arXiv:1805.00777v1  [cs.RO]  2 May 2018
Potentials and Limitations of Deep Neural
Networks for Cognitive Robots
Doreen Jirak and Stefan Wermter
Knowledge Technology, Department of Informatics, University of Hamburg, Germany
Abstract—Although Deep Neural Networks reached remark-
able performance on several benchmarks and even gained scien-
tific publicity, they are not able to address the concept of cognition
as a whole. In this paper, we argue that those architectures
are potentially interesting for cognitive robots regarding their
perceptual representation power for audio and vision data.
We identify crucial settings for cognitive robotics where deep
neural networks have as yet only contributed little compared
to the challenges in this area. Finally, we highlight the rather
unexplored area of Reservoir Computing for sequence learning.
This new paradigm of learning recurrent neural networks in a
fast and robust way qualifies to be an integral part of cognitive
robots and may inspire novel developments.
I. INTRODUCTION
With the advent of both data availability and accelerated
GPU computations, deep neural network architectures (DNN)
have flourished over the past years and showed tremen-
dous success on several benchmarks for computer vision,
and audio and speech. However, DNN models demand high
training times for sensitive parameter tuning and network-
specific optimization like regularization techniques, which is
a rather engineering approach. Although DNNs are motivated
by neurobiological principles observed in the visual cortex, the
complete picture of the computations boils down to learning
filters detecting specific frequency characteristics like edges
in images or voice pitch in audio data. As the learning is
carried out in a supervised manner, the use of a huge amount
of labeled data has a crucial impact on high accuracy and
other evaluation measures. The area of cognitive robotics
is highly interdisciplinary and advances in sensor and robot
technology allow the in- tegration of human-like capabilities.
The aim of cognitive skills in robots is motivated by their
increased integration in human’s everyday life, whether as
a companion in human- robot interaction (HRI) or as an
autonomous system in health care and industry. DNNs can
only address part of what actually cognitive systems need
and which challenges they face. The focus on supervised
benchmarking shifts research on the multiple, interconnected
facets of cognitive robots to a rather engineered approach
accompanied by constraints on the data and the environment.
Critical investigations on DNNs also reveal that due to their
high dependence on apriori knowledge, they are easily fooled
[1] and fail to cope with, e.g., real-world speech data from
children. Recent advances of (deep) unsupervised networks
closely related to neural information principles like sparse
autoencoders or self organizing maps may stimulate again
joining state-of-the-art algorithms successfully applied to robot
scenarios with deep architectures. In the present paper, we
will discuss the pros and cons of DNNs in cognitive robotics.
We highlight their usefulness but also limitations, discussing
also the potential of Reservoir Computing in the context of
cognitive robots.
II. LAB CONDITIONS VS. THE REAL WORLD
A challenging aspect of cognitive systems and specifically
cognitive robots is the high parallelization of tasks with a vast
amount of incoming stimuli which need to be filtered to trigger
a suitable (re)action. We are able to focus our attention onto a
specific object or sound but we can also be easily distracted as
is observable in the ventriloquist effect, where humans’ sound
perception is influenced by a doll being an additional visual
stimulus. Researchers thus developed algorithms to capture
the most salient cues in the scene but either on datasets [2]
or under very constrained lab conditions [3]. Although imple-
mentations using deep networks showed promising results in
those limited cases, DNN will have problems with changing
conditions in real-world environments due to retraining issues
inherent in closed set classification system. A recent robotic
experiment on indoor exploration proposed a framework based
on DNN to exploit the robot capabilities to learn the unknown
terrain in a human-like fashion using the turtlebot platform
[4]. For the robot vision, a Kinect device was mounted
delivering depth information. Although one may argue that
Kinect sensors are error-prone and question their reliability in
a robust image capturing, the availability of the authors’ code
enable other researchers to explore further the advantages and
pitfalls of the approach, identifying further problems and pos-
sible solutions enriching discussions in the cognitive robotic
community for this important task. Another important factor is
that learning in humans is not subject to one specific task for a
specific point in time and space but is continuous throughout
life. Although DNNs produce stable feature representations
which may code for more abstract concepts similar to neural
codes along the brain hierarchy, yet they are less flexible and
do not allow online adaptation nor any memory mechanisms.
A new trend, how- ever, to avoid retraining is using transfer
learning: weights of a deep network successfully trained on
a set of e.g. images are transfered to solve a similar task. A
fruitful combination for robot applications could be to join
those strategies with open set classification algorithms like
the growing-when-required networks (GWR, [5]), which allow
insertion and deletion of labels. The benefit of a GWR is that it
provides a computational method to the question how actually
learning is guided in the absence of any labels. The robust
representations emerging from a deep neural network can
thus complement the flex- ible classification in environments
beyond the restricted lab conditions.
