=Paper=
{{Paper
|id=Vol-3227/Bouhadar-et-al.PP7
|storemode=property
|title=Prediction: An Algorithmic Principle Meeting Neuroscience and Machine Learning Halfway
|pdfUrl=https://ceur-ws.org/Vol-3227/Bouhadar-et-al.PP7.pdf
|volume=Vol-3227
|authors=Younes Bouhadjar,Caterina Moruzzi,Melika Payvand
|dblpUrl=https://dblp.org/rec/conf/hlc/BouhadjarMP22
}}
==Prediction: An Algorithmic Principle Meeting Neuroscience and Machine Learning Halfway==
https://ceur-ws.org/Vol-3227/Bouhadar-et-al.PP7.pdf
Prediction: An Algorithmic Principle Meeting
Neuroscience and Machine Learning Halfway
Younes Bouhadjar1,† , Caterina Moruzzi2,*,† and Melika Payvand3,†
1
Institute of Neuroscience and Medicine (INM-6), & Institute for Advanced Simulation (IAS-6), & JARA BRAIN Institute
Structure-Function Relationships (INM-10), & Peter Grünberg Institute (PGI-7,10), Jülich Research Centre and JARA,
Jülich, Germany & RWTH Aachen University, Aachen, Germany
2
Department of Philosophy, University of Konstanz, Universitätsstraße 10, 78464 Konstanz, Germany
3
Institute of Neuroinformatics, University of Zurich and ETH Zurich, Winterthurerstrasse 190, 8057 Zurich, Switzerland
Abstract
In this paper, we support the relevance of the collaboration and mutual inspiration between research in
Artificial Intelligence and neuroscience to create truly intelligent and efficient systems. In contrast to the
traditional top-down and bottom-up strategies designed to study and emulate the brain, we propose an
alternative approach where these two strategies are met halfway, defining a set of algorithmic principles.
We present prediction as a core algorithmic principle and advocate for applying the same approach
to identify other neural principles which can constitute core mechanisms of new Machine Learning
frameworks.
Keywords
Prediction, Neuroscience, Reasoning, Algorithmic Principles, Computation, Bottom-up, Top-down
1. Introduction
The impressive accomplishments by Machine Learning (ML) research in the last decade sub-
stantiate the argument that, given enough computing power and scale, almost any kind of task
can be successfully achieved with just statistical correlation of data [1, 2, 3, 4]. Yet, leveraging of
this type of computation does not seem to be enough in order to solve two hurdles that research
in Artificial Intelligence (AI) still has to confront: (i) Functional: the current ML models lack
the ability to achieve abstraction and generalization, e.g., classification of images in unfamiliar
environments, or making predictions on out-of-distribution data [5, 6, 7]; (ii) Technological: ML
models typically require a lot of data to train and consume a substantial amount of energy [8].
In this paper, we support the claim that by exploiting the knowledge we have about the brain,
as a proof of the existence of an efficient intelligent machine, we can provide insight to ML
International Workshop on Human-Like Computing - International Joint Conference on Learning & Reasoning 2022
*
Corresponding author.
†
These authors contributed equally.
$ y.bouhadjar@fz-juelich.de (Y. Bouhadjar); caterina.moruzzi@uni-konstanz.de (C. Moruzzi); melika@ini.uzh.ch
(M. Payvand)
https://younesbouhadjar.github.io (Y. Bouhadjar); https://caterinamoruzzi.weebly.com (C. Moruzzi);
https://services.ini.uzh.ch/people/melika (M. Payvand)
0000-0003-4367-8236 (Y. Bouhadjar); 0000-0002-9728-3873 (C. Moruzzi); 0000-0001-5400-067X (M. Payvand)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
computational models for solving these problems.
Functionally, the shift from behaviorism to cognitivism and the study of mental processes
was aided and inspired by the emerging field of computer science. Many researchers shared
the vision to undertake the project of building Thinking Machines. Conversely, the inspiration
for many discoveries in the AI research field comes from the observation of underlying com-
putational mechanisms in the animal and human brain [9, 10, 11, 12]. The first computational
model of an artificial neuron was developed in 1943 by McCulloch and Pitts, and it was based
on their understanding of the structure and functionality of a biological neuron [13, 14]. More
than fifty years later, the creation of deep neural networks was inspired by the hierarchical
structure of the brain [15]. The observation of reward prediction brain processes was the basis
for the development of reinforcement learning techniques [16]. Technologically, many of
the solutions used by the current AI hardware for more efficient implementations are inspired
by and/or compatible with the solutions found by the biological counterparts. For example,
pruning the neural network parameters [17], reducing parameter precision [18], reducing the
precision of the activation functions [19], and exploiting the sparsity of the activations [20] and
network parameters [21].
