The many contributions made during the past three decades to computer-assisted analysis and generation of music with the aid of Connectionist architectures can be seen to have occured in two waves, in parallel with developments in Connectionist research itself. During the first wave, the founding principles of Connectionism were introduced (Rumelhart et al., 1986) through the idea of Parallel Distributed Processing according to which mental phenomena occur as a result of simultaneous interactions between simple elementary processing units, as opposed to the then prevailing notion of Sequential Symbolic Processing which explained the same phenomena in terms of sequential interactions between complex goal-specific units. Its significance is largely theoretical, with a few experimental and empirical results to support the feasibility of the theory. Following several years of reduced interest, the second wave further strengthened the claims made by its precursor through a series of successful high-impact real-world applications. This was owing to both the proposal of newer theories, and the availability of greater computational power and vast amounts of data that enabled the demonstration of the efficacy of these theories nearly two decades on (Bengio, 2009; LeCun et al.,2012). The innovations that came about as a result of these two phases trickled down to several application domains (Krizhevsky et al., 2012; Hinton et al., 2012;Collobert et al., 2011) of which music is one (Todd and Loy, 1991; Griffith and Todd,1999; Humphrey et al., 2012). This section reviews notable contributions among the many that demonstrated the application of connectionism to symbolic music modelling during these two waves in order to present a historical perspective together with an overview of the techniques employed.

The First Wave

The first set of notable approaches which apply Connectionism to the analysis and generation of symbolic music were proposed in the years following the publication of the influential text on Parallel Distributed Processing (Rumelhart et al.,1986). While the breadth of contributions to the field during this period is indeed vast, I present a brief historical perspective only on work involving Feedforward Neural Networks, Recurrent Neural Networks and Boltzmann Machines, and refer the reader to (Rumelhartet al., 1986; Medler, 1998) for more in-depth and comprehensive reviews. Many of the inventions and algorithms proposed during this period persisted through the decades that followed and significantly impacted research in Artificial Intelligence, and the now thriving field of Machine Learning. These were the years that saw the maturation of the previously proposed Perceptron (Rosenblatt, 1958) into the Multi-Layer Perceptron (also known as the Feedforward Neural Network) and the invention of the Backpropagation algorithm (Rumelhart et al., 1988) which offered a simple and efficient means to train this model on data, thus leading to a surge in its popularity. The architecture of the Feedforward Neural Network (FNN) was further adapted to deal with sequential data into the Recurrent Neural Network(RNN) (Elman, 1990; Jordan, 1986), and likewise, the Backpropagation algorithm extended into the Backpropagation Through Time (BPTT) (Werbos, 1990) to train this new architecture. Other algorithms were also proposed around the same time to carry out real-time learning in the RNN architecture (Williams and Zipser, 1989). Another significant innovation from this period is the Boltzmann Machine family of models (Smolensky, 1986; Hinton et al., 1984), which consists of undirected graphical models that learn joint probability distributions of sets of visible and latent variables through a process of minimisation of an energy function associated with configurations of these variables. Probabilistic inference can be carried out in these models to determine conditional distributions, typically of interest in various prediction tasks.

Contributions to Connectionist theory and Artificial Intelligence, such as the above, generated interest in their adoption into several application domains that foresaw their potential benefits. This included the computer-assisted analysis and synthesis of music. One of the first systems for this purpose, known as HARMONET (Hild et al., 1992), was designed for harmonising chorales in the style of J S Bach. It consists of a symbolic (rule-based) component together with a recurrent neural network, and generates four part harmonisations of a given chorale melody. The role of the neural network is to generate human-like harmonisations within the rules dictated by music theory, which when taken literally tend to result in “aesthetically offensive musical output”. HARMONET divides the harmonisation task into three subtasks. In the first, a harmonic skeleton of the chorale melody is generated for every quarter note of the given melody (which essentially involves determining the bass voice of the chorale) using a recurrent neural network. The network takes as inputs harmonies generated at previous time-steps, and also the local context and global position (with respect to the beginning of the melody) of the note at the current time-step to generate a harmony for it. A novel representation for the pitch of each musical note was introduced at this stage which encodes the harmonic functions that contain the note, thus introducing hand-crafted musicological information as input to the network. This is followed by the generation of the alto and tenor voices taking into account the given soprano voice in the melody, and the bass voice generated in the previous step. Finally, ornamenting eighth notes are added to the result at each chord by another network which takes into account the local harmonic context. The system was evaluated by an audience of music professionals who judged the quality of the harmonisations. By treating each of the possible harmonizations of the first network above as classes and changing its output units to softmax (Specht, 1990), the system can be used for predicting harmonic expectation over time.

