Institut für
Grundlagen der Informationsverarbeitung (7080)
Lecturer:
O.Univ.Prof. Dr. Wolfgang Maass
Office hours: by appointment (via email)
Email: maass@igi.tugraz.at
Homepage: www.igi.tugraz.at/maass/
Assoc. Prof. Dr. Robert Legenstein
Office hours: by appointment (via email)
Email: robert.legenstein@igi.tugraz.at
Homepage: www.igi.tugraz.at/legi/
Marblestone, Adam, Greg Wayne, and Konrad Kording (2016). "Toward an Integration of Deep Learning and Neuroscience." Frontiers in Computational Neuroscience, 10. [PDF@Frontiers].
hypothesizes that biological neuronal systems may utilize learning processes that share similarities with deep learning techniques. We will discuss in this semester this paper and selected references therein.Abstract: Neuroscience has focused on the detailed implementation of computation, studying neural codes, dynamics and circuits. In machine learning, however, artificial neural networks tend to eschew precisely designed codes, dynamics or circuits in favor of brute force optimization of a cost function, often using simple and relatively uniform initial architectures. Two recent developments have emerged within machine learning that create an opportunity to connect these seemingly divergent perspectives. First, structured architectures are used, including dedicated systems for attention, recursion and various forms of short and longterm memory storage. Second, cost functions and training procedures have become more complex and are varied across layers and over time. Here we think about the brain in terms of these ideas. We hypothesize that (1) the brain optimizes cost functions, (2) these cost functions are diverse and differ across brain locations and over development, and (3) optimization operates within a prestructured architecture matched to the computational problems posed by behavior. Such a heterogeneously optimized system, enabled by a series of interacting cost functions, serves to make learning dataefficient and precisely targeted to the needs of the organism. We suggest directions by which neuroscience could seek to refine and test these hypotheses.
(1) R. C. O’Reilly, D. Wyatte, and J. Rohrlich
(2014). Learning through time in the thalamocortical loops.
arXiv:1407.3432v1. https://arxiv.org/abs/1407.3432
(2) W. Lotter, G. Kreiman, and D. Cox (2016).
Unsupervised learning of visual structure using predictive
generative networks. arXiv:1511. http://arxiv.org/abs/1511.06380
(3) R. C. O’Reilly, T. E. Hazy, J. Mollick, P.
Mackie, and S. Herd (2014). Goaldriven cognition in the brain:
A computational framework. arXiv:1404.7591v1. https://arxiv.org/abs/1404.7591
(4) J. O. Rombouts, S. M. Bohte, and P. R.
Roelfsema (2015). How attention can create synaptic tags for the
learning of working memories in sequential tasks. PLOS
Computational Biology  DOI:10.1371/journal.pcbi.1004060. http://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1004060
(5) G. Wayne, and L. F. Abbott (2014).
Hierarchical control using networks trained with higherlevel
forward models. Neural Comput 26(10):21632193. doi:
10.1162/NECO_a_00639. http://www.ncbi.nlm.nih.gov/pubmed/25058706
(6) T. J. Sejnowski, H. Poizner, G. Lynch, S.
Gepshtein, and R. J. Greenspan (2014). Prospective Optimization.
Proc IEEE Inst Electr Electron Eng. 2014 May;102(5):799811.
DOI: 10.1109/JPROC.2014.2314297. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4201124/
(7) E. Jonas, and K. Kording (2016). Could a
neuroscientist understand a microprocessor?. bioRxiv, 055624. http://www.biorxiv.org/content/early/2016/05/26/055624.abstract
(8) I. Sutskever, J. Martens, G. E. Dahl, and G.
E. Hinton (2013). On the importance of initialization and
momentum in deep learning. Proceedings of the 30 th
International Conference on Machine Learning, Atlanta, Georgia,
USA, 2013. JMLR: W&CP volume 28. http://www.cs.toronto.edu/~fritz/absps/momentum.pdf
(9) J. Yosinski, J. Clune, and Y. Bengio (2014).
How transferable are features in deep neural networks? In
Advances in Neural Information Processing Systems 27: 33203328.
http://papers.nips.cc/paper/5347howtransferablearefeaturesindeepneuralnetworks
(10) C. Gülcehre, and Y. Bengio (2016). Knowledge
matters: Importance of prior information for optimization.
Journal of Machine Learning Research, 17(8), 132. www.jmlr.org/papers/volume17/gulchere16a/gulchere16a.pdf
Date 
Speaker 
Talks 

24.10.2016
16:1518:00 
Maass, Legenstein 
Introduction 
PDF 
21.11.2016
15:4518:00 
Absenger, Mulle 
Goaldriven cognition in
the brain: A computational framework. R. C. O’Reilly, T.
E. Hazy, J. Mollick, P. Mackie, and S. Herd (2014) 
PDF 
28.11.2016 15:4518:00  Marchetto, Raggam 
Learning through time in
the thalamocortical loops. R. C. O’Reilly, D. Wyatte,
and J. Rohrlich (2014) 

Harb, Micorek  Prospective Optimization. T. J.
Sejnowski, H. Poizner, G. Lynch, S. Gepshtein, and R. J.
Greenspan (2014) 
PDF 

12.12.2016 15:4518:00  Steger, Zöhrer 
Unsupervised learning of visual structure using predictive generative networks. W. Lotter, G. Kreiman, and D. Cox (2016)  PDF 
Wohlhart, Müller 
Could a neuroscientist understand a microprocessor?. E. Jonas, and K. Kording (2016)  PDF 

09.01.2017 15:4518:00  Legenstein 
A brief introduction into
deep learning 

Lindner, Narnhofer 
Knowledge matters: Importance of prior information for optimization. C. Gülcehre, and Y. Bengio (2016)  PDF 

23.01.2017
15:4518:00 
Topic, Eibl 
On the importance of initialization and momentum in deep learning. I. Sutskever, J. Martens, G. E. Dahl, and G. E. Hinton (2013)  PDF 
Fuchs, Ainetter 
How transferable are features in deep neural networks? J. Yosinski, J. Clune, and Y. Bengio (2014)  PDF 