See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/269711306

Teaching Computational Neuroscience

Article in Cognitive Neurodynamics · December 2014

DOI: 10.1007/s11571-015-9340-6. · Source: arXiv

CITATIONS READS

3 87

1 author:

Peter Erdi
Kalamazoo College
93 PUBLICATIONS 984 CITATIONS

SEE PROFILE

All content following this page was uploaded by Peter Erdi on 09 February 2016.

The user has requested enhancement of the downloaded file.

 Noname manuscript No.
(will be inserted by the editor)

Teaching Computational Neuroscience

Reviews:
Britt Anderson: Computational Neuroscience and Cognitive Modelling A Student’s Introduction to Methods
and Procedures. Sage, 2014
Hanspeter Mallot:Computational Neuroscience: A First Course, Springer, 2013
Thomas Trappenberg: Fundamentals of Computational Neuroscience. Oxford Univ. Press, 2010

Péter Érdi
arXiv:1412.5909v1 [q-bio.NC] 18 Dec 2014

Received: date / Accepted: date

Abstract The problems and beauty of teaching com- into sections reflecting the hierarchical organization of
the nervous system: Overviews, The Synaptic Level,
putational neuroscience are discussed by reviewing three
new textbooks. The Network Level, Neural Maps, Systems. The flag-
ship conference of the emerging discipline was organized
Keywords Computational neuroscience · Education
by Jim Bower [3] from 1992. As the Organization of the
PACS 87.19.L Computational Neuroscience website (www.cnsorg.org/)
describes: ”. . . Computational neuroscience combines math-
Mathematics Subject Classification (2010) 00A17 ·
ematical analyses and computer simulations with ex-
97U20
perimental neuroscience, to develop a principled under-
standing of the workings of nervous systems and apply
1 Teaching Computational Neuroscience: it in a wide range of technologies.”.
Diverse Perspectives
With a somewhat different perspective (shifting the
Computational Neuroscience is a discipline developed emphasis from structure-based models to functional as-
rapidly in the last twenty-five years. Roughly speaking pects) Cosyne (Computational and Systems Neuroscience)
it has two different meanings. First, how to use compu- conferences (http://www.cosyne.org/) have been or-
tational (more precisely theoretical and mathematical) ganized since 2003 initiated by Anthony Zador, Alex
methods to understand neural phenomena occurring at Pouget, and Zachary Mainen. As the conference website
different hierarchical levels of neural organization. Sec- specifies: ”Cosyne topics include (but are not limited
ond, how the brain computes (if at all). to): neural coding, natural scene statistics, dendritic
computation, neural basis of persistent activity, non-
While of course computational neuroscience has its linear receptive field mapping, representations of time
predecessors (from ”mathematical biophysics”, via ”cy- and sequence, reward systems, decision-making, synap-
bernetics” to ”theoretical neurobiology”, most likely tic plasticity, map formation and plasticity, population
the first book with the very title of ”Computational coding, attention, and computation with spiking net-
Neuroscience” was edited by Eric Schwartz [12]. The works. Participants include pure experimentalists, pure
chapters written by celebrated authors were grouped theorists, and everything in between.”
P. Érdi
Center for Complex Systems Studies
Kalamazoo College Computational neuroscience is a prototype of in-
Kalamazoo, MI 49006 terdisciplinary science, and scientist following different
1200 Academy Street traditions would agree that the education is a key com-
and
Institute for Particle and Nuclear Physics
ponent (i) to train new generation of future experts;
Wigner Research Centre for Physics, Hungarian Academy of (ii) help experimentalists (biologists, clinical scientists,
Sciences psychologists etc.) to understand what computational
E-mail: perdi@kzoo.edu model can (and cannot) do.
2 Péter Érdi