III. CHILDREN VS. ADULT LEARNING
Developmental psychology is an important research area
to gain an understanding on how we acquire cognitive skills
starting from first imitation tasks to controlled motor acts, from
first babbling to words and sentences to finally reasoning,
consciousness and empathy (theory of mind). In contrast to
deep networks being trained on billions of e.g. images for
object recognition, infants learn rather slow and are provided
only with a very limited excerpt of the world. DNNs tuned
for benchmarking as was carried out during the past years
is very different from developmental learning and the ac-
quisition of cognitive skills. Although their performance on
speech recognition tasks is indubitable, DNN models show low
performance when the speech resources are getting noisy (e.g.
repetitions) or are from children, since their language differs
from adults. In addition, the way children learn is driven by
high intrinsic motivation [6], curiosity [7] and their physical
interaction with the world (embodiment). The latter was shown
to be essential in exploring affordances for the associative
learning between objects and grasp types [8], keeping also
into account parental scaffolding [9]. The link between motor
activation in the brain and language acquisition speaks in favor
of embodied learning which demands to ground experiments
in real robots (see e.g. [10]) rather than a set of labeled data.
Until now, these important aspects to acquire cognitive skills
are absent in learning systems based on DNNs which may
be explained by their difficulty in interpretation. A literature
research on this topic revealed that only a few trends emerged
towards closing this gap. One approach used a generative
deep neural network based on the time course of learning
object categories in children [11]. Essentially, the authors
demonstrated that despite using linear neurons, their model
was able to capture the refinement of semantic labels and with
an additional analysis on the temporal dynamics of gradient-
based learning, that this differentiation emerges naturally in
hierarchical networks. Another recent paper [12] identified a
gap between de- velopmental learning and deep neural net-
works along the hierarchical processing of object recognition
on the perceptual level for the acquisition of action knowledge
on the cognitive level where affordances play a crucial role.
In the light of embodied cognition, affordances are grounded
in sensorimotor experiences. The authors [12] argued in favor
of unsupervised learning techniques as labels of sensorimotor
experiences do not exist. Though deep networks do not allow
for flexible, online adaptation of e.g. grasping when affor-
dances may change (think about the change of perspective of a
mug handle), the main argument integrating deep unsupervised
models is that they can serve as an enhancement strategy for
learned sensorimotor experience as in humans during sleep.
In conclusion, although still other developmental learning
mechanisms as the mentioned curiosity are missing and can
not be captured by DNNs, they might be a useful tool for
representational learning, encouraging a novel research area
called ”deep developmental learning” [12]. Regarding the
interpretability issues in DNN, a deep net- work in a one-
shot word learning scenario was employed and the results
investigated by means of methods from cognitive psychology,
namely the ”cognitive bias” which allows children to eliminate
word hypotheses from a vast possible word space [13]. The
network evaluation, which used training on the popular Ima-
geNet database, underpinned the “shape bias” hypothesis, i.e.
that humans tend to label objects similar in shape as one entity.
Joining DNNs with cognitive psychology may open the way
to further investigate the positive as well as negative aspects of
biases in human learning, complementing behavioral studies
which found the basis for many cognitive robotic scenarios.
IV. RESERVOIR COMPUTING FOR SEQUENTIAL LEARNING
IN ROBOT SCENARIOS
Learning sequential tasks is inevitably necessary for the
development of cognitive systems, may it be navigation,
planning or communication with language and gestures. Deep
neural architectures do not possess an explicit time resolution
to capture time correlation nor timescales inherent in sentences
and actions. A popular technique to overcome the missing
temporal link in, e.g., CNNs is to use 3D convolution kernels
for image stacks representing videos. However, the benefit of
such kernels on the performance compared to the standard
2D kernel is not yet satisfiable answered [14]. Due to the
inherent correlations in sequential tasks, Re- current Neural
Networks (RNN) are a widely used tool as they provide local
network memory and are thus able to compute different time
aspects (e.g. continuous-time RNN). RNNs in their diverse
implementations were successfully used in cognitive robotic
scenarios (see e.g. [15], [16]). Due to the error computations
propagated through the whole network using gradient descent,
RNN training suffers from vanishing or exploding gradients,
trapping into local minima, and network bifurcations. To
overcome these issues, RNNs need a lot of retraining and are
thus computationally demanding. Alternatively, the Reservoir
Computing (RC) paradigm in- troduced a simplified training
by the conceptual separation of a high-dimensional reservoir
of randomly connected neurons providing rich input repre-
sentations which are read out by simple linear models and
thus updating the network weights becomes obsolete. Popular
implementations are the Liquid State Machines [17] (LSM)
and Echo State Networks [18] (ESN) the latter demonstrating
competitive performance in predictions tasks for (chaotic)
timeseries [19]. Despite the reduced computational effort, only
a few applications for robots were developed until today.
