R. Kopiez, A. C. Lehmann, I. Wolther & C. Wolf (Eds.)
Proceedings of the 5th Triennial ESCOM Conference
8-13 September 2003, Hanover University of Music and Drama, Germany
COMPETITIVE VS FEED-FORWARD MODELING OF MUSIC RECOGNITION TASKS
Alessandro DAusilio1, Frederic G. Piat2, Alessandro Londei1,3, Marta Olivetti Belardinelli1,Dept. of Psychology, University of Rome La Sapienza, Italy 2 Pierre et Marie lab for Artificial Intelligence University of Paris VI, France 3 ECONA: Interuniversitary Center for Research in Cognitive Processing in Natural and Artificial Systems, Italy

ABSTRACT

Background. Most of the connectionist models designed for musical stimuli recognition, seems to be unrealistic, for two main reasons: first of all, most are based on a tonal a priori knowledge; second, because they do not include a competitive learning phase. Aims. Our previous researches highlight salience, operatively defined as redundancy of interval or rhythmic parameters, as an important dimension to recall melodies from memory. This research is aimed at demonstrating that musical superficial characteristics are a sufficient basis for music categorization. Method. Different input schemes and neural networks models have been tested on a categorization task. Results. Results obtained within different theoretical frameworks, give us the opportunity to reconsider how cognitive music processing looks like. Conclusion. Neural networks can easily replicate human capabilities in a large variety of experimental tasks, even if each model has its own style. In our case it appeared that the competitive strategy obtained better results in detecting superficial cues, and in using them to separate and recall incoming patterns. In this work, salience appeared as a relevant variable for music recognition by naives subjects, and the evidence is strengthened by the fact that the results were collected by means of a set of different neural network architectures.
implies an exact sequence and temporal coding, the latter on the other side, an exact pitch representation. The first coding scheme is based on a six notes window, each corresponding to three neurons. The first codes the time passed since the previous note attack, (for the first note this value is always set to zero), the second one is the note duration and the third node stands for the pitch contour (each note is a value ranging from 88 to +88, relative to the previous one) For the second coding scheme, instead, each of the 18 input nodes codify a 10 msec. time slice, its activation is computed in terms of melodic contour, as explained before. In both cases the input pattern slides (one note in the first, 10 msec. in the second), to form the new pattern
Figure 1: Input coding n 1.

1. STIMULI

The 48 musical stimuli have been constructed by composer Fabio Cifariello Ciardi according to these specifications: two variables, tonality and salience, both on two levels, presence and absence. Salience has been defined as: redundancy or repetition of melodic and/or rhythmic patterns. Each category has 12 musical pieces.

1.1. Input Codification

To accomplish our task, we need a precise codification of musical events. Basically we want to loose as few information as possible, so that we can always rebuild the original musical piece from the input code. This necessity was due to our hypothesis, as a matter of fact we are looking for both timing and melodic redundancy. The former Figure 2: Input coding n 2.
ISBN 3-931852-67-9 ISSN 1617-6847

Node num. 17 18

Code 0.0.2083 0.2083 2.000 0.2083 0.4167 -1.000 0.4167 0.0.2083 0.2083 2.000 0.2083 0.4167 -2.000

Code 2

As we already said, our hypothesis has not been simulated before, so we needed to compare different approach to gain more knowledge on such a field.

3. PRESENT RESULTS

Early simulations clearly showed how the input codification and data preprocessing are of the greatest importance. We collected a great amount of data, which strongly urge us to continue in such a way. To be more precise, more tests are needed with other coding scheme because present results appear to be sometimes ambiguous. Below its shown a SOMs classification made with coding scheme 1. Here is clearly visible how Tonal Salient pieces are easily classified, on the other hand the net completely fails with non Tonal non Salient class. But more interesting is its behaviour with the other two classes. Although both achieve an average percentage, nT - S pieces get an higher hits number. According with human data Salience seems to facilitate the overall network performance.
Table 1: First input given to the net, in both coding.

2. MODELS

Data were collected with different architectures and approaches, more precisely a Kohonens Self Organizing Map, Piats ARTIST model which is based on the ART algorithm, and a Perceptron with and without Back-propagation rule.
2.1. Why To Use Different Coding Scheme?
To perform our simulation we had to solve many problems generated by the fact that the stimuli were build for a quite different task. As a matter of fact they have been widely tested on humans, with a classic recall paradigm. Subjects were presented with a sample of stimuli and, after a varying period of time, they were asked to listen the whole set and to recognize the stimuli they heard before. As a consequence stimuli have different time length and different number of notes. Obviously it is possible neither to feed the net with an entire piece at once, nor with a jumping window ( note1.note5; note6.note10;. ). This a pilot study on which topic few experiments have been conducted with means of neural network models, as a consequence we though the best way was to test various coding to understand which was better, in terms of general performance, and which was more psychologically relevant.
Figure 3: code n1 with a SOM model.