In this paper three recent textbooks are reviewed.I tor (Anderson mentions the joke with the physicists’s
am going to argue that while there is no reason to be- ”spherical chicken” approach). The example makes it
lieve that it is possible or even necessary to write a clear that before doing anything students should un-
single textbook, there are some overarching concepts derstand the state variables of any system under study.
and methods of computational neuroscience. However, In this simple mechanical example they are the velocity
computational neuroscience is taught (i) to students, and acceleration. The most common introduction to the
not only with different background, but also with dif- methodology of solving differential equations numeri-
ferent goals; (ii) by teachers with diverse backgrounds; cally is Euler’s method calculating the approximation
(iii) by reflecting upon different intellectual forerunner of the derivative by using small, but finite changes in
disciplines; (iv) occasionally by emphasizing the impor- the variable.
tance of some specific levels of the neural hierarchy and
(v) by emphasizing the importance of different mathe- From Action Potential to Programming Neurons: Inte-
matical methods. grate and Fire

2 The Motivational Approach This chapter briefly introduces action potential; and
the integrate and fire model.The explanation of the un-
Computational Neuroscience and Cognitive Modeling derlying physics is restricted to the definition of Ohm’s
[1] is written by Britt Anderson, a psychologist with law and Kirchoff’s law. It uses a Java applet for solving
MD at University of Waterloo. His motivation is clearly the Hodgkin-Huxley equations in the condensed form
stated: ”I find many psychological undergraduates are Eq. 1
interested in computational modelling, but that there is
not a lot of literature or texts to help them get started. dv(t)
To try and improve that, I took the notes from a semi- τ = RI(t) − V (t), (1)
dt
nar I have taught over the last few years, and wrote a
textbook.. . . ”.
and explains the meaning of dV /dt, τ , I(t). The solu-
tion of the equation tells the evolution of the voltage, or
Actually among the four parts of the book only the
membrane potential over time. Again, the author sug-
first two (1: Modeling Neurons, 2: Neural Networks) be-
gests to write a spreadsheet program to solve integrate
long to computational neuroscience by using the term
and fire models.
in a narrow sense, the other two parts (3. Probabil-
ity and Psychological Models, 4. Cognitive Modeling as
Logic and Rules) are rather being taught in Cognitive Hodgkin and Huxley: The Men and Their Model
Science classes.
Part I about neurons is concluded by a chapter of
the pioneers, and the emphasis is to explain conduc-
2.1 Modeling Neurons tance-based models. A single current is substituted by
the sum of individual currents flowing on specific ion
What is a Differential Equation?
channels, giving some hint to the interpretation of the
gating variables. Concerning the educational level, there
This chapter is clearly written for those whom this
is an explanation in a nice box with the title ”What Are
is the first encounter with differential equations. It ex-
the Parentheses for?”, where the reader should under-
plains the the difference between analytical and numer-
stand that V (t) is a number, actually the value of the
ical solutions, by solving the algebraic equation x2 = 4,
function V at the time-point t.
and shows that an easy computational technique of
numerical (as opposed to analytic) calculations is a
spreadsheet program. Of course, students should under- Programming
stand what is a D in the DE, so derivatives as instanta-
neous slopes, and the elementary jargon are discussed. In addition to spreadsheet programming, Anderson
suggests to learn some Python. While the book is cer-
Numerical Application of a Differential Equation tainly cannot substitute a programming class, it men-
tions some elementary concepts of programming, such
There is very simple, non-technical introduction to as declaring variables, controlling structures with loops
the idealized behavior of spring, i.e harmonic oscilla- and conditions.
Teaching Computational Neuroscience 3

2.2 Neural Networks exercises to understanding asynchronous updating al-

gorithms, and to convince themselves about the algo-
Neural Network Mathematics: Vectors and Matrices rithm’s converge.