The work initiated in the context of HARMONET was later extended to create MELONET (Feulner and Hörnel, 1994) – a system comprised of a hierarchy of neural networks operating at different time-scales which models melodies assequences of harmony-based motifs and varies one of the chorale voices generated by HARMONET. It uses, what are known as delayed-update neurons in a recurrent network which, by integrating their inputs over a certain time-span reflectlong-term information about the melody input. It works hand-in-hand with HARMONET to generate the said variations. In a subsequent publication, a committee of such neural networks, each of which has learned a specific harmonisation style, was used to recognise different styles of harmonisation according to how expected it is to each network (Hörnel and Menzel, 1998).

Chorale harmonization has also been the focus in (Bellgard and Tsang, 1994) where, in contrast to HARMONET, the approach relies solely on a connectionist model — the Boltzmann Machine (BM) (Hinton et al., 1984; Smolensky, 1986). Four-part writing in practice is regarded here as being the result of a unique set of choices made by the composer between various competing harmonization techniques, which is not clearly defined in practice and is thus essentially an imprecise and noisy process. This is where the stochastic nature of the model employed is highlighted as an advantage over the deterministic nature of models from previous work i.e., HARMONET, CHORAL (Ebcio ̆glu, 1988). The task is viewed as one of pattern completion (or gap-filling) where a given chorale melody is only the partial specification of a complete piece of information which is the harmonized chorale. A BM learns local harmonic constraints through a series of overlapping time-windows extracted at each time-step in the chorale. Harmonization is achieved in an identical fashion, but with the learned model slid (in time) along a given chorale melody. Its visible units are comprised of a mixture of multinomial and binomial units which represent three octaves of musical pitch, musical rest and phrase-control variables. The energy-function associated with a Boltzmann machine to assess the quality of learning in the model is also used to assess the quality of the harmonies generated by the model. Sliding the BM in time along a temporal input gives, what is referred to by the authors as the Effective Boltzmann Machine (EBM).

As music is inherently temporal in nature, recurrent neural networks (RNNs) are a natural choice for modelling musical structure. In one of the first applications of neural networks to music (Todd, 1989), a special case of the RNN known as the Jordan network (Jordan, 1986) was made to memorize and interpolate between melodies of different styles. The network consists of an input, hidden and an output layer. A part of the input layer consist of a set of plan units that indicate the style of the melody being learned or produced. The rest is a set of context units which maintain the memory of the sequence produced so far by combining the effects of the most recently predicted output in the sequence (which is fed back as input) and an exponentially decreasing sum of all of the network’s previous inputs. The input layer is fully connected to the hidden layer which is in turn fully connected to the output layer. The network models sequences of pitches and durations, and uses a fixed size time-window of notes in its input context units and predicts the same number of notes as those in the input time-window for the next time-step which are fed back to its input layer. All melodies are transposed into the key of C, and a binary one-hot representation (a vector containing all 0s and a single 1 corresponding to a particular value) is used for pitch. A time-slice representation is used for duration where the length of a note is given by the number of consecutive evenly spaced time-slices (of eighth-note duration), with additional information about its onset. The purpose of the network is to memorize melodies that it has come across, associating each melody with a plan so that it can also interpolate between melodies when plans are interpolated, and change melodies dynamically as well when plans are changed.