Some highlights in the literature using the RC principles
include navigation [20], [21] and language acquisition [22].
Especially the latter research shows that RC is a promis-
ing framework combining principles from neuroscience and
developmental psychology for the highly complex cognitive
task of learning and understanding language. Also, only little
is known about the potential of RC algorithms for visual
tasks. As images and videos are highly complex in structure
and introduce varying lighting conditions and perspectives, a
solution combining the robust feature capabilities provided by
DNNs with ESNs capturing the inherent temporal structure
showed good performance on a gesture recognition task for
a set of command gestures [23]. The coupling of DNNs with
RC was also highlighted recently regarding future research
directions [24]. Methods based on RC can also be exploited
for robot control [25] and locomotion [26] but coupling RC
with more complex robot scenarios is still in its infancy.
We believe, that the area of cognitive robotics would sub-
stantially benefit from further research on the potentials of RC
frameworks complementing other computational methods like
DNNs for robot perception or even substituting existing RNN
methods exploiting the simplified yet robust training. Also,
analysis of ESNs regarding their ”edge of stability” would
unify hypotheses on cognition underlying chaotic computa-
tions with performance maximization [27], which can lead to
effective network design.
V. CONCLUSION
The argument in favor of cognitive systems is diametral to
the current benchmarking and competition of accuracies using
deep neural networks. Although those network architectures
are beneficial for perceptual or supervised learning involved
in the acquisition of cognitive skills, we highlighted some
limitations and suggest alternative computations for future ap-
plications. This brief communication focused on recent neural
network approaches and Reservoir Computing to stimulate
discussions on the integration of those methods into cognitive
robots. Our intention was not to exclude other important
methods like probabilistic models or dynamic neural fields. We
rather argue, that advances in technology in humans everyday
life demand tighter coupling in the community regarding
current progress in the neural network and machine learning
area with cognitive models proven to be successful in the
development of cognitive robots.
REFERENCES
[1] A. Nguyen, J. Yosinski, and J. Clune, “Deep neural networks are easily
fooled: High confidence predictions for unrecognizable images,” in The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
June 2015.
[2] K. Xu, J. L. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhutdinov, R. S.
Zemel, and Y. Bengio, “Show, attend and tell: Neural image caption
generation with visual attention,” in Proceedings of the 32nd Interna-
tional Conference on Machine Learning - Volume 37, ser. ICML’15.
JMLR.org, 2015, pp. 2048–2057.
[3] P. Barros, G. I. Parisi, and S. Wermter, “Emotion-modulated attention
improves expression recognition: A deep learning model,” Neurocom-
puting, vol. 253, pp. 104–114, Mar 2017.
[4] L. Tai, S. Li, and M. Liu, “Autonomous exploration of mobile robots
through deep neural networks,” International Journal of Advanced
Robotic Systems, vol. 14, no. 4, 2017.
[5] S. Marsland, J. Shapiro, and U. Nehmzow, “A self-organising network
that grows when required,” Neural Networks, vol. 15, no. 8-9, pp. 1041–
1058, 2002, [There is a reference to a similar piece of work omitted from
this paper. It is: A. Baraldi and P. Blonda, ‘A survey of fuzzy clustering
algorithms for pattern recognition: Part II’, IEEE Trans. Systems, Man
and Cybernetics - Part B: Cybernetics, vol. 29, no. 6, pp. 786-801, Dec.
1999.].
[6] P.-Y. Oudeyer and F. Kaplan, “What is intrinsic motivation? a typology
of computational approaches,” Frontiers in Neurorobotics, vol. 1, p. 6,
2009.
[7] J. Gottlieb, P.-Y. Y. Oudeyer, M. Lopes, and A. Baranes, “Information-
seeking, curiosity, and attention: computational and neural mechanisms.”
Trends in cognitive sciences, vol. 17, no. 11, pp. 585–593, Nov. 2013.
[8] E. Oztop, N. S. Bradley, and M. A. Arbib, “Infant grasp learning: a
computational model,” Experimental Brain Research, vol. 158, no. 4,
pp. 480–503, Oct 2004.