4. CONCLUSIONS

Even if these initial results support our hypothesis, many other simulations are needed for many reasons: 1. Better inputs psychological validity. 2. Better matching with human data. 3. Compare other neural network models 4. Find the best algorithm to quantify, in an empirical way, variable Salience. 5. Extend the model to other stimuli. Point one is referred to the different levels of complexity we can observe in musical pieces, actually we consider only part of them, but people dont. Secondly, the percentage obtained with nT nS, seems to be quite unrealistic. As regards point three,

2.2. Why To Use Different Models?
Every neural network model has its own best application, and is built on a particular theorical framework. Our task is not only to replicate human data, as a matter of fact we need also a way to measure if every single stimulus is a good example of its class, and obtain, if possible, the inter-class stimuli distances.
we feel confident that comparing many approaches would be useful to better understand how they really works in the musical domain. The fourth issue instead, could be a starting point for further experiment on such hypothesis. Finally the fifth point is probably the most achieving, because it involves the possibility to quantify how a listener will understand and remember of pieces he attended to.
13. Leman M. (1992). The theory of tone semantics: Concept, foundation, and application. Minds and Machines, 2(4), 345363. 14. Leman M. (1995). A model for retroactive tone-center perception. Music Perception, Summer vol. 12(4), 439471. 15. Olivetti Belardinelli M., Rossi-Arnaud C. (1999), Recollection and familiarity in recognition memory for musical themes. In Vandierendonck A., Brysbaert M., Van Der Goten K. (Eds). XI Conference of European Society for Cognitive Psychology, Academia Press. 16. Olivetti Belardinelli M. (2000). Structuring factors in musical processing. Proceedings of the 6th International Conference on Music Perception, Keele University. 17. Olivetti Belardinelli M., Rossi-Arnaud C., Pitti G., Vecchio S. M. (2000). Looking for the anchor points for musical memory. Proceedings of the 6th International Conference on Music Perception, Keele University. 18. Petroni N. C., Tricarico M. (1997). Self-organizing neural networks and the perceptual origins of the circle of fifths. In Leman M. Music gestalt and computing. Studies in cognitive and systematic musicology. Springer Verlag, Berlin, Germany. 19. Piat F. (1999). ARTIST: adaptive resonance theory to internalize the structure of tonality. Unpublished PhD Thesis, University of Texas at Dallas. 20. Piat F. (2000). ARTIST: a connectionist model of music acculturation, Proceedings of the 6th International Conference on Music Perception, Keele University. 21. Stevens C., Latimer C. (1997). Music Recognition: an illustrative application of a connectionist model. Psychology of Music, 25, 161-185. 22. Tillmann, B., Bharucha, J., Bigand, E. (2000). Implicit learning of tonality: A self organizing approach. Psychological Review, 107, 885-913. 23. Todd P. M., Loy D. G. (Eds) (1991). Music and Connectionism. MIT Press/Bradford Books, Cambridge, MA.

5. REFERENCES

1. Berz W. L. (1995). Working memory in music: a theoretical model. Music Perception, Spring vol. 12(3), 353-364. 2. Bharucha, J. J. (1987). Music cognition and perceptual facilitation: A connectionist framework. Music Perception 5(1), 130. 3. Carpenter, G. & Grossberg, S. (1987). ART2: Selforganization of stable category recognition codes for analog input patterns, Applied Optics 26(23), 4919 4930. 4. Duch W., Jankowski N. (1998). Survey of neural transfer functions. Neural Computing Surveys 1, 61-101. 5. Gjerdingen, R. (1990), Categorisation of musical patterns by self-organizing neuronlike networks, Music Perception 7(4), 339370. 6. Gjerdingen, R. (1992). Learning syntactically significant temporal patterns of chords: a masking field Embedded in an ART3 architecture. Neural Networks, 5, 551-564. 7. Griffith N. (1993). Modelling the acquisition and representation of musical tonality as a function of pitchuse through self-organizing artificial neural networks. Unpublished PhD Thesis, University of Exeter. 8. Griffith N., Todd N. P., (Eds.) (1999). Musical Networks: parallel distributed perception and performance. Cambridge, MA: MIT Press. 9. Japkowicz N. (2001). Supervised versus unsupervised binary-learning by feedforward neural networks. Machine Learning, 42, 97-122. 10. Kohonen, T. (1989). A self-learning musical grammar, or Associative memory of the second kind. In Proceedings of the International Joint Conference on Neural Networks (pp. 1-5). New York: IEEE. 11. Large W. E. (1994). Dynamic representation of musical structure. Unpublished PhD Thesis, Ohio State University. 12. Large W. E., Palmer C., Pollack J. B. (1995). Reduced memory representations for music. Cognitive Science, 19, 53-96.