An often cited work in connectionist music composition is that of Mozer (1991), where an RNN named CONCERT is empoyed for learning structure in melodies to generate novel variations on them. In contrast to the above described approach in (Todd, 1989), this network uses an Elman RNN (Elman, 1990) and also contains a learning stage (absent in the other) where the backpropagation through time (BPTT) algorithm (Werbos, 1990) is applied to tune the weights of the network to the prediction task. The task is to predict the next note, given the previous one and the state of its hidden layer in the most recent time-step which accounts for the notes further back in time that are not dealt with explicitly. The shortcomings of the network’s architecture in dealing with long-term memory and global structure of a musical piece are addressed by taking into account the notes in the melody at multiple time-resolutions, and also employing an additional parameter that enabled controlling its sensitivity to recent versus not so recent notes in a melody. With the generation of aesthetically pleasing melodies being the focus of the network, the task-unaware one-hot representation of notes in it is abandoned (or retained only for the sake of interpreting results) in favour of a perceptually motivated one, based on earlier empirical observations by Shepard (1982). The model was evaluated by having it extend a C major diatonic scale, learn the structure of diatonic scales, learn random walk sequences of pitches, learn specific kinds of phrase patterns and generating new melodies in the style of J S Bach.

A different approach inspired by the Target-note Technique in Bebop jazz is explored by Toiviainen (1995), wherein given a typical jazz chord progression an auto-associator network emulates the creativity of an improviser. The melodies generated by the model rely on the starting notes at any given point in time, together with the current chord to determine the possible melodic patterns, and the next chord in the progression to determine the possible target notes to follow. Several constraints that reflect the typical practices in jazz improvisation, such as the relationship between the musical pitch of a note in the melody and the root of the current chord, typical chord-types occurring in jazz progressions, typical syncopation in improvised melodies, etc. influenced the design choices for the architecture of the network. The network relies on the Hebbian learning rule for updating its connections while learning from data. A moving time-window approach was adopted for representing time, where each window corresponded to one half-measure. Thus in each step of its operation during the generative process, the network generated a melody of length equal to a half-measure, which was fed back into it in order to generate the next one, and so on. The fact that such a network learns to generate music from examples in a dataset, much like a typical jazz musician who improvises based on the repertoire that she/he has paid attention to overtime is what motivates this approach. The author concludes that “the melodies produced by the network resemble those of a beginning improviser”, based on a qualitative assessment of its generations learned from excerpts of solos played by the trumpet player Clifford Brown, over chord changes in George Gershwin’s “I’ve Got Rhythm”.

The above list of connectionist systems for the analysis and synthesis of symbolic music consists of notable contributions among those that laid the foundations for future work on the subject. It is, by no means exhaustive, and there exist several others that explore other musical phenomena with connectionist architectures considered beyond the scope of this review. I point the inquisitive reader to (Todd and Loy, 1991; Griffith and Todd, 1999) for a comprehensive summary of work carried out in the field during, what I refer to here as, the first wave of connectionism.

The Second Wave

The second wave of interest in neural networks and connectionism, which has prevailed for nearly a decade (with hardly any signs of subsiding) at the time of writing of this post, can be said to have come about towards the end of what is generally known as the AI Winter (Hendler, 2008). Its success has been attributed to the culmination of three key factors — theoretical and empirical advances in connectionist research, the presence of very powerful hardware in modern computers, and the availability of vast amounts of data. This wave brought with it several new innovations in connectionist architectures and algorithms which also fueled a revival in the study and application of older ones brought about by its precursor. The theoretically known, but often practically infeasible concept of a deep neural network (a feedforward neural network with more than one hidden layer) was made into reality during this period with the introduction of new methods for pre-training these networks layer-by-layer in an unsupervised fashion before training on a certain task in a supervised manner (Bengio et al., 2007; Hinton et al., 2012). The Restricted Boltzmann Machine (RBM), a generative unsupervised model which was, in part responsible for this turnaround, was extended in many different ways to serve as a supervised learning model and a classifier (Salakhutdinov et al., 2007; Larochelle and Bengio, 2008), a sequence learning model (Sutskever and Hinton, 2007; Sutskever et al., 2009; Taylor et al., 2007) and generalised to handle different types of data (Welling et al., 2004). The RBM, in turn, soared in popularity thanks to the Contrastive Divergence algorithm (Hinton, 2002; Tieleman,2008) which made it possible to train this model more efficiently than was previously possible. Likewise, the limitation of recurrent neural networks in modelling very long-term memory was also addressed to increase their effectiveness as sequence models (Martens and Sutskever, 2011). A previously proposed architecture to address the same issue of long-term memory — the Long Short Term Memory (LSTM) network (Hochreiter and Schmidhuber, 1997) was also revisited and is now even more widely used as a sequence model, with proposals of other models inspired by it (Chung et al., 2014). Another architecture that underwent a breakthrough is the Convolutional Neural Network which is now the de facto standard for object recognition and related image recognition and classification tasks (Krizhevsky et al., 2012). All these advances had a significant impact on three application areas — Natural Language Processing, Speech Processing and Computer Vision (Lecun et al., 2015), the very tasks in which the failure of Artificial Intelligence to perform well in the past was an important reason for a drop in interest in the field, i.e. the AI Winter.

