Generative music system using quantum reservoir computing

Abstract

An interactive generative music system trains a Large Music Model (analogous to a Large Language Model like ChatGPT) and uses quantum reservoir computing to generate polyphonic (16a-c) output composit

Description

The present invention relates to a system, in particular an interactive system, for generating music. In particular, the system makes use of quantum reservoir 5 computing. BACKGROUND TO THE INVENTION Recently there has been a lot of interest in “generative artificial intelligence”. In particular, the Generative Pre-trained Transformer (GPT) architecture has been used to convincingly generate text in Large Language Models (LLMs). The best 10 known example of such a system is OpenAI’s ChatGPT. These LLMs are trained on vast amounts of text data. They are essentially trained to recursively estimate the next likely token (e.g. word) to follow an input sequence. They are “pre-trained” in the sense that training happens once, to build the 15 model, which can then be used to generate text in response to prompts. There is a distinct “training” phase and a “generating” phase. These LLMs consume very large amounts of computational resource to process user input and generate output. Nevertheless, supported by powerful hardware, systems like ChatGPT are able to respond quickly enough for interactive use. 20 Similar architectures have been used to create generative music models analogous to LLMs - so, Large Music Models (LMMs). Examples include MuseNet by OpenAI and Music Transformer by Magenta. These models can generate somewhat convincing musical “compositions" in response to a prompt in the form of a short input sequence of notes. However, we note that these are 25 very short results, often no more than 3 minutes long. GPT architectures such as ChatGPT (for text), and MuseNet and Music Transformer (for music), are the current state-of-the-art which have tended to replace architectures based on recurrent neural networks (RNNs) used previously in generative Al applications. However, music requires considerably more memory and computing power than text to process, for data preparation, machine learning and generation. In broad outline the reasons for this can be understood as arising from the complexity inherent in music. In simple embodiments a piece of music could be 5 defined as what can be written in a musical score (or encoded, for example, in a MIDI file). This includes basic things like the volume, pitch and duration of each note and the sequence in which the notes occur. A musical score of course goes much further than this with markings such as staccato, marcato, tenuto, etc. on individual notes, and performance directions applying to 10 passages. However, these markings are inherently imprecise and open to interpretation by the performer - indeed that it a major part of the discipline of musical performance. Music incorporating all of this can of course be digitally encoded as audio, but this ends up being a huge amount of data in relative terms. For example, the sampling rate of 44,100 kHz at 16-bit resolution is used 15 for compact disc quality audio, so that 705,600 bits are required to encode one second. Compare this to the 56 bits required to store a 7-letter word in ASCII format. Moreover, music requires multi-levelled structural representations where, for example, a sequence of notes constitutes a melody, or a sequence of chords a riff. We shall refer to these as ‘tokens of a musical lexicon’. 20 Polyphonic music, i.e. multiple voices playing simultaneously, is exponentially more complicated computationally in terms of the number of weights which need to be updated in training and the number of calculations which need to be made in generation. Training Large Music Models (LMMs) and then using them to generate music 25 is therefore essentially unrealistic with current technology unless considerable simplifications are made, and models will not run fast enough for interactive use, limiting their utility. Quantum computing technology has the potential to handle large amounts of information at considerable speed. In due course, the technology could be 30 transformative and allow computational problems to be solved which are impossible for classical computers. However, despite rapidly progressing technology, at this time quantum computing remains at an early research prototype stage in most applications. Quantum computers today have been called “Noisy Intermediate-Scale Quantum” (NISQ) devices, which have a small number of qubits (tens or hundreds), and limited fidelity due to noise and errors in quantum operations. These devices are generally seen as a stepping stone 5 towards the development of the next phase - the “Fault Tolerant Quantum Computing” (FTQC) era - rather than industrially useful devices in their own right. Reservoir computing is a technique for computation whereby input signals are mapped into high-dimensional spaces through a fixed, non-linear system called 10 a “reservoir”. A reservoir can be a physical system, for example, a container of liquid (literally, a reservoir) where inputs are disturbances introduced onto the surface of the liquid for example by electric motors, and then the ripples on the surface form the output which can be measured and further processed by a digital computer. More generally though, the “reservoir” can be any fixed non- 15 linear system which can map inputs to high-dimensional outputs. Various types of quantum reservoir have been proposed. The most prevalent model at this time is the gate-based quantum reservoir. The reservoir in this case is a quantum circuit made up of quantum gates. These have been proven to work on current NISQ devices. 