left the chair with a yellow bag

22-22420-22510 Danny left the chair with a green bag

22-54050-54160 Danny approached the chair with a blue telescope

22-55780-55850 Danny approached the chair with a blue telescope

29-24110-24210 Danny looked at the chair with a blue telescope

The vectors of the kind shown in Fig. 2 that are

relevant in the sentence, are then associated with the

three Nengo SPA units

by means of equa-

tions (13), (14), (15). All SPA units are populations of

spiking neurons, which vectors evolve in time follow-

ing equation (1). This evolution is shown in Fig. 3 for

a small number of examples. All plots show the evo-

lution in time of the three vectors related with Nengo

SPA units

. The crucial aspect for the pur-

pose of the disambiguation is that the ﬁnal shape of

the vectors is such that between

and

the most

similar to

will be the SPA unit associated with the

correct word W

. This ﬁnal similarity can be appreci-

ated in the four examples of Fig. 3.

Table 1 presents the quantitative results of the

model over all the processed LAVA sentences.

The total set of sentences has been divided into

two categories, those with W

=bag and those with

=telescope. In the ﬁrst set the possible correct

heads of the prepositional phrase W

can be person

or chair, while in the second set the possibilities are

bag, person, and chair. For each of the sets the ma-

trix of errors is reported, showing the fractions of lex-

ical element that the model has predicted as head of

the prepositional phrase, given the correct head word.

The overall accuracy of the model is good, over 80%,

and slightly lower when W

=telescope.

As in the case of the experiments by Berzak et al.

(2015), the most signiﬁcant source of failures are poor

object detection. Objects in the LAVA corpus are of-

ten rotated and presented at angles which turns out to

be difﬁcult to recognize.

Neural Semantic Pointers in Context

451

00022-9570-9660 00022-18590-18700 00022-22420-22510

00022-54050-54160 00022-55780-55850 00029-24110-24210

Figure 2: The vectors ~p

(C)

of probabilities to ﬁnd an object of category C ∈ { person, bag, telescope, chair } in the

discrete horizontal positions X computed for 6 different images of the LAVA corpus. Vectors are generated by applying the

deep convolutional neural network f

(·) to all the submatrices of a given image according to equation (6). Images are labeled

with their corresponding codes in the LAVA corpus.

00022-9570-9660

=bag, W

=person

00022-22420-22510

=bag, W

=chair

00022-54050-54160

=telescope, W

=chair

Figure 3: Evolution in time of the Nengo SPA neural populations associated with the three key words in the disambiguation

tasks. In the example in the left the noun under the head of the preposition is bag, and the head of the prepositional phrase

is person, it is visible how at the end of the evolution the SPA vector associated with bag become more similar to the vector

associated with person, with respect to the vector associated with chair. Exactly the opposite happen in the example in the

center, where this time the noun under the head of the preposition, bag has chair as head of the prepositional phrase. The

scenes corresponding to the three examples can be seen in Fig. 2.

NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications

452

Table 1: Overall and detailed accuracy obtained by the

model when tested on the LAVA dataset. The results are

grouped for the two possible W

, and for each one the ma-

trix of errors is shown, with the true W

as rows, and the

predicted W

∗

as columns. The overall accuracy is of 0.81.

=bag

\ W

person chair

person .73 .27

chair .00 1.00

accuracy 0.87

=telescope

\ W

bag person chair

bag .25 .75 .00

person .00 .96 .04

chair .07 .20 .73

accuracy 0.77

It turns out moreover that some object classes, like

the telescope and the bag, are much more difﬁcult to

be recognized. It can be observed in Fig. 2 that ob-

jects of the classes bag and telescope are the most

difﬁcult to be recognized due to their small size and

to the fact that hands tend, in most cases, to largely

cover them. Conversely objects of the classes person

and chair are generally well detected and generate a

much more accurate probability vector. We have as-

sessed this source of error by evaluating the pure vi-

sual recognition accuracy, which is of 80% for person

objects, of 79% for chair object, of 67% for the bag

object, and as low as 60% for telescope.

Note that we deliberately avoided to include in

the model a state-of-the-art deep learning model that

would have easily achieved better recognition rates,

but loosing biological plausibility.

Moreover, with our model we have been able to

evaluate the disambiguation performances that takes

into account uncertainty in the visual process. As

seen in the performances shown in Table 1, disam-

biguation is more reliable that the pure visual object

recognition.