The collaboration and mutual inspiration between AI and brain studies can bear fruit yet
another time, by helping the AI research to tackle the functional and technological problems it
is facing. To achieve this aim, in section 2 we propose an alternative strategy to more traditional
top-down and bottom-up approaches to the emulation of the human brain. We propose to meet
these approaches halfway and start from an algorithmic principles level connecting the two.
In section 3, we identify prediction - an important mechanism which had already received
attention in the AI and cognitive field - as one of these principles, crucial in order to endow
machines with more human-like capacities.
2. Algorithmic Principles: A Middle Layer
Figure 1 illustrates the articulation of mental processes and functions which stem from a mind,
i.e., a functional entity that supports intelligent behavior [22]. This functional entity needs
to have a substratum, which can either be biological or artificial. In the quest for building
human-level intelligence, the most straightforward strategy would seem that of emulating the
functioning of the organic brain. In performing this attempt, two opposite approaches are
traditionally followed: top-down or bottom-up [23].
Top-down strategies start from the analysis of the brain’s behavior and its higher cognitive
functionalities, e.g. learning, reasoning, and memory. Bottom-up approaches attempt instead
at understanding and replicating the organizing principles of the brain, such as modularity,
spatial awareness, self-organization, and temporal coding, or, at an even lower abstraction level,
the biological structure and physiological features of the brain. Both approaches may face
criticism, though. Top-down approaches are subject to interpretation biases, as framing the
behavior of the system-under-analysis consequently influences the choice of strategy followed
for reproducing it [24]. On the other hand, bottom-up approaches are comparatively more
resource-expensive, and lack high-level functional guidance.
As an alternative strategy to top-down and bottom-up approaches towards the emulation of
Figure 1: Articulation of Mental Processes and Functions. In the diagram, prediction is presented as one
of the algorithmic principles that bridge cognitive faculties and organizing principles.
the human brain, we propose to meet these approaches halfway, and start from an algorithmic
principles level connecting the two. The intermediate layer is defined by using insights from
both top-down and bottom-up studies, and it expresses the higher-order cognitive functions
of the brain, while simultaneously explaining the biological functions and neural system’s
organizing principles. Our approach is reminiscent of Marr’s three levels [23], but advocates
for a new interpretation that can be accessible to practitioners in different fields.
At this stage, our proposal is speculative in nature: it is intended as an invitation to engage in
a discussion about how an alternative third way, between top-down and bottom-up perspectives
on learning and cognition, could facilitate collaborations between different fields of research
aimed at endowing machines with human reasoning skills. We leave to further research a
detailed consideration of how this approach would be implemented in practice.
Focusing on the algorithmic principles layer, as suggested by this paper, will help to identify
how the structure of the brain might give rise to its cognitive functionality. It helps neuro-
scientists to abstract away from the many details of biology to find the organizing principles
of intelligence, and provides insights to ML researchers to move towards more human-like
intelligence. In the remaining part of the paper, we focus on prediction, as an exemplary
mechanism that governs the interaction between top-down expectations and bottom-up sensory
input and which can help neuroscience and ML research meet halfway.
3. Prediction as a Primary Computation Performed by the Brain
We define prediction as an inference of future states learned by means of prior experiences, which
can result in the fast detection and adaptation to changes in the environment. Prediction has
been identified by many as a requirement for planning, decision-making, motor and cognitive
control, counterfactual reasoning, and the improvement of behavior on the basis of experiences
[25, 26, 27, 28, 29, 30, 24]. To discuss this principle with a simple example, let us say you are
thirsty. You get up from the couch, enter the kitchen and open the fridge, because you predict
we will find a fresh bottle of water there. The brain predicts and prepares actions on the basis
of prior experiences: your action of opening the fridge, and the cognitive and motor control
aspects responsible for this action, were primed by the expectation to find the water bottle in
there.
The neuronal structure responsible for this functionality in the animal cortex has been
extensively studied. The cortex consists of a repeating modular structure with hierarchical
connectivity of a population of neurons. The majority of these neurons are pyramidal neurons
characterized by a distinct multi-compartment structure receiving inputs from different direc-
tions. The bottom-up sensory information is received in compartments closer to the neuron
body (known as the basal dendrites), and the top-down feedback information is received in
compartments further away from the body (known as the apical dendrites) [31]. The spatio-
temporal features of the sensory input is detected by the coincident activation of the neighboring
synapses in the dendrites. In case of sufficient activation, a dendritic action potential (dAP) is
generated [32, 33, 34]. The activation of a dAP results in a long-lasting depolarization of the
neuron body, or soma. The depolarized state of these neurons make them more likely to fire as
a response to an input.