How to teach the elements of linear algebra for stu- Programming

dents without technical background? The first question,
of course, Anderson addresses: ”What is a Vector?”.In Octave programming language is suggested to learn
addition to statement that it is a list of numbers, there how to write functions, and to implement and use delta
is also some hint about the geometric way of thinking on functions.
vectors, row and column vectors, with a smooth transi-
tion to the introduction to the concept or matrices and
elementary vector and matrix operations, including in- 2.3 Cognitive Modeling
ner products, dot product and matrix multiplication.
Part III is about ”Probability and Psychological Mod-
An Introduction to Neural Networks els”, and Part IV is ”Cognitive Modeling as Logic and
Rules”. While I see the importance of these concepts,
The first question to be discussed is: ”What neural in my own teaching experience it belongs to a Cognitive
networks are?”. Is is never easy to clarify the subtle science class, and will not be addressed in this review.
connection between biological real neural, and abstract
artificial neural networks. The author makes some ref-
erences to the early history, starting with just mention- 3 The Systems Neuroscience Approach
ing McCulloch and Pitts, John von Neumann and his
cellular automata, and more generally the emergence Hanspeter Mallot’ Computational Neuroscience [11] grew
of global structures from local interactions. It might be our from his notes on lectures delivered to his gradu-
a good exercise for psychology majors to play a little ate students in neuroscience. The book provides an es-
with cellular automata. sential introduction to computational neuroscience at
membrane, cellular, network and systems levels. . . .
One of the best part of the book is the simple expla-
nation of Perceptron. While it is true that one might
decide not to teach any artificial network in a computa- 3.1 Excitable Membranes and Neural Conduction
tional neuroscience class, its historical and conceptual
importance justifies its place.The geometrical interpre- The books starts with description of membrane poten-
tation of the perceptron learning rule helps to under- tials, by mentioning but not writing down the Nernst
stand how linear algebra works. The chapter concludes equation. It continues with the Hodgkin-Huxley theory.
with the description of the delta rule, and with argu- Related to modeling conductance changes it is shown by
ments about the necessity of having other neural net- an example what is a differential equation and its solu-
works, such as multilayer perceptron and support vec- tion. By explaining the physiology and kinetic models
tor machine. of potassium and sodium channels, the original (spa-
tially homogeneous) Hodgkin-Huxley model, a 4D ODE
Auto-associative Memory and the Hopfield Net is specified, and a hint is made that such kinds of equa-
tions ca be solved numerically only. The two-dimensional
Perceptron and its extensions were designed to solve approximation of the model (the FitzHugh-Nagumo model),
classification problems. Looking back to the historical at least its qualitative properties can be studied ana-
development, it is reasonable to ask about the simi- lytically, . The book does not go into the details how
larities and differences between Perceptron and Hop- changes in the input parameter leads to transition from
field nets? Hopfield transferred a mathematical tech- equilibrium to oscillatory response.
nique from physics (no more details were mentioned). It is always difficult to decide how to teach concepts
Essentially it mentions some remarkable features, as er- such as spatio-temporal propagation of action poten-
ror correction and the ability to pattern completions. tials. The explanation requires some knowledge of some
Here the book became a little more difficult, students basic concepts of partial differential equations. There is
should understand some elementary concepts of dynam- a section on passive conduction to give an intuitive in-
ical systems, and the the algorithm for calculating the troduction to cable equation, and to combine it with
outputs of a Hopfield network. Students should make the kinetic equations to channel models
4 Péter Érdi

