Many studies in neuropsychology have highlighted that expert musicians, who started learning music in childhood, present structural differences in their brains with respect to non-musicians. This indicates that early music learning affects the development of the brain. Also, musicians’ neuronal activity is different depending on the played instrument and on the expertise. This difference can be analysed by processing electroencephalographic (EEG) signals through Artificial Intelligence models. This paper explores the feasibility to build an automatic model that distinguishes violinists from pianists based only on their brain signals. To this aim, EEG signals of violinists and pianists are recorded while they play classical music pieces and an Artificial Neural Network is trained through a cloud computing platform to build a binary classifier of segments of these signals. Our model has the best classification performance on 20 seconds EEG segments, but this performance depends on the involved musicians’ expertise. Also, the brain signals of a cellist are demonstrated to be more similar to violinists’ signals than to pianists’ signals. In summary, this paper demonstrates that distinctive information is present in the two types of musicians’ brain signals, and that this information can be detected even by an automatic model working with a basic EEG equipment.
Abstract of the scientific paper
Coro, G., Masetti, G., Bonhoeffer, P., & Betcher, M. (2019, September). Distinguishing Violinists and Pianists Based on Their Brain Signals. In International Conference on Artificial Neural Networks (pp. 123-137). Springer, Cham.
https://link.springer.com/chapter/10.1007%2F978-3-030-30487-4_11
Hubble Asteroid Hunter III. Physical properties of newly found asteroids
Distinguishing Violinists and Pianists Based on Their Brain Signals
1. Distinguishing Violinists and
Pianists based on their Brain
Signals
GIANPAOLO CORO, GIULIO MASET TI
I N S T I T U T E O F I N F O R M A T I O N S C I E N C E A N D T E C H N O L O G I E S " A L E S S A N D R O F A E D O " ( I S T I ) , N A T I O N A L R E S E A R C H C O U N C I L O F
I T A L Y ( C N R )
PHILIPP BONHOEFFER, MICHAEL BETCHER
A S S O C I A Z I O N E I L P O G G I O M O N T E C A S T E L L I
CORR AUTH: CORO@ISTI.CNR.IT
2. Context
Music influences the development of the brain from childhood to
adulthood (Chen-Hafteck et al. 2018)
The pattern of brain architecture, brain’s plasticity, and behaviour
development are affected by early music learning, perhaps with music-
specific neural networks (Peretz and Zatorre, 2005)
Structural differences exist between the brains of adult musicians and non-musicians (Gaser et al. 2003, Vaquero et al.
2016)
Musicians’ neural activity changes depending on the played instrument (Zatorre et al. 2007)
Sound is also processed by the musician’s auditory circuitry, and the brain adjusts movements based on the
processed information and visually processes and interprets symbols if the musician is reading music (Critchley and
Henson, 2014)
Structural differences are related to complex motor, auditory, and multi-modal skills but their nature is still
unclear
3. Evidence in EEG Signals
Music-related brain activity is correlated with the signals produced by different concurrent neuronal aggregations from
several areas of the brain.
EEG and music therapy and music production:
Music has been deduced from brain signals in neurotherapy (Vaquero et al., 2016), stress control (Subhani et al., 2018), brain activity
monitoring (Baier et al., 2007), and music generation (Potard and Schiemer, 2004)
Musicians’ brain signals have similarities in terms of their correlation with the way sound is produced (Petsche et al., 2005)
EEG and music listening:
Listening to different types of music causes changes of energy intensity in different brain regions (Sun et al., 2013)
Musicians and non-musicians have different EEG response to music listening (Liang et al., 2011; Ribeiro and Tomaz, 2018)
EEG and musical expertise:
Musical expertise is correlated with altered functional connectivity and brain plasticity (Steward et al., 2003; Oechslin et al. 2010;
Paraskevopoulos et al. 2015)
4. Issues
Most analytical approaches are based on neurophysiological
analytical frameworks.
They require complex and expensive equipment
They aim at explicitly modelling very complex and unknown
phenomena
Typically, they can distinguish a musician from a non-musician but they
don’t identify the type of musician
(from Gu et al., 2018)
(from Chen et al. 2013)
5. Our Goal
A preliminary investigation on the differences between the EEG
signals of expert violinists and pianists when playing:
1. Similarities between musicians’ brain signals depend on the played
instrument and on musical expertise
2. EEG signals allow to automatically distinguish violinists from pianists
Model
Violinist Pianist
3. Discriminant information can be detected even using basic equipment and models
4. Musical expertise influences the performance of the model
5. Provide Open Science tools to check, revise, and extend our experiment with different categories and data
6. A Peculiar Laboratory – Il Poggio http://www.ilpoggiomontecastelli.com/
..In Montecastelli Pisano (Tuscany, Italy)
A project of Science and Music
conceived as a melting pot for creative
people and for music and science:
• A violinmaking workshop
• Concerts and festivals
• Scientific conferences
• Top-level musicians
• Top-level instruments (also from 1560 to
1800)
Aims at answering to questions on music and
musicians:
Why does a musician have to practice the same technical
passage for hours and hours? What happens in the brain to
require this daft endless repetition? What is a good violin?
