Philosophers and philosophically-minded scientists have advanced several competing accounts of explanation in cognitive science. Arguably, none of these accounts captures the methods and concepts of cognitive science in all of their heterogeneous glory. This talk is an exercise in synthesis for the purposes of conceptual housekeeping. Starting with Marr’s celebrated three-level framework, I explicate concepts such as ‘computation’, ‘algorithm’, and ‘implementation’, and clarify their role in cognitive scientific explanation. To this end, I will deploy resources from other frameworks, including the recently-popular framework of mechanistic explanation.

I'll talk about a scientific problem in linguistics and a technical computational problem in machine learning and their relationship.
The scientific problem is the acquisition of syntactic structure and how this takes places during the early years of first language acquisition; and the technical problem is about machine learning of certain types of grammars, especially context free grammars.
In recent years, substantial progress has been made in the mathematical and computational theory of how these grammars can be learned just from strings:
I will review  this work and discuss how this can shed light on language acquisition and in particular on the notorious argument from the poverty of the stimulus, the role of innate biases, and related methodological issues to do with the application of such abstract mathematical ideas to scientific problems.

In the first part of this talk I will report on some recent research on the use of repetition in language and music. A corpus analysis of classical melodies shows that, when a melodic pattern is repeated with an alteration, the alteration tends to lower the probability of the pattern - for example, by introducing larger intervals or chromatic notes (notes outside the scale). A corpus analysis of written English text shows a similar pattern: in coordinate noun-phrase constructions in which the first and second phrases match syntactically (e.g. "the black dog and the white cat"), the second phrase tends to have lower lexical (trigram) probabilities than the first. A further pattern is also observed in coordinate constructions in language: the tendency towards "parallelism" (syntactic matching between the first and second coordinate phrases) is much stronger for rare constructions than for common ones (the "inverse frequency effect"). (There is some evidence for this phenomenon in music as well.) I will suggest that these phenomena can be explained by Levy and Jaeger's theory of Uniform Information Density (UID): repetition is used to smooth out the "spikes" in information created by rare events.

"Is there an Easy First bias in jazz improvisation?"
Klaus Frieler (University of Music FRANZ LISZT Weimar, Germany)

"Harmonic Syntax: A uniform framework for elaborations, substitutions, and form"
Daniel Harasim (École polytechnique fédérale de Lausanne, Switzerland)

"Towards Computational Music Theory"
Sebastian Klaßmann (University of Cologne, Germany)

"Cognitively Plausible Machine Learning for Music Signal Processing"
Hendrik Purwins (Accenture, Germany)

"Insect Foraging vs Human Speech : Any common principles in behavioral modeling?"
Panagiotis Sakagiannis (University of Cologne, Germany)

Abstracts

Is there an Easy First bias in jazz improvisation?
Klaus Frieler (University of Music FRANZ LISZT Weimar, Germany), Martin Norgaard (Georgia State University, USA), and Roger Beaty (Penn State University, USA)

Psycholinguistics research has identified several biases in speech, which are claimed to have strong influences on language production and formation (MacDonald, 2013). One of these biases is called “Easy First”, which states that some words and phrases are easier to retrieve from memory and therefore appear earlier in spoken utterances This retrieval is facilitated when words and phrases are more frequent, syntactically less complex as well as more conceptually salient to the speaker. Since jazz improvisation and speech production are both created in real time, it suggests an Easy First bias could also be present in jazz improvisation. To test this hypothesis, we carried out a corpus study based on the Weimar Jazz Database (WJD) with 456 monophonic jazz solo transcriptions (about 200,000 tones in total). We consider interval n-grams as “word” equivalents (using a range of lengths from 3 to 10) and phrases, which are annotated in the WJD, as analogue to sentences in speech production. To operationalize “easiness” in melodies, we used two approaches. First, retrieval complexity was measured as the surprise (negative logarithm of occurrence frequency) of n-grams. Secondly, a set of (simple) complexity measures were defined including interval variety, pitch variety, mean interval size, number of direction changes, and zig-zaggity as well as a certain combined measure. Looking at all n-grams at the first 15 starting positions in all phrases of all solos, revealed that indeed the complexity of consecutive n-grams is significantly increasing for all considered n-gram lengths, with effect sizes of about 0.1 to 0.5. However, investigating single phrases brought out a large range of possible trends, from negative to positive, with a slight preference for increasing complexity. To check whether this observed Easy First bias is related to real time production or simply occur due to stylistic conventions, we created a simulated copy of the WJD based on a first order Markov model for intervals with exactly the same structure in terms of number of solos, solo lengths, and phrases. The simulated data showed virtually no Easy First bias on the corpus level (effect sizes close to zero), but a similar distribution of trends on the single phrase level. This may be attributed to the fact that the raw interval distribution of jazz solos, which is dominated by movement in small steps, is partly responsible for the observed single phrase trends, as the most common “words” in jazz consists of combination of the most common intervals. In conclusion, we suggest that there is indeed an Easy First bias present in jazz solo improvisations due to a subconscious memory retrieval bias for easy material but that this process is modulated by overarching intentions such as aesthetic and virtuosic goals or stylistic constraints. For example, the Easy First bias is much more prominent in bebop solos, which rely heavily on pre-conceived patterns to create long lines, than in solos from traditional jazz, which feature short melodic licks. As this is only a very first report on this effect, we plan to investigate it in more depth in the future.