3.2 Receptive Fields and the Specificity of Neuronal input images and activity distributions, The eigenfunc-
Firing tions of convolution is explained in details, using both
real and complex number notations.
How neurons integrate activities in space and time is an The basic theory of Fourier decomposition is based
important topic understanding the functions of specific on the fact that most functions can be expressed with
neurons. The superposition principle prescribes that the linear superposition of sine and cosine functions.
stimuli coming from different space points can be added. Via convolution theorem the construction leads to al-
The rigorous understanding of the concept of receptive gorithms of finding the coefficients of the Fourier se-
field function Φ(x, y) requires advanced math, such as ries. Generalizations for non-periodic functions and for
the theory of linear operators, b̧elongs to functional higher dimensions were also given.
analysis. The phenomenon of lateral inhibition, i.e. the
reduction of the activity of neighboring cells demon-
strated already by Mach, and related to the notion 3.4 Artificial Neural Networks
of convolution. Its understanding needs some insights
about integral equations. Tuning curves describe the I decided not to review ANN in details. From educa-
response of a neuron for different stimulus parameters. tional point of view the mathematics explained here
is the dot product and matrix operations. From the
Simple receptive fields are symmetric or show rota- perspective of applications classification and associative
tional symmetry, and they are well modeled by Gaus- memory are discussed.
sian function. To describe orientation-dependent effects
the celebrated Gabor function is extensively used. Ga-
bor function is generated by multiplying sinusoidal and 3.5 Coding and Representation
Gaussian functions.
This chapter contains two parts, population code and
The relationship between the set of all stimulus and retinotopic mapping.
the set of possible excitatory responses in the general
case is described by nonlinear operators. Static nonlin- There are different possibilities of coding with neu-
earities in receptive fields (such as threshold, saturation ral activity. Intensity code means that the activity of a
and compression)are illustrated. There is some expla- neuron is a monotonous function of a coded parameter.
nation of the concepts of Volterra kernels, for purely Channel coding and mostly population coding is used
temporal systems. more frequently. While population code needs more neu-
ron to encode some parameter value of a stimulus, still
Motion detection is identified to calculate velocity it looks superior to others, e.g. it leads to better reso-
and direction. The computational problem was sug- lution.
gested to be solved by designing coincidence and/or The section about information content of popula-
correlation detectors. From educational perspective it tion codes is used to explain the basic notions of infor-
requires the knowledge of some elementary probability mation theory. The celebrated concept of the center of
theory. The author jumped suddenly to introduce auto- gravity estimator for reading a population code is de-
and cross-correlation functions, and hopefully most stu- scribed.
dents have the background to understand them. It might
have been a reasonable question to ask what type of The last section about retinotopic mapping is writ-
computations can be implemented by single cells, and ten based on the research the author did with Werner
what should remain for neural networks. von Seelen (who actually wrote the Foreword of the
book) in the nineteen-eighties. In terms of mathematics
the question is how to describe the coordinate transfor-
3.3 Fourier Analysis for Neuroscientists mation from retina to visual cortex. Conformal maps
and log-polar mapping are the appropriate tools to de-
I read this chapter with great interest . It starts with ex- scribe the geometry of retinotopy.
amples of light spectra, acoustics, vision and magnetic
resonance tomography. Terms, such as spectrum and
Fourier decomposition appear. Mathematical concepts 4 The Computational Approach
from complex numbers to Gaussian convolution kernels
have been reviewed. Eigenfunctions and eigenvalues are Thomas Trappenberg’s Fundamentals of Computational
introduced as characteristics of the operator connecting Neuroscience [13] is a comprehensive textbook for ad-
Teaching Computational Neuroscience 5