[9] E. Ugur, Y. Nagai, H. Celikkanat, and E. Oztop, “Parental scaffolding as
a bootstrapping mechanism for learning grasp affordances and imitation
skills,” Robotica, vol. 33, no. 5, p. 11631180, 2015.
[10] D. Marocco, A. Cangelosi, K. Fischer, and T. Belpaeme, “Grounding
action words in the sensorimotor interaction with the world: experiments
with a simulated icub humanoid robot,” Frontiers in Neurorobotics,
vol. 4, 2010.
[11] A. M. Saxe, J. L. McClelland, and S. Ganguli, “Learning hierarchical
category structure in deep neural networks,” in Proceedings of the 35th
annual meeting of the Cognitive Science Society, 2013, pp. 1271–1276.
[12] O. Sigaud and A. Droniou, “Towards deep developmental learning,”
IEEE Transactions on Cognitive and Developmental Systems, vol. 8,
no. 2, pp. 99–114, June 2016.
[13] S. Ritter, D. G. T. Barrett, A. Santoro, and M. M. Botvinick, “Cognitive
psychology for deep neural networks: A shape bias case study,” 2017.
[14] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and
L. Fei-Fei, “Large-scale video classification with convolutional neural
networks,” in 2014 IEEE Conference on Computer Vision and Pattern
Recognition, June 2014, pp. 1725–1732.
[15] M. Oubbati, B. Kord, P. Koprinkova-Hristova, and G. Palm, “Learning
of embodied interaction dynamics with recurrent neural networks: some
exploratory experiments,” Journal of Neural Engineering, vol. 11, no. 2,
2014.
[16] S. Murata, Y. Yamashita, H. Arie, T. Ogata, S. Sugano, and J. Tani,
“Learning to perceive the world as probabilistic or deterministic via in-
teraction with others: A neuro-robotics experiment,” IEEE Transactions
on Neural Networks and Learning Systems, vol. 28, no. 4, pp. 830–848,
April 2017.
[17] W. Maass, T. Natschl¨ager, and H. Markram, “Real-time computing
without stable states: A new framework for neural computation based
on perturbations,” Neural Computation, vol. 14, no. 11, pp. 2531–2560,
11 2002.
[18] H. Jaeger, “The echo state approach to analysing and training recurrent
neural networks - with an erratum note,” German National Research
Center for Information Technology, Tech. Rep., 2001.
[19] M. Lukoˇseviˇcius, “A practical guide to applying echo state networks,”
in Neural Networks: Tricks of the Trade: Second Edition, G. Montavon,
G. B. Orr, and K.-R. M¨uller, Eds.
Springer Berlin Heidelberg, 2012,
pp. 659–686.
[20] S. Dasgupta, F. W¨org¨otter, and P. Manoonpong, “Information dynamics
based self-adaptive reservoir for delay temporal memory tasks,” Evolving
Systems, vol. 4, no. 4, pp. 235–249, 2013.
[21] E. A. Antonelo and B. Schrauwen, “On learning navigation behaviors
for small mobile robots with reservoir computing architectures,” IEEE
Transactions on Neural Networks and Learning Systems, vol. 26, no. 4,
pp. 763–780, April 2015.
[22] X. Hinaut, M. Petit, G. Pointeau, and P. F. Dominey, “Exploring the
acquisition and production of grammatical constructions through human-
robot interaction with echo state networks,” Frontiers in Neurorobotics,
vol. 8, p. 16, 2014.
[23] D. Jirak, P. Barros, and S. Wermter, “Dynamic gesture recognition using
echo state networks,” in Proceedings of the European Symposium of
Artificial Neural Networks and Machine Learning, 2015, pp. 475–480.
[24] A. Goudarzi and C. Teuscher, “Reservoir computing: Quo vadis?” in
Proceedings of the 3rd ACM International Conference on Nanoscale
Computing and Communication, 2016, pp. 13:1–13:6.
[25] A. Polydoros, L. Nalpantidis, and V. Krger, “Advantages and limitations
of reservoir computing on model learning for robot control,” 2015.
[26] F. Wyffels and B. Schrauwen, “Design of a central pattern generator us-
ing reservoir computing for learning human motion,” in 2009 Advanced
Technologies for Enhanced Quality of Life, July 2009, pp. 118–122.
[27] F. M. Bianchi, L. Livi, and C. Alippi, “Investigating echo-state networks
dynamics by means of recurrence analysis,” IEEE Transactions on
Neural Networks and Learning Systems, vol. PP, no. 99, pp. 1–13, 2016.