Recent studies show that the dAPs can encode contextual information representing prior
experiences or top-down expectations [35, 36]. The contextual information allows pyramidal
neurons to predict the most likely next outcome [37, 38] or compare top-down expectations
from higher-level areas with bottom-up sensory signals [39]. In case of a prediction error, the
cells fire in a non-sparse manner, signaling a “mismatch”. This activity derives the learning and
in turn reduces the error [40].
Pyramidal neurons are thus powerful computational units that not only passively sum up
incoming input, but can also signal predictions as a result of the dendritic activity. As these
neurons are grouped within different sensory and hierarchical areas, they allow for simultaneous
predictions of different features, which can be of low-level details, such as predicting sensory
information, or of high-level details, such as predicting concepts or thought processes. As you
open the fridge to grab the bottle of water, the visual areas of the cortex predict the shape
and color of the bottle, the somatosensory areas predict tactile sensations, and other areas
predict more complex features such as the weight of the bottle. These predictions are unified by
higher-level areas into an abstract concept.
In the light of top-down studies, prediction emerged as a building block of cognition. Using
this knowledge has helped to guide bottom-up studies toward an interpretation of neuronal
mechanisms that were otherwise obscure. This, in turn, solidifies the concept of prediction and
presents it as an algorithmic principle implemented by the brain.
We advocate for applying the same approach to identify other neural principles. These
principles can constitute core mechanisms of new machine learning frameworks, bridging the
gap in the performance between machines and brains. In the quest toward the goal of emulating
human-level intelligence in artificial systems, what is needed is a shared platform where
methodologies of different disciplines can meet and cross-pollinate each other. Algorithmic
principles, we argue, may constitute such a platform.
Acknowledgments
The idea for this paper was developed during the 2022 edition of “The CapoCaccia Workshops
toward Neuromorphic Intelligence". We would like to thank the organizers of the workshop,
as well as the Office of Naval Research for supporting the participation to it through the ONR
Global grant (grant number: N62909-22-1-2029).
References
[1] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan,
P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learners, Advances in
neural information processing systems 33 (2020) 1877–1901.
[2] A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, M. Chen, Hierarchical text-conditional image
generation with clip latents, arXiv preprint arXiv:2204.06125 (2022).
[3] S. Reed, K. Zolna, E. Parisotto, S. G. Colmenarejo, A. Novikov, G. Barth-Maron, M. Gimenez,
Y. Sulsky, J. Kay, J. T. Springenberg, et al., A generalist agent, arXiv preprint
arXiv:2205.06175 (2022).
[4] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre,
D. Kumaran, T. Graepel, et al., Mastering chess and shogi by self-play with a general
reinforcement learning algorithm, arXiv preprint arXiv:1712.01815 (2017).
[5] K. Kawaguchi, L. P. Kaelbling, Y. Bengio, Generalization in deep learning, arXiv preprint
arXiv:1710.05468 (2017).
[6] J. Miüller, R. Weibel, J.-P. Lagrange, F. Salgé, Generalization: state of the art and issues,
GIS and Generalization (2020) 3–17.
[7] C. Moruzzi, Artificial creativity and general intelligence, Journal of Science and Technology
of the Arts 12 (2020) 84–99.
[8] A. Mehonic, A. J. Kenyon, Brain-inspired computing needs a master plan, Nature 604
(2022) 255–260.
[9] J. Fan, L. Fang, J. Wu, Y. Guo, Q. Dai, From brain science to artificial intelligence, Engi-
neering 6 (2020) 248–252.
[10] D. Hassabis, D. Kumaran, C. Summerfield, M. Botvinick, Neuroscience-inspired artificial
intelligence, Neuron 95 (2017) 245–258.
[11] N. Savage, Marriage of mind and machine, Nature 571 (2019) S15–S17.
[12] S. Ullman, Using neuroscience to develop artificial intelligence, Science 363 (2019) 692–693.
[13] S. Chakraverty, D. M. Sahoo, N. R. Mahato, Mcculloch–pitts neural network model, in:
Concepts of Soft Computing, Springer, 2019, pp. 167–173.
[14] F. Rosenblatt, The perceptron: a probabilistic model for information storage and organiza-
tion in the brain., Psychological review 65 (1958) 386.
[15] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, nature 521 (2015) 436–444.
[16] W. Dabney, Z. Kurth-Nelson, N. Uchida, C. K. Starkweather, D. Hassabis, R. Munos,
M. Botvinick, A distributional code for value in dopamine-based reinforcement learning,
Nature 577 (2020) 671–675.
[17] S. Han, H. Mao, W. J. Dally, Deep compression: Compressing deep neural networks with
pruning, trained quantization and Huffman coding, arXiv preprint arXiv:1510.00149 (2015).
[18] Edge TPU (2019). URL: https://cloud.google.com/edge-tpu/.
[19] S. Gupta, A. Agrawal, K. Gopalakrishnan, P. Narayanan, Deep learning with limited
numerical precision, in: International conference on machine learning, PMLR, 2015, pp.