vanced graduate students, It tries to teach the mini- first the McCulloch-Pitts model.One challenge is to ex-
mally necessary neurobiology a computational science plain the the response of neurons for different (constant,
student should know. In terms of using mathematics the and time-dependent) inputs. While the integrate-and-
author took a pragmatic approach, mentioned mathe- fire models neglect the details of subthreshold dynam-
matical concepts when explicitly used. Some elemen- ics, the Izikievich neuron is a nice compromise between
tary textbook notions of linear algebra, calculus, nu- biological plausibility and computational efficiency. All
merical solutions, probability and information theory. the previous models are deterministic, and cannot de-
Descrition of MATLAB were put into Appendix.I found scribe the inherent random character of neural firing.
the book very carefully and clearly written and close to Somehow stochasticity should be put to the models.
the spirit I am using in classrooms. The book actually does not mention the methods grew
up from the classical paper of Gerstein and Mandelbrot
The book has three main parts (Basic Neurons, Ba- [4] which considers the membrane potential change as
sic Networks, System-Level Models) following the spirit a random walk. Spike-time variability can be explained
of bottom up modelling perspective preceded by a nice with the introduction of stochastic threshold or reset.
introduction to computational neuroscience. Of course, The noisy integrator puts the randomness to the input.
it is somewhat a matter of personal taste how to write
briefly about such questions as ”What is computational One of the main challenge is to understand the na-
neuroscience?”, ”What is a model?”, ”Is there a brain ture of neural code: what is the relationship between fir-
theory?”, ”A computational theory of the brain”. ing patterns and behavior. A somewhat different ques-
tion is about the relationship between sensory stimuli
and neural activity patterns. Rate code (going back
4.1 Basic Neurons
to Lord Adrian’s finding), correlation codes and coin-
This part starts with a chapter on neurons and conductance-cidence detectors are explained, maybe a little more
based models, with a little biology about cellular mor- about temporal coding is missing.
phology, synapses, ion channels and membrane poten-
tial. After writing about synaptic mechanisms, it ex- While we are still discussing about (more or less)
plains the problem of modelling synaptic response, maybe single neurons, there is a section about population dy-
a little too early. This sub-chapter seems to be a lit- namics) to be studied again in a later chapter. The
tle dense, and between the explanation of Kirchhoff’s state of the population of neurons can be characterized
law and Ohm’s law there is a short introductions to by some population activity vector. Actually the vague
differential equations. The next sub-chapter explains concept of ”cell assembly” (or local pool, based on the
the generation of action potential and Hodgkin-Huxley terminology used in the book) suggests that cells in a
equations followed by some elements of numerical in- certain pool have common features. I miss something
tegration, and implementation with MATLAB. Cable more about mesoscopic neurodynamics. (We worked
theory uses partial differential equations (actually it slowly twenty years ago on a model of large popula-
is a parabolic, and not a hyperbolic equation, as it is tions of neurons. What we saw is that a ”mean field”
written.). The spatial discretization of neurons lead to approach is not enough, and the behavior at least as
compartmental models, and the reader get some refer- for an ”average cell” (which is a hypothetical neuron
ence to the two most frequently used neuron simulators receiving the average synaptic input calculated by a
(GENESIS, NEURON). statistical approach) is also necessary. The chapter fin-
ishes with two clearly written sections; one describes
The next chapter comprises two reasonable topics: the different activation (or transfer) functions, and the
why and how to simplify neurons, and how to treat other networks with non-classical synapse taking into
the dynamics of neural populations. Trappenberg writes account nonlinear interactions among neurons.
correctly: one trivial reason is to make possible calcu-
lations with large number of neurons, but conceptu- What might be more complicated than one neuron?
ally more important to get an insight into the skele- Two neurons connected by a synapse. While the con-
ton mechanism of the generation of emergent network nectivity matrix of a network is able to store memory
properties. It somewhat arbitrary, what is a good or- traces, learning is related to the mechanism of synaptic
der to explain different models. Here first we see leaky modifiability to be described by (generalized) Hebbian
integrate-and-fire neuron, than spike response model, learning algorithms. As the book correctly reflects, Heb-
Izikievich neuron and McCulloch-Pitts neuron. (I like bian learning mechanisms explained first psychological
to teach following the historical development, teaching phenomena, such as learning by associations, and condi-
6 Péter Érdi