5 CONCLUSIONS

We described a biologically plausible neural cognitive

model able to resolve linguistic ambiguities in a sen-

tences by selecting the interpretation of an ambiguous

sentence matching the content of a given image. The

model has been based on Nengo, using SPA for rep-

resenting concepts.

The component of our model dealing with vi-

sual object recognition is based on deep convolutional

networks, which are less biological plausible than

Nengo. This solution is motivated by the marginal

signiﬁcance of object recognition in our objectives,

and by the well known performances of deep convo-

lutional networks.

Our neural model has been evaluated on the subset

of sentences in the LAVA dataset, in which the prepo-

sition with is responsible for the contextual ambigui-

ties.

Our model achieve an ability to resolve linguist

ambiguities of the kind described, in the LAVA

dataset, over 80%. It turns out that the most signif-

icant source of failures is poor object detection, as

in the case of Berzak et al. (2015), however, the ob-

tained disambiguation accuracy is greater that the vi-

sual recognition error.

REFERENCES

Airenti, G. and Plebe, A. (2017). Editorial: Context in com-

munication: A cognitive view. Frontiers in Psychology,

8:115.

Barsalou, L. W. (1983). Ad hoc concepts. Memory and

Cognition, 11:211–217.

Bates, E., Dal, P. S., and Thal, D. (1995). Individual dif-

ferences and their implications for theories of language

development. In Fletcher, P. and Whinney, B. M., edi-

tors, Handbook of child language, pages 96–151. Basil

Blackwell, Oxford (UK).

Berzak, Y., Barbu, A., Harari, D., Katz, B., and Ullman,

S. (2015). Do you see what i mean? visual resolution

of linguistic ambiguities. In Conference on Empirical

Methods in Natural Language Processing, pages 1477–

1487.

Bornstein, M. H. and R.Cote, L. (2004). Cross-linguistic

analysis of vocabulary in young children: Spanish,

dutch, french, hebrew, italian, korean, and american en-

glish. Child Development, 75:1115–1139.

Eliasmith, C. (2013). How to build a brain: a neural ar-

chitecture for biological cognition. Oxford University

Press, Oxford (UK).

Eliasmith, C. and Anderson, C. H. (2003). Neural Engi-

neering Computation, Representation, and Dynamics in

Neurobiological Systems. MIT Press, Cambridge (MA).

Goodwin, C. and Duranti, A. (1992). Rethinking context:

an introduction. In Duranti, A. and Goodwin, C., ed-

itors, Rethinking context: Language as an interactive

phenomenon, pages 1–42, Cambridge (UK). Cambridge

University Press.

Hindle, D. and Rooth, M. (1993). Structural ambiguity and

lexical relations. Cognitive Linguistics, 19:103–120.

Neural Semantic Pointers in Context

453

Machery, E. (2015). By default: Concepts are accessed in

a context-independent manner. In Laurence, S. and Mar-

golis, E., editors, The Conceptual Mind: New Directions

in the Study of Concepts. MIT Press, Cambridge (MA).

Mazzone, M. and Lalumera, E. (2009). Concepts: Stored

or created? Minds and Machines, 20:47–68.

Plate, T. (2003). Holographic reduced representations.

CSLI Publication, Stanford, (CA).

Plebe, A. and De La Cruz, V. M. (2020). Neural representa-

tions in context. In Pennisi, A. and Falzone, A., editors,

The Extended Theory of Cognitive Creativity – Interdis-

ciplinary Approaches to Performativity, pages 285–300.

Springer, Berlin.

Plebe, A. and Grasso, G. (2019). The unbearable shallow

understanding of deep learning. Minds and Machines,

29:515–553.

Searle, J. R. (1978). Literal meaning. Erkenntnis, 13:207–

224.

Stark, S. M., Reagh, Z. M., Yassa, M. A., and Stark, C. E.

(2018). What’s in a context? cautions, limitations, and

potential paths forward. Neuroscience Letters.

Thagard, P. (2011). Cognitive architectures. In Frankish, K.

and Ramsey, W., editors, The Cambridge Handbook of

Cognitive Science, pages 50–70. Cambridge University

Press, Cambridge (UK).

VanRullen, R. (2017). Perception science in the age of deep

neural networks. Frontiers in Psychology, 8:142.

NCTA 2020 - 12th International Conference on Neural Computation Theory and Applications

454