This revival of interest in connectionist research inspired a body of work that deals with a diverse set of musical tasks using symbolic music. One such application was in modelling melodies by capturing short melodic motifs in them using a Time Convolutional RBM (TC-RBM) (Spiliopoulou and Storkey, 2011). In contrast to other RBM-based sequence models (Sutskever et al., 2009; Taylor and Hinton, 2009), the TC-RBM does not make use of any recurrent connections and relies on the idea of convolution through time over fixed-length subsequences within a window centered at each time-step (Lee et al., 2009). Furthermore, a weight-sharing mechanism which features in this model helps it achieve translation invariancealong time, which is desirable as motifs can occur anywhere in a musical piece. The approach models both the pitch and duration of notes, and uses an implicit representation of time by discretising it in eighth-note intervals. A two-fold evaluation of this model was carried out with the model on the Nottingham Folk Music Database. A qualitative evaluation involved the analysis of the latent distributed representations learned by the TC-RBM when presented with musical data in its visible layer, which were found to convey information about the scale, octave and chords. In a quantitative evaluation, the model was made to predict the next k time-steps given a fixed-length context. The prediction log-likelihood was computed approximately by sampling from the model, and the Kullback-Leibler divergence was used to determine the closeness of the model’s predictions to the empirical distribution.

As a continuation of a previously proposed probabilistic grammar based approach for generating Jazz solos known as the Impro-visor (Keller and Morrison, 2007), a Deep Belief Network (Hinton et al., 2006) (DBN, a probabilistic generative model made up of a stack of the aforementioned Restricted Boltzmann Machines) was experimented with for the same purpose (Bickerman et al., 2010). As modelling entire melodies, or solos requires dealing with long-term dependencies that are not feasible with a non-recurrent model such as the DBN, only 4-bar jazz licks (short, coherent melodies) are modelled at each time-step. As in some of the approaches outlined above, a sliding-window is used to model temporal information, with a window-size of one measure (4 beats) of the piece of music, and a step-size of 1 beat. The visible (input) layer of the DBN simultaneously modelled the joint distribution of the chromatic pitch-class, duration and onset, and octave of the melody note, and the chord underlying the melody, thus allowing the model to associate chords with various melodic features which is a key factor to consider in jazz music. The model was trained generatively using the Contrastive Divergence algorithm (Hinton, 2002; Tieleman, 2008) on a large corpus of 4-bar jazz licks. With the DBN being a stochastic generative model, novel jazz licks could be sample done beat at a time from it in generative mode. While it could be demonstrated that the model does indeed generate the desired licks, the authors conclude in favour of their previous grammatical approach to lick generation over the DBN stating the subjective quality of the generated licks and the large training time of the DBNs to support this choice.