1737–1746.
[20] C. Gao, A. Rios-Navarro, X. Chen, S.-C. Liu, T. Delbruck, EdgeDRNN: Recurrent neural
network accelerator for edge inference, IEEE Journal on Emerging and Selected Topics in
Circuits and Systems 10 (2020) 419–432.
[21] C. Gao, T. Delbruck, S.-C. Liu, Spartus: A 9.4 TOp/s FPGA-based LSTM accelerator
exploiting spatio-temporal sparsity, arXiv preprint arXiv:2108.02297 (2021).
[22] J. E. Laird, C. Lebiere, P. S. Rosenbloom, A standard model of the mind: Toward a common
computational framework across artificial intelligence, cognitive science, neuroscience,
and robotics, Ai Magazine 38 (2017) 13–26.
[23] D. Marr, Vision: A computational investigation into the human representation and pro-
cessing of visual information, MIT press, 2010.
[24] M. György Buzsáki, The brain from inside out, Oxford University Press, 2019.
[25] A. Bubic, D. Y. Von Cramon, R. I. Schubotz, Prediction, cognition and the brain, Frontiers
in human neuroscience 4 (2010) 25.
[26] A. Clark, Whatever next? predictive brains, situated agents, and the future of cognitive
science, Behavioral and brain sciences 36 (2013) 181–204.
[27] K. Friston, The free-energy principle: a unified brain theory?, Nature reviews neuroscience
11 (2010) 127–138.
[28] J. Hawkins, S. Blakeslee, On intelligence: How a new understanding of the brain will lead
to the creation of truly intelligent machines, Macmillan, 2007.
[29] C. Moruzzi, Climbing the ladder: How agents reach counterfactual thinking, in: ICAART
2022: 14th International Conference on Agents and Artificial Intelligence, 2022, pp. 555–
560.
[30] R. S. Sutton, Learning to predict by the methods of temporal differences, Machine learning
3 (1988) 9–44.
[31] N. Spruston, Pyramidal neurons: dendritic structure and synaptic integration, Nature
Review Neuroscience 9 (2008) 206–221.
[32] G. J. Stuart, N. Spruston, Dendritic integration: 60 years of progress, Nature Neuroscience
18 (2015) 1713–1721.
[33] G. Major, M. Larkum, J. Schiller, Active properties of neocortical pyramidal neuron
dendrites, Annual Review Neuroscience 36 (2013) 1–24.
[34] M. London, M. Häusser, Dendritic computation, Annual Review Neuroscience 28 (2005)
503–532.
[35] N. Takahashi, C. Ebner, J. Sigl-Glöckner, S. Moberg, S. Nierwetberg, M. E. Larkum, Ac-
tive dendritic currents gate descending cortical outputs in perception, Nature Neu-
roscience 23 (2020) 1277–1285. URL: https://doi.org/10.1038/s41593-020-0677-8. doi:10.
1038/s41593-020-0677-8.
[36] C. J. Gillon, J. E. Pina, J. A. Lecoq, R. Ahmed, Y. N. Billeh, S. Caldejon, P. Groblewski,
T. M. Henley, I. Kato, E. Lee, J. Luviano, K. Mace, C. Nayan, T. V. Nguyen, K. North,
J. Perkins, S. Seid, M. T. Valley, A. Williford, Y. Bengio, T. P. Lillicrap, B. A. Richards,
J. Zylberberg, Learning from unexpected events in the neocortical microcircuit (2021).
URL: https://doi.org/10.1101/2021.01.15.426915. doi:10.1101/2021.01.15.426915.
[37] J. Hawkins, S. Ahmad, Why neurons have thousands of synapses, a theory of sequence
memory in neocortex, Frontiers in Neural Circuits 10 (2016) 23. URL: https://doi.org/10.
3389/fncir.2016.00023. doi:10.3389/fncir.2016.00023.
[38] Y. Bouhadjar, D. J. Wouters, M. Diesmann, T. Tetzlaff, Sequence learning, prediction, and
replay in networks of spiking neurons, arXiv preprint arXiv:2111.03456 (2021).
[39] M. Larkum, A cellular mechanism for cortical associations: an organizing principle for the
cerebral cortex, Trends in Neurosciences 36 (2013) 141–151. URL: https://doi.org/10.1016/j.
tins.2012.11.006. doi:10.1016/j.tins.2012.11.006.
[40] A. Payeur, J. Guerguiev, F. Zenke, B. A. Richards, R. Naud, Burst-dependent synap-
tic plasticity can coordinate learning in hierarchical circuits, Nature Neuroscience
24 (2021) 1010–1019. URL: https://doi.org/10.1038/s41593-021-00857-x. doi:10.1038/
s41593-021-00857-x.