20 It is an object of the present invention to harness state-of-the-art “NISQ” quantum reservoirs to provide an interactive system for generating music. STATEMENT OF INVENTION According to the present invention there is provided an interactive system for generating music, the system comprising: 25 first input means for inputting a passage of music; an input layer of a neural network, the input layer taking input from the first input means and outputting a projection into multi-dimensional vector space; a quantum reservoir in the form of a quantum circuit comprising a plurality of quantum gates, the quantum reservoir being connected at its input to the output of the input layer; and an output layer of a neural network, the output layer taking input from the 5 output of the quantum reservoir and outputting a prediction of the next event in the passage of music, in which a second input means is provided, the second input means allowing interactive control of at least one parameter of the quantum reservoir, and in which the output from the output layer is fed back to the input layer 10 to continuously generate music. In preferred embodiments, a passage of music is a sequence of tokens of a given musical lexicon, such as for example, musical notes, which may be encoded in the MIDI format or similar. In some embodiments the passage of music may be in an audio encoding format in digital, analogue or quantum 15 codification schemes. For clarity, we describe the system in terms of musical notes as a token, but the present invention applies to any such tokens of a musical lexicon; e.g., melodies, riffs, harmonic progressions, themes, phrases, and rhythmic patterns, to cite but a few. The output may be the next note in the sequence. The next note (more 20 generally the next “event” or next token) may be defined for example as a single note of a given pitch and duration in a simple embodiment. By using a quantum reservoir, the generative network of the invention is found to be able to handle the complexity inherent in music at speeds making it suitable for interactive use. Furthermore, the combination of the first input 25 means and the second input means allows an interactive session with the generative network, in which’ the user can input a musical passage to start the generative process, , and also adjust parameters of the quantum reservoir in real time while the generation is taking place. The system may be configured to continue to generate music for a pre-set length of time, or indeed to continue to 30 generate music until stopped. In different applications, the user could for example enter a short passage on the input means, hear a generated continuation of the passage of similar length, and then enter a further passage. In another example the user could enter a short passage and then use the system to generate the rest of a piece of music. In another example the user 5 could enter a short passage and the system will just continually generate music without end, stopping only when stopped by the user. The system can be considered as a performance tool or musical instrument in its own right, with unique control by a user over the generated music. The input layer is preferably a neural network layer with fixed randomised 10 weights. In other words, the input layer does not have its weights updated during training. The purpose of the input layer is to map the discrete elements in the input passage onto a multi-dimensional vector. For example, in one embodiment the input layer maps each note to a N-dimensional vector. The input weights therefore contain N randomised values within bounds determined 15 by an input scale coefficient. For example, if the input scale coefficient is 0.5 then the input weights contain N random values between -0.5 and 0.5. In one embodiment, N = 32. Le. the input layer maps each note to a 32-dimensional vector. Unique events (notes) within the passage of music may be represented as an 20 aggregate of pitch and duration. In more complex embodiments, more properties, e.g. dynamics, timbre, etc., may be encoded in an event. Because of the random-weighted projection into the N-dimensional vector at the input layer, notes with the same pitch, but different durations, for example, are spread out in N-dimensional vector space. 25 The output layer is preferably a pre-trained neural network layer. In other words, the weights of the output layer are set in a training phase by repeatedly applying known passages of music to the network and then updating the weights in the output layer depending on an error vector (i.e. the difference between the output of the network and the known next note in the training passage). These training 30 techniques are well-known in the field of machine learning. Note that the training data is typically applied to the whole network (the input layer which connects to the quantum reservoir which connects to the output layer) to obtain the error vector, but only the output layer has its weights updated. The input layer and quantum reservoir remain fixed (unaffected by training). This very significantly reduces the computational resource required for training, and yet the network as a whole can be optimised to the training data. 5 Then, in the generation phase, the output layer remains fixed, i.e. the weights do not get updated further. The purpose of the output layer is to map quantum states at