The approaches described above use non-recurrent models which have largely been superceded, when it comes to the modelling of sequential data, by recurrent models that are a more natural fit for temporal data. In an attempt towards style-independent polyphonic music generation in (Boulanger-Lewandowski et al., 2012), an RNN-RBM is made to model sequential information directly from the piano-roll notation (Orio, 2006b). The reason for dealing with this notation is to avoid making any kind of prior assumptions regarding the nature of the modelling task that would simplify it, thus leaving much for the model to determine by itself. The RNN-RBM is a stochastic model and can be understood as a sequence of RBMs, which at each time-step of the sequence are conditioned by the hidden layer of an RNN. Thus in addition to the RNN modelling sequential information, the RBM models correlations between variables (MIDI note values) that occur simultaneously at each time-step. The latter is often ignored in standard RNNs, and can be viewed as an advantage of this model given sufficient data since it also entails the need for a greater number of model parameters. The model is targeted at the task of automatic music transcription and is thus required to model time in seconds incontrast to other symbolic music modelling approaches that represent time relative to the musical score, thus requiring an additional step of alignment between the audio and symbolic formats. Time, in this model, is represented in terms of consecutive slices of the quantised musical signal. It is trained using the mini-batch gradient descent and the Backpropagation Through Time algorithms. It was found that this model outperforms others addressing the same task. This work has also inspired other very close extensions with the same goal, that claim improved performance (Goel et al., 2014; Lyu et al., 2015).

A previous approach by Eck and Schmidhuber (2002) for modelling Blues music with a Long Short-Term Memory (LSTM) RNN can be said to have influenced the above described one (Boulanger-Lewandowski et al., 2012) in its choice to not incorporate any prior musicological information in order to simplify the modelling task. As mentioned earlier, the LSTM is an enhanced version of the basic RNN and has been shown to be able to successfully model longer temporal dependencies than the latter. Here, once again, successive slices of the musical signal are treated as time-steps. A quantisation step-size of 8 notes per measure was used, and thus the 12-bar blues musical segments used for training the model were each 96 time-steps in length. The first experiment carried out with this model involved having it learn and generate a musical chord structure, from which the authors conclude that this is a fairly straightforward task for the model, and also expected given its previous success in tasks involving counting. In the second experiment both melody and chords are learned, leading to a conclusion that the LSTM is indeed able to generate a blues melody constrained by the learned chord structure that sounds better than a random walk across the pentatonic scale and are faithful to the examples in the training set. The evaluation in this case is left to the listener who is encouraged to visit a webpage containing the pieces of music generated by the network.

A more recent study with the LSTM (Franklin, 2006) carried out further experiments with this model on jazz-related tasks. Here, various note representations were studied in order to incorporate musical knowledge into the network. This can be contrasted with the approach adopted in (Eck and Schmidhuber, 2002; Boulanger-Lewandowski et al., 2012) that avoids making any music theoretic assumptions. A pitch representation based on major and minor thirds known as the circle-of-thirds representation, and a duration representation known as the modular-duration representation which extends that proposed in (Mozer, 1991) were used to train the dual pitch/duration LSTMs. Two experiments were carried out. The first focused on short musical tasks, and only sequences of musical pitch were considered. These included outputting in sequence the four chord tones given a dominant seventh chord as input, determining whether or not a given sequence of notes are ordered chromatically, and reproducing a specific 32 note melody of the form AABA given only the first note as input. A single network was used for all these tasks. In the second experiment, which focused on long musical tasks, the objective was to learn the melody of the song Afro Blue composed by the jazz percussionist Mongo Santamaria. Two separate networks are used to learn musical pitch sequences and note duration sequences respectively. The study concludes in favour of the LSTM and a detailed qualitative analysis of the results with respect to the authors’ expectations.

Lambert et al. (2015) trained a two-layered RNN on the Mazurka dataset (MAZ), an audio dataset of expressively performed piano music. The first layer for the system is a Gradient Frequency Neural Network (GFNN) (Large et al., 2010), which uses nonlinear oscillators to model metre perception of a periodic signal. The second layer contains LSTM units which model the output of the GFNN and predict rhythmic onset as a time-series activation function. This work builds on previous experiments involving a symbolic data in which the authors find that the LSTM performs time-series modeling significantly better when GFNNs are used exclusively (Lambert et al., 2014a,b). Their GFNN-LSTM model was able to predict rhythmic onsets with an f-measure of 71.4%.

As stated before, there seems to be very little work focusing on connectionist models for information theoretic music modelling. One such attempt is presented in (Cox, 2010), where the relationship between entropy and meaning in music inspired by (Meyer, 1956, 1957) is explored with the help of Recurrent Neural Networks that estimate instantaneous entropy for music with multiple parts in the analysis of a string quartet piece composed by Joseph Haydn. The model considered here contains two components – a long-term model (LTM), and a short-term model (STM) (Conklin and Witten, 1995). The parameters of each model are learned through exposure to appropriate data. The LTM models global stylistic characteristic acquired by a listener over a longer time-span. The STM models context-specific information, available in a melody while it is being processed by the listener, in the generation of expectations. Predictions made by each modelare combined using ensemble methods, and this has been shown previously to improve the quality of predictions over individual models in the past (Conklin and Witten, 1995; Pearce, 2005). The work demonstrates that the entropies as predicted by the model are sensitive to the effects of cadences, resolutions, textural change, and interruptions in music.

References

1. Matthew I. Bellgard and C P Tsang. Harmonizing Music the Boltzmann Way. Connection Science, 6(2):281–297, 1994.
2. Yoshua Bengio, Pascal Lamblin, Dan Popovici, Hugo Larochelle, et al. Greedy Layer-wise Training of Deep Networks. Advances in Neural Information Processing Systems, 19:153, 2007.
3. Yoshua Bengio. Learning Deep Architectures for AI. Foundations and Trends® in Machine Learning, 2(1):1–127, 2009.
4. Greg Bickerman, Sam Bosley, Peter Swire, and Robert Keller. Learning to Create Jazz Melodies using Deep Belief Nets. In International Conference On Computational Creativity, 2010.
5. Nicolas Boulanger-Lewandowski, Yoshua Bengio, and Pascal Vincent. Modeling Temporal Dependencies in High-Dimensional Sequences: Application to Polyphonic Music Generation and Transcription. In International Conference on Machine Learning, 2012.
6. Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling. arXiv preprint arXiv:1412.3555, 2014.
7. Ronan Collobert, Jason Weston, Léon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. Natural Language Processing (almost) from Scratch. The Journal of Machine Learning Research, 12:2493–2537, 2011.
8. Darrell Conklin and Ian H Witten. Multiple viewpoint systems for music prediction. Journal of New Music Research, 24(1):51–73, 1995.
9. Greg Cox. On the Relationship Between Entropy and Meaning in Music: An Exploration with Recurrent Neural Networks. In Annual Conference of the Cognitive Science Society, pages 429–434, 2010.
10. Kemal Ebcio ̆glu. An Expert System for Harmonizing Four-Part Chorales. Computer Music Journal, pages 43–51, 1988.
11. Douglas Eck and Juergen Schmidhuber. Finding Temporal Structure in Music: Blues Improvisation with LSTM Recurrent Networks. In IEEE Workshop on Neural Networks for Signal Processing, pages 747–756. IEEE, 2002.
12. Jeffrey L Elman. Finding Structure in Time. Cognitive science, 14(2):179–211, 1990.
13. Johannes Feulner and Dominik Hörnel. MELONET: Neural Networks that Learn Harmony-based Melodic Variations. In Proceedings of the International Computer Music Conference, pages 121–121. INTERNATIONAL COMPUTER MUSIC ASSOCIATION, 1994.
14. Judy A Franklin. Recurrent Neural Networks for Music Computation. INFORMS Journal on Computing, 18(3):321–338, 2006.
15. Kratarth Goel, Raunaq Vohra, and JK Sahoo. Polyphonic Music Generation by Modeling Temporal Dependencies using a RNN-DBN. In Artificial Neural Networks and Machine Learning–ICANN 2014, pages 217–224. Springer, 2014.
16. Niall Griffith and Peter M Todd. Musical Networks: Parallel Distributed Perception and Performance. MIT Press, 1999.
17. James Hendler. Avoiding Another AI Winter. IEEE Intelligent Systems, 2(23):2–4,2008.
18. Hermann Hild, Johannes Feulner, and Wolfram Menzel. Harmonet: A Neural Net for Harmonizing Chorales in the Style of JS Bach. In Advances in Neural Information Processing Systems, pages 267–274, 1992.
19. Geoffrey E Hinton, Terrence J Sejnowski, and David H Ackley. Boltzmann Machines: Constraint Satisfaction Networks that Learn. Carnegie-Mellon University, Department of Computer Science Pittsburgh, PA, 1984.
20. Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8):1771–1800, 2002.
21. Geoffrey E Hinton, Simon Osindero, and Yee-Whye Teh. A Fast Learning Algorithm for Deep Belief Nets. Neural Computation, 18:1527–1554, 2006.
22. Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdel-rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara N Sainath, et al. Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups. Signal Processing Magazine, IEEE, 29(6):82–97, 2012.
23. Sepp Hochreiter and Jürgen Schmidhuber. Long Short-term Memory. Neural computation, 9(8):1735–1780, 1997.
24. Dominik Hörnel and Wolfram Menzel. Learning Musical Structure and Style with Neural Networks. Computer Music Journal (CMJ), 22(4):44–62, 1998.
25. Eric J Humphrey, Juan Pablo Bello, and Yann LeCun. Moving Beyond Feature Design: Deep Architectures and Automatic Feature Learning in Music Informatics. In International Society for Music Information Retrieval Conference, pages 403–408, 2012.
26. Michael I Jordan. Serial Order: A Parallel Distributed Processing Approach. Technical report, Institute for Cognitive Science, University of California San Diego, 1986.
27. Robert M Keller and David R Morrison. A Grammatical Approach to Automatic Improvisation. In Sound and Music Computing Conference, pages 11–13, 2007.
28. Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Neural Information Processing Systems, volume 1, page 4, 2012.
29. Andrew Lambert, Tillman Weyde, and Newton Armstrong. Beyond the Beat: Towards Metre, Rhythm and Melody Modelling with Hybrid Oscillator Networks. In Joint 40th International Computer Music Conference and 11th Sound & Music Computing conference, Athens, Greece, 2014a.
30. Andrew Lambert, Tillman Weyde, and Newton Armstrong. Studying the Effect of Metre Perception on Rhythm and Melody Modelling with LSTMs. In Tenth Artificial Intelligence and Interactive Digital Entertainment Conference, 2014b.
31. Andrew J. Lambert, Tillman Weyde, and Newton Armstrong. Perceiving and Predicting Expressive Rhythm with Recurrent Neural Networks. In 12th Sound & Music Computing conference, Maynooth, Ireland, 2015.
32. Edward W. Large, Felix V. Almonte, and Marc J. Velasco. A Canonical Model for Gradient Frequency Neural Networks. Physica D: Nonlinear Phenomena, 239(12):905–911, June 2010.
33. Hugo Larochelle and Yoshua Bengio. Classification using discriminative restricted Boltzmann machines. In International Conference on Machine Learning, pages 536–543. ACM Press, 2008.
34. Yann A LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient Backprop. In Neural networks: Tricks of the trade, pages 9–48. Springer, 2012.
35. Yann Lecun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, (521):436–444, 2015.
36. Honglak Lee, Roger Grosse, Rajesh Ranganath, and Andrew Y Ng. Convolutional deep belief networks for scalable unsupervised learning of hierarchical representations. In International Conference on Machine Learning, pages 609–616. ACM, 2009.
37. Qi Lyu, Zhiyong Wu, and Jun Zhu. Polyphonic Music Modelling with LSTM-RTRBM. In ACM Conference on Multimedia, pages 991–994. ACM, 2015.
38. David A Medler. A Brief History of Connectionism. Neural Computing Surveys, 1:18–72, 1998.
39. James Martens and Ilya Sutskever. Learning Recurrent Neural Networks with Hessian-free Optimization. In Proceedings of the 28th International Conferenceon Machine Learning (ICML-11), pages 1033–1040, 2011.
40. Leonard B Meyer. Emotion and Meaning in Music. University of Chicago Press,1956.
41. Leonard B Meyer. Meaning in music and information theory. The Journal of Aesthetics and Art Criticism, 15(4):412–424, 1957.
42. Michael C Mozer. Connectionist music composition based on melodic, stylistic and psychophysical constraints. Music and Connectionism, pages 195–211,1991.
43. Nicola Orio. Music Retrieval: A Tutorial and Review, volume 1. Now Publishers Inc., 2006a.
44. Marcus Thomas Pearce. The Construction and Evaluation of Statistical Models of Melodic Structure in Music Perception and Composition. PhD thesis, City Uni-versity London, 2005.
45. Frank Rosenblatt. The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain. Psychological review, 65(6):386, 1958.
46. David E. Rumelhart, James L. McClelland, and the PDP Research Group, editors. Parallel Distributed Processing: Explorations in the Microstructure of Cognition,Vol. 1: Foundations. MIT Press, 1986.
47. Ruslan Salakhutdinov, Andriy Mnih, and Geoffrey Hinton. Restricted Boltzmann Machines for Collaborative Filtering. In International Conference on Machine Learning, pages 791–798. ACM, 2007.
48. Roger N Shepard. Geometrical Approximations to the Structure of Musical Pitch. Psychological Review, 89(4):305, 1982.
49. Paul Smolensky. Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Vol. 1. chapter Information Processing in Dynamical Systems: Foundations of Harmony Theory, pages 194–281. MIT Press, 1986.
50. Donald F Specht. Probabilistic neural networks. Neural networks, 3 (1):109–118,1990.
51. Athina Spiliopoulou and Amos Storkey. Comparing probabilistic models for melodic sequences. In Machine Learning and Knowledge Discovery in Databases, pages 289–304. Springer, 2011.
52. Ilya Sutskever and Geoffrey E Hinton. Learning Multilevel Distributed Representations for High-dimensional Sequences. In International Conference on Artificial Intelligence and Statistics, pages 548–555, 2007.
53. Ilya Sutskever, Geoffrey E Hinton, and Graham W Taylor. The Recurrent Temporal Restricted Boltzmann Machine. In Advances in Neural Information Processing Systems, pages 1601–1608, 2009.
54. Graham W Taylor, Geoffrey E Hinton, and Sam Roweis. Modeling Human Motion using Binary Latent Variables. In Advances in Neural Information Processing Systems, pages 1345–1352, 2007.
55. Graham W Taylor and Geoffrey E Hinton. Factored Conditional Restricted Boltzmann Machines for Modeling Motion Style. In International Conference on Machine Learning, pages 1025–1032. ACM, 2009.
56. Tijmen Tieleman. Training restricted boltzmann machines using approximations to the likelihood gradient. In International Conference on Machine Learning, pages 1064–1071. ACM, 2008.
57. Peter M Todd. A Connectionist Approach to Algorithmic Composition. Computer Music Journal, 13(4):27–43, 1989.
58. Peter M Todd and D Gareth Loy. Music and Connectionism. MIT Press, 1991.
59. Petri Toiviainen. Modeling the target-note technique of bebop-style jazz improvisation: An artificial neural network approach. Music Perception, pages 399–413,1995.
60. Max Welling, Michal Rosen-Zvi, and Geoffrey E Hinton. Exponential Family Harmoniums with an Application to Information Retrieval. In Advances in Neural Information Processing Systems, pages 1481–1488, 2004.
61. Paul J Werbos. Backpropagation through time: what it does and how to do it. Proceedings of the IEEE, 78(10):1550–1560, 1990.
62. Ronald J Williams and David Zipser. A Learning Algorithm for Continually Running Fully Recurrent Neural Networks. Neural Computation, 1(2):270–280, 1989.

This above post is an excerpt from my doctoral thesis (with minor modifications so that it makes sense outside the context of the manuscript), accepted in July, 2016.