Who defines the quality? Would a computer analysis on
acoustic signals be better than the human ear?
7. EEG Device Used
• NeuroSky EEG dry biosensor embedded in the NeuroSky
MindWave wearable headphone-like tool
• EEG from the frontal-pole (Fp1) position of the cerebral cortex
• Raw mono-dimensional signal sampled at 512 Hz
Pros:
Fair precision
Low cost
Built-in noise reduction filter and power spectrum
calculation (1s windows)
Lightweight and portable – does not obstruct musician’s
movements
Cons:
One channel only at the Fp1 position
Fp1 activity is correlated also to other
movements (e.g. eyes’ ones) and only indirectly
to music
The model has to recognize the indirect effects of
playing different instruments on Fp1 EEG signals
Post-processing of the signal spectrum:
Band-pass filtered (between 0.5Hz and 100Hz) to reduce motion artefacts (Patki et al. 2012; Bertrand et al. 2013)
Classified according to common subdivision ranges of brain signals frequency bands: (low-high) Gamma (40-100
Hz), (low-high) Beta (12-40 Hz), (low-high) Alpha (8-12 Hz), Theta (4-8Hz), and Delta (0.5-4 Hz) waves.
8. EEG Signals and Experimental Setup
Musicians:
Nine males: 4 Pianists, 4 Violinists, 1 Cellist
Expert musicians were professionals and started music training in childhood
Music - Two pieces for each player:
Easy-familiar: a piece the player knew well and that was not demanding
Challenging-unfamiliar: a more difficult piece requiring higher concentration
All pieces were different from each other, because this maximised the players’
comfort and allowed for better concentration on the music
Equipment and environment:
Musicians played while wearing the NeuroSky Mindwave
The hall was reserved for the experiments: The players were
isolated from distractions
11. Collected Data
40 minutes of audio and brain signals were recorded:
https://data.d4science.net/NDJT
Segments of EEG signals with 1s to 30s were prepared:
30 datasets with different segment cuts: 1s, 2s, .., 30s
8 features per second for each segment - averaged power spectrum in the ranges: Delta, Theta, low-Alpha, high-Alpha,
low-Beta, high-Beta, low-Gamma, high-Gamma
Training and test sets:
80%-20% rule with 10-fold cross-validation
Non-adjacent signals average cross-correlation in the 30 datasets: 0.2 [0.008;0.4] – poor intra-subject correlation
the played pieces were articulated and did not contain repeated musical sequences
Two collections of training-test sets:
1 – Overlapping: both involve signals belonging to the same musician, i.e. they included a potential intra-subject correlation.
2 – Disjoint: a leave-one-out procedure removes from the test set the musician’s signals that were involved in the training set
12. Classification Model
A multi-layer Feed-Forward ANN was used to
classify a segment of brain signal as belonging
either to a violinist or a pianist.
Pros:
FF-ANNs are suited for classification tasks when
a spectrum portion is processed as a “snapshot”
(Cutugno et al., 2005; Coro et al. 2018)
Suited when reaching the highest possible
classification performance is not required
Good for feasibility assessments like “is there
discriminant information in brain signal for
musicians’ classifications?”
Cons:
Lower performance with respect to Deep Neural
Networks
Impossible to reconstruct the analytical form of
the learned function – detected patterns remain
unknown
13. Parameters to estimate
Optimal segment length containing
discriminant information
Optimal ANN topology: number of layers
and neurons per layer
Optimal separation threshold in the ANN
output
?
?
?
?
14. Supports
• Big Data processing
• Sharing of inputs and results
• Saves the provenance of
experiments
• Supports Open Science
• Input/Out
• Parameters
• Provenance
Cloud Computing Platform
WPS
REST
Workspace
Dataminingprocessesandframeworks
Capabilities of DataMiner
25 Virtual Machines with 16 virtual CPU cores, 32GB RAM and 100GB
of storage
~ 420 algorithms provided across ~40 VREs
~ 30 contributing institutes
~ 40,000 requests per month
~ 4000 scientists/students in 44 countries using VREs
Programming languages: R, Java, Python, Fortran, Linux-compiled,
Scala, .Net
Coro, G., Panichi, G., Scarponi, P., & Pagano, P. (2017). Cloud computing in a
distributed e‐infrastructure using the web processing service standard. Concurrency
and Computation: Practice and Experience, 29(18), e4219.
15. ANN Training
The cloud computing platform tested every parameter combination:
• Segment lengths: 1s-30s
• Number of layers: 2-*
• Number of neurons: 2-*
• ANN logistic sigmoid activation function with 10 thresholds tested in [0-1]
Input: concatenated power spectrum falling in the signal window
E.g.: 20s -> 8 features*20 frames= 160 features
Output: 0 for pianists and 1 for violinists
Topology search strategy: growing approach
Training-Test executed on the Overlapping sets with different groups of musicians
Performance calculation: 10-fold cross validation
Overall ~700 000 models were trained
17. Optimal Architecture
Optimal segment length: 20s
Optimal topology:
160 input neurons
2 hidden layers:
100 neurons in the first hidden layer
20 neurons in the second
1 output neuron
Optimal classification threshold: 0.7
160 Neurons
100 Neurons
20 Neurons
1 Neuron
0.7
18. Performance on Expert Musicians only
Average Accuracy Maximum Accuracy Minimum Accuracy
Easy-familiar pieces only
65% 70% 50%
Challenging-unfamiliar pieces only
65% 75% 50%
All pieces
72.5% 87.5% 50%
Individual accuracies are comparable between them
The highest performance is reached when using all pieces
Discriminant and complementary information is contained in both types of pieces
19. Introducing a Cellist
Average Accuracy Maximum Accuracy Minimum Accuracy
All pieces – no Cellist
72.5% 87.5% 50%
All pieces – with Cellist as Violinist
80% 90% 70%
All pieces – with Cellist as Pianist
58.3% 70.8% 45%
The Cellist increases the performance when indicated as a Violinist
The ANN gets “confused” when indicating him as a Pianist
The Cellist’s signals resemble more the Violinists’ signals
Although the movements of violinists and cellists are different, neurophysiological studies
have highlighted similar cortical activation patterns in the right hemisphere, whereas higher
activity in the left hemisphere for Pianists (Elbert et al., 2005; Langheim et al., 2002)
20. Dependency on Musicians’ Expertise
Average Accuracy Maximum Accuracy Minimum Accuracy
Experts only (Cellist as Violinist)
80% 90% 70%
Expert musicians + Good musician
81.6% 91.7% 75%
Expert, Good, and Basic musicians
74.3% 85.7% 64.3%
Introducing the good musician increases the performance
Introducing the basic player reduces the performance of 8.9% relative; 7.3% absolute
Using the Disjoint training/test sets, the performance decreases from 81.6% to 79.5%
This strengthens the hypothesis that performance is poorly affected by the low intra-
subject correlation
21. Compliance with the Open Science Paradigm
Make scientific research, data and dissemination accessible to all levels of an inquiring society, amateur or professional. Support
e-Intensive Science and the three Rs of Science (Re-usability, Repeatability, Reproducibility).
We used the D4Science e-Infrastructure (www.d4science.org) as the underlying computational and OS platform:
Training set available on a distributed storage system: https://data.d4science.net/NDJT
Open source algorithms:
http://svn.research-infrastructures.eu/public/d4science/gcube/trunk/data-analysis/EcologicalEngine/src/main/java/org/gcube/dataanalysis/ecoengine/models/
Free-to-use cloud computing services and Web interfaces to train and test ANNs:
• https://services.d4science.org/group/scalabledatamining/data-miner?OperatorId=org.gcube.dataanalysis.wps.statisticalmanager.synchserver.
mappedclasses.transducerers.FEED_FORWARD_NEURAL_NETWORK_TRAINER
• https://services.d4science.org/group/scalabledatamining/data-miner?OperatorId=org.gcube.dataanalysis.wps.statisticalmanager.synchserver.
mappedclasses.transducerers.FEED_FORWARD_NEURAL_NETWORK_REGRESSOR
OS service features:
• OGC-Web Processing Service standard interface for the services
• Computational provenance tracking for each execution
• Collaboration facilities: data, results, and algorithms sharing
• Re-usability, Repeatability, Reproducibility of services and data
22. Conclusions
• It is possible to build an EEG classifier for violinists and pianists using basic equipment and modelling
• Musicians’ brain signal similarities likely depend on the played instrument and on the musicians’ expertise
We made Open Science tools available to make the experiment extensible:
A Cloud computing platform to search in a large parameters space
Services to repeat/reproduce the experiment with new data
WPS interfaces to reuse ANN across domains and software
Interfaces to enrich the platform with new models and approaches
Further steps suggested by the results:
Use more powerful equipment
Use deep neural networks
Embed neurophysiological information
Correlate musicians’ brain signals with the way sound is produced:
Cellists and violinists have a constant interaction with a continuous sound
Pianists interact with digital-like sounds.
Beta waves are correlated with logical rational processing, whereas Alpha waves are associated with intuitive processes involving internal focus of attention and with top-down sensory inhibition
Gamma -> emotional tension
Beta -> Logic processing
Alpha -> Intuition, attention and half-sleep
Theta and Delta -> deep sleep
Generally, motion artefacts are a common source of noise when using dry sensors, although most of the artefact energy is concentrated in the frequencies under 5Hz
These artefacts can be reduced by using a high-pass filter (over 0.5Hz) and a filter based on contact impedance variations
8 features overall: delta theta low Alpha high Alpha low Beta high Beta low Gamma high Gamma