the output of the quantum reservoir to musical notes. In preferred embodiments, multiple output layers may be trained, with different levels of optimisation. I.e. some output 10 layers will be trained to very closely or exactly approximate the function determining the next notes in the training data, and some output layers will be trained to more loosely approximate that function. One way to control the level of optimisation is simply to stop the training at different times. If the number of training epochs is reduced then the loss value will tend to be greater. During 15 training, the weights can be saved at fixed intervals (or even, for example, after every single training epoch). The saved weights during training will correspond with progressively more optimised output layers. Rather than fixed intervals, in some embodiments the optimisation may be measured (for example using validation data) and weights saved at particular measured levels of 20 optimisation. A third input means may be provided to interactively control the optimisation level of the output layer. For example, in embodiments, the third input means may select between more- and less- optimised pre-trained output layers. This is done by making use of the weights that were saved during training, before 25 the output layer weights were completely optimised. Where the output layer is unoptimized or less optimised, the characteristics of the generated music tend to be less bound by the rules implied by the training data, and become heavily dependent on the hidden layer which is the quantum reservoir. Parameters of the quantum reservoir may be controlled by the 30 second input means. In particular, where the parameters of the quantum reservoir are changed from the original parameters used in the network during training, the generated music is affected in certain ways. Hence, interaction with the parameters of the quantum reservoir can be leveraged for musical creativity. In particular, in embodiments, one parameter of the quantum reservoir which 5 may be adjusted is the number of shots. The number of shots determines the accuracy of the probability distribution of outcomes from a quantum circuit measurement. During generation, changing the number of shots can be considered as tweaking the memory capacity of the model. Another parameter which can be adjusted is parameters associated with qubit 10 rotations in the quantum circuit. By altering qubit rotations in the circuit during generation, the output pattern of music is affected. This can be considered as a “temporal distortion” in the generated music. In some embodiments, multiple output layers are provided for processing the output of the quantum reservoir simultaneously. Each output layer may be pre-15 trained to generate a different voice of polyphonic music. Preferably, one of the output layers is pre-trained as described above to predict the next note in the input sequence, whereas the other output layers are pre-trained to predict the next note in accompaniment or harmony tracks. The continuation of each track is mapped in parallel, and multiple voices in polyphonic music can be realised 20 by providing multiple pre-trained output layers. Critically, each of the pre-trained output layers takes exactly the same input from the quantum reservoir, i.e. the computation done in the quantum reservoir is re-used and the complexity of the quantum reservoir does not increase with the number of voices. Only the number of output layers is increased. This means that the level of computational 25 complexity in both the training and generation stages scales linearly with the number of voices. The quantum reservoir may be a noisy quantum reservoir. In other words, there may be an external noise source. This is usually considered a disadvantage in quantum computing, but current and near-term quantum devices are generally 30 “noisy”. However, in this application the noisy nature of the quantum reservoir does not adversely affect the generation of music, and may even be considered as an advantage since the naturally occurring noise adds to the overall dynamics and computational resource of the quantum reservoir. Embodiments of the invention can be made today, using currently-available quantum computing resources. 5 The key property that makes a reservoir capable of processing sequences effectively is its ability to create a high-dimensional, non-linear, and temporal expansion of the input. The dimensionality of a quantum reservoir can be considered as 2n states for an n-qubit configuration. Therefore, increasing the number of qubits results in 10 exponentially higher dimensionality, i.e. degrees of freedom in its measured response. This means that current quantum reservoirs with fairly modest numbers of qubits provide good performance. In the quantum reservoir, the non-linear dynamics are introduced by quantum operations such as qubit rotations and entanglements. Various types of non-15 linear transformations of input can be achieved using different quantum gate operations. For example, rotation gates designated RY, conditional not gates (CX), Hadamard (H) and Pauli-X (X) gates are all familiar to the skilled person. To take time into account, the reservoir network must contain recurrence. This makes any input fed into the network reverberate for a longer period, i.e. 20 functioning as a short-term memory. The output at a given time depends on the current input and also on the fading memory of past inputs interacting with it. Hence, the network has a sense of order. Recurrence can be introduced by