tioning. Going downwards on the ladder, than the phys- the spirit of this review I don’t discuss in details. The
iological basis of plasticity is given. As it is well known, single and multilayer perceptron and support vector
the identification of the cellular mechanism of synap- machine are clearly presented from classroom teaching
tic plasticity (related to the long term potentiation and perspective.
long term depression) was given a quarter century later
after postulating the Hebbian mechanism. (Actually for Cortical maps between two regions are used to rep-
me it was very useful to learn about the work of Alan resent features. One of the main principles of neural
Fine and Ryosuke Enoki about their plasticity exper- organizations states that maps often have topographic
iments by combining paired-pulse facilitation and cal- character, i.e. neighboring areas represent neighboring
cium imaging of single synapses.) At even lower level, features.The normal ontogenetic formation and the plas-
the biochemical pathways related to the role of NMDA tic response of these maps for certain lesions have a
receptors and calcium ions is discussed. common mechanism to be modeled by self-organizing
algorithms. The famous Willshaw - von der Malsburg
While Hebb described his rule verbally, now there model and the Kohonen algorithm are the paradigmatic
are many variations of the mathematical forms of the examples of these algorithms, Maps get newer train-
phenomenological learning rules. Again, as a matter of ing input patterns, but they should preserve their ba-
taste, Trappenberg starts with a newest rule (spike tim- sic properties. The problem, called stability-plasticity
ing dependent plasticity); the book does not mention dilemma [6].
what it is now well known: there was an (early discov-
ery of the phenomenon by Levy and Steward 1983 [10] In certain applications it is plausible to use con-
it came too early). I am comfortable the way of pre- tinuous space coordinates instead of discrete neurons,
sentation (but we did a different approach in the sec- so the appropriate dynamic model to describe spatio-
ond edition of Michael Arbib’s wonderful Handbook of temporal activity propagation would be a partial differ-
Brain Theory [2]. Also, the last (very well written) sec- ential equation. In the general case the system is non-
tion on synaptic scaling and weight distributions should autonomous, since the input is time-dependent. How-
have been written within the section on mathematical ever, the activity dynamic in a certain point depends on
formulation. other spatial coordinate to be taken into account an in-
tegral (kernel) Gaussian function and the Mexican-hat
4.2 Neural networks functions are characteristic examples nnnnnof centre-
surround interaction kernels. The dynamics lead to dif-
There is a very difficult question without having a clear ferent qualitative states. One of them, is characterized
answer: What to teach about neural organization? Trap- by a stable active region even after the removal of ex-
penberg discusses large-scale neuroanatomy and hierar- ternal input, so memory can be stored. Mathematically
chical organization, the layered structure and the mod- the construction leads to continuous attractors. Again,
ular organization of the cortex (according to my biased from educational perspective it is disadvantageous that
view, John Szentágothai (and also Valentino Braiten- we have not yet read about the simpler concept of point
berg) might have mentioned. attractors. Inferior temporal cortex, prefrontal cortex
implementing working memory and hippocampus in-
I am not totally sure whether random networks should volved in space representation are illustrative examples
be studied emphatically Actually Anninos and his col- of the computational concept.
leagues are mentioned, but more precisely their per-
spective was to consider neural networks with the fea- The main goal of the chapter (Recurrent associa-
ture described as ”randomness in the small and struc- tive networks and episodic memory) is to show how re-
ture in the large”. The remaining part of the chap- current neural networks might be the anatomical sub-
ter about more physiological spiking networks based on strate of the associative (specifically auto-associative)
Izikievich is plausible, and appropriate for demonstra- memory. Auto-associative memory basically does pat-
tive simulations. From educational perspective it is not tern completion. The reentrant connections among the
convincing that spectral analysis and fast Fourier trans- pyramidal cells of the hippocampal CA3 region imple-
formation mentioned without telling anything about ment a network capable of storing and retrieving, (at
them. least static) input patterns. While there were forerun-
ners and followers, John Hopfield [8,9] popularized an
The chapter about feed-forward networks belongs abstract network implementing point attractors mostly
mostly to artificial neural networks, and according to for the physicist’s community (and it cannot be denied
Teaching Computational Neuroscience 7

that Google Scholar gives altogether twenty-thousand Obviously it is very difficult to write a short chapter
citation for the two papers.). about the cognitive brain. As a matter of taste what
should be the illustrative examples, The first two ex-
Section 8.2 nicely describes the standard math com- amples Trappenberg chose attentive vision are the in-
ing from the physicists theory of phase transitions of terconnecting workspace hypothesis. I find very impor-
the spin glasses. A remarkable feature of the attractor tant and also well-written the section about the antic-
neural networks is their noise tolerance of the stored ipating brain. A large school of computational neuro-
memories. While the original work (and the underlying science adopts the top down computational approach.
mathematical theorems) are based on rather restrictive The starting point is that the brain builds a model of
assumptions, simulation results generally permit the ex- the external world and generates hypothesis about the
tension of the theory. In this book some important fur- consequences of possible actions. The formal framework
ther studies, such as the effects of dilution (i.e the de- to formulate the predictive power of is probabilistic The
viation from the fully connected networks, and for non- subsection on probabilistic reasoning and Bayesian net-
random, but correlated, input patterns are discussed. work briefly discusses a technique of the analysis. The
There is one neglected problem with attractor neural main corpus of the book is ended by the adaptive res-
networks, that Trappenberg also does not mention: real onance theory, Any author has the right to decide how
neural systems processes time-dependent inputs, so the to finish a book.
differential equations set to describe the dynamic be-
havior are non-autonomous. Non-autonomous systems
don’t have attractors in the general case.