Harmonic Syntax: A uniform framework for elaborations, substitutions, and form
Daniel Harasim (École polytechnique fédérale de Lausanne, Switzerland)

Music is hierarchically structured, both in how it is perceived by listeners and how it is composed. Such structure can be elegantly captured using grammatical models similar to those used to study natural language.
In this talk, I present the application of context-free grammars for the analysis of chord sequences using examples of Jazz standards like "Afternoon in Paris". This framework of harmonic syntax provides a language to describe structure-building relations of harmonic sequences in a uniform way. It includes prototypical patterns such as II V I progressions, chord substitutions, and higher-level organization such as AABA forms. As a modeling toolbox, harmonic syntax thus proves to be valuable for music theorists as an objective means to study subjective listening experiences as well as for empirical musicologist to quantitatively study composition and music cognition.

Cognitively Plausible Machine Learning for Music Signal Processing
Hendrik Purwins (Accenture, Germany)

First, I will present (biologically inspired) sparse approximation methods for sound classification and synthesis. In Scholler & Purwins (JSTSP 2011), the shapes and lengths of the atoms of a sound dictionary are learned and a spike representation of the signal is used for sound classification.
Secondly, I will talk about a statistical music representation based on cognitive principles (unsupervised, life-long, and one-shot learning, cognitive-perceptual top-down control) that are scalable in complexity.
Applications are given for the generation of stylistically similar and musically interesting variations of a given piece of audio. (see http://tinyurl.com/zpx22mo, Marchini & Purwins 2011; Marxer & Purwins IEEE TASP, 2016).
Third, I will introduce the Musical Brain Computer Interface and present found neural correlates of selective attention to voices in polyphonic music, based on ERP analysis in a multi-streamed oddball experiment (Treder, Purwins et al., JNE 2014).
I will also give an outlook on deep learning for music analysis (Purwins et al IEEE STSP, 2019,arxiv.org/abs/1905.00078).

Insect Foraging vs Human Speech : Any common principles in behavioral modeling?
Panagiotis Sakagiannis (University of Cologne, Germany)

Can we build models of biological behavior following always the same principles regardless of species and complexity? With this question in mind, we walk through an insect foraging model and comment on insights possibly relevant to human speech.

We discuss the talks of the first day in relation to the workshop aim.

The presentation discusses convergences between computer-based systems developed around live interactive composition and situated music with processes that have emerged from models in computational neuroscience and music information retrieval. Firstly, I'll describe a conceptual point of view according to which a theory of mind can be applied to the development of interactive computation models that produced sound material in real-time. Then, new compositional trends such as the use of extended instrumental techniques is associated with the development of new interfaces for music expression, methods for sound synthesis using digital instruments, and devices for human-machine interaction. These aspects, once interconnected, might lead to new research approaches in analyzing the nature of the sound phenomena and the various possibilities of extracting information from sound pressure waves via spectral analysis. This emphasis on sound as such, combined with new technologies for its synthesis and combination in complex compositions, let creative approaches emerge that move away from composing a particular musical piece to designing a potential space of musical expression, where the particular will be defined through interaction between the musical system and its environment. It can then be speculated that innovative comparative music and language research could follow, striving for more general and process-oriented theories of mind, brain and behavior. This in turn enables comparative research on action, language, and music. Such a development would be based on the dynamics of interaction and embodiment. Overall, it is an action-oriented approach to cognition, perception, and emotion in the spirit of motor theories of cognition. These direct research on music cognition and computational modeling away from score only approaches and towards studying musical behavior and the music-making of agents. To illustrate this perspective, I present a set of live interactive compositions that experiment with such a situated aesthetics.

The human vocal apparatus can produce a vast range of frequencies that can be articulated over time, generating rhythmic and prosodic structure.  At the same time, discrete elements as words can also be transmitted, thus generating a dual structure in the sequence, which contains serially ordered elements (words or syllables, or notes) upon which is superimposed a temporal (prosodic or rhythmic) structure.
How are serial order and temporal-rhythmic structure processed in the nervous system? Inspired by the neurophysiology of recurrent loops in the primate cortex, and their massive link with the striatum, Dominey (1995), (Dominey et al 1995) developed a model of sensorimotor sequence learning. The recurrent loops render the system naturally sensitive to serial and temporal structure of sensorimotor sequences (Dominey 1998a, Dominey 1998b), and to the prosodic structure of language (Blanc et al 2003, Blanc & Dominey 2003, Dominey & Ramus 2000).  These networks are sufficiently powerful to recognize and generate grammatically structured sentences (Hinaut & Dominey 2013, Hinaut et al 2015), and narratives (Mealier et al 2017), which should have their analogs in music structure (Thompson-Schill et al 2013).
I will present research that supports the hypothesis that intrinsic recurrent connections in cortex, and plasticity in corticostriatal projections provide the neural substrate for the perception and generation of language and music.

• Blanc, J. M., Dodane, C., & Dominey, P. F. (2003). Temporal processing for syntax acquisition: A simulation study. In Proceedings of the Annual Meeting of the Cognitive Science Society (Vol. 25, No. 25).
• Blanc JM, Dominey PF. 2003. Identification of prosodic attitudes by a temporal recurrent network. Brain Res Cogn Brain Res 17: 693-9
• Dominey PF. 1995. Complex sensory-motor sequence learning based on recurrent state representation and reinforcement learning. Biol Cybern 73: 265-74
• Dominey PF. 1998a. Influences of temporal organization on sequence learning and transfer: Comments on Stadler (1995) and Curran and Keele (1993). Journal of Experimental Psychology: Learning, Memory, and Cognition, 24: 14
• Dominey PF. 1998b. A shared system for learning serial and temporal structure of sensori-motor sequences? Evidence from simulation and human experiments. Brain Res Cogn Brain Res 6: 163-72
• Dominey PF, Arbib MA, Joseph JP. 1995. A Model of Corticostriatal Plasticity for Learning Oculomotor Associations and Sequences J Cogn Neurosci 7: 25
• Dominey PF, Ramus F. 2000. Neural network processing of natural language: I. Sensitivity to serial, temporal and abstract structure of language in the infant. Language and Cognitive Processes 15: 40
• Hinaut X, Dominey PF. 2013. Real-time parallel processing of grammatical structure in the fronto-striatal system: a recurrent network simulation study using reservoir computing. PloS one 8: 1-18
• Hinaut X, Lance F, Droin C, Petit M, Pointeau G, Dominey PF. 2015. Corticostriatal response selection in sentence production: Insights from neural network simulation with reservoir computing. Brain and language 150: 54-68
• Mealier A-L, Pointeau G, Mirliaz S, Ogawa K, Finlayson M, Dominey PF. 2017. Narrative Constructions for the Organization of Self Experience: Proof of Concept via Embodied Robotics Frontiers in Psychology: Language
• Thompson-Schill S, Hagoort P, Dominey PF, Honing H, Koelsch S, et al. 2013. Multiple levels of structure in language and music  In Language, music, and the brain: A mysterious relationship, pp. 289-303: MIT Press

While we may consider ourselves to be intelligent goal-directed entities whose behaviour is the product of careful deliberation, much of our behaviour appears in fact to consist of stringing together everyday or routine activities in a more or less automatic fashion with little awareness of the individual actions involved. Tasks such as dressing, grooming and commuting are illustrative of this mode of action. At the same time, this routine behaviour may be punctuated by more deliberate goal-directed acts (e.g., searching the cupboards for a new packet of sugar when the sugar bowl is empty), either when things do not go as planned or when a situation requires special care and attention.

In this presentation I will review evidence for the view that action production involves two distinct systems – a routine/automatic system and a non-routine/deliberative system – with the non-routine system controlling action indirectly by modulating the routine system. I will then describe a computational model of routine sequential action production in which hierarchically structured action schemas compete for the control of behaviour. The model provides an account of the interaction between top-down intentions and bottom-up affordances, and has been used to explain both slips and lapses in the routine behaviour of neurologically healthy individuals when distracted and the more elaborate errors of neurological patients following various forms of brain injury. I will then illustrate how the model may be extended to support; a) error detection and recovery through monitoring mechanisms; and b) non-routine tasks by allowing deliberate top-down excitation of action schemas.

Language production may also be viewed as an everyday activity, in that for the most part we are able to participate in fluent conversations without awareness of the specific syntactic structures selected. Language production similarly involves the semantically mediated production of basic elements over time subject both to serial order constraints and higher-level hierarchical organisation. Moreover language is also subject to deliberate control when special care is required (e.g., when making a technical argument or when choice of words is otherwise critical). Given these and other parallels between action and language, I will then consider how the model might be applied in the domain of language production.

In the final part of the talk I will consider implications and extensions of the model, including: a) how the original model might be run in reverse to support event comprehension; b) how the language production model might consequently be run in reverse to support language comprehension; and c) how the model might be applied to the music production domain (specifically with regard to top-down and bottom-up interactions in both playing to music and improvisation).

We discuss the talks of the second day in relation to the workshop aim. In addition, we discuss issues raised in the eintire workshop.