I find somewhat arbitrary both the ”where” and

the ”how” of discussions of the dynamic systems and 5 Computational Neuroscience: A Wonderful
chaotic attractors. I might be biased, but I think a sep- Subject to Teach and Learn
arate chapter on dynamic would be useful. Between the
point attractors and chaotic regimes the very important The discipline of Computational Neuroscience discusses
neural oscillators should have been studied. The Cohen mathematical models, computational algorithms, and
- Grossberg theroem is extensively described. I would simulation methods that contribute to our understand-
have liked to see Hirsch’s celebrated paper [7], which ing of neural mechanisms. Students can learn the basic
gives a more generalized, but still operative study of the concepts and methods of computational neuroscience
problem. Also, both the structural basis and the func- research, and absorb some knowledge of neurobiologi-
tional significance of the chaotic neural network (see e.g cal concepts and mathematical techniques.
[14] might have been discussed.
What are the goals of teaching computational neu-
roscience? The first goal is to teach WHY mathemati-
4.3 System-Level Models cal and computational methods are important in under-
standing the structure, function and dynamics of neural
Section nine (Modular networks, motor control, and organization. The second goal is to explain HOW neu-
reinforcement learnig) and ten (The cognitive brain) ral phenomena occurring at different hierarchical levels
about system level models, and I find them somehow can be described by proper mathematical models.
the less convincing part of the book. I acknowledge
the important aspects of the discussed topics. Modu- The books under review took some different per-
lar mapping networks, coupled attractor networks, se- spective, had different goals, and it is good to see that
quence learning, complementary memory systems, mo- teachers have alternatives. and it is not the end, There
tor learning and control, reinforcement learning are all are a couple of other books around, I simply selected
important topics. However, I somehow missed the co- books, which from one or another reasons are on my
herent story line, Mixture of experts is clearly a topic in desk. A new great book [5]. is left for another time.
machine learning. ”Sequence learning”, gave a partial
answer what to do with (at least some specific) time-
dependent inputs. It would have been possible to show Acknowledgements Thanks for my numerous teaching as-
sistants over the years. I had many conversation with them
either motor sequence learning, or to discuss the role of about the method of teaching of this discipline. The author
hippocampus in sequence learning. also thanks to the Henry Luce Foundation to let him to be a
Henry R Luce Professor.
8 Péter Érdi

References

1. Anderson, B.: Computational Neuroscience and Cogni-

tive Modelling. A Student’s Introduction to Methods and
Procedures. SAGE Publ. Ltd (2014)
2. Arbib, M.: The Handbook of Brain Theory and Neural
Networks. MIT Press (2002)
3. Bower, J.: 20 Years of Computational Neuroscience.
Springer Verlag (2013)
4. Gerstein, G.L., Mandelbrot, B.: Random walk models for
the spike activity of a single neuron. BIOPHYS-J 4(1),
41–68 (1964)
5. Gerstner, W., Kistler, W., Naud, R., Paninski, L.: Neu-
ronal Dynamics - From single neurons to networks and
models of Cognition. Cambridge University Press (2014)
6. Grossberg, S.: How does a brain build a cognitive code?
Psychological Review 87, 1–51 (1980)
7. Hirsch, M.: Convergent activation dynamics in continu-
ous time networks. Neural Networks 2, 331349 (1989)
8. Hopfield, J.J.: Neural networks and physical systems with
emergent collective computational properties. Proc. Nat.
Acad. Sci. (USA) 79, 2554–2558 (1982)
9. Hopfield, J.J.: Neurons with graded response have collec-
tive computational properties like those of two-sate neu-
rons. Proc. Nat. Acad. Sci. (USA) 81, 3088–3092 (1984)
10. Levy, W., Steward: Temporal contiguity requirements for
long-term associative potentiation/depression in the hip-
pocampus. Neuroscience 8, 791797 (1093)
11. Mallot, H.: Computational Neuroscience. A First Course.
Springer (2013)
12. Schwartz, E. (ed.): Computational neuroscience. MIT
Press, Cambridge, Mass (1990)
13. Trappenberg, T.: Fundamentals of Computational Neu-
roscience. Oxford Univ. Press (2010)
14. Tsuda, I.: Dynamic link of memorychaotic memory map
in nonequilibrium neural networks. Neural Networks 5,
313–326 (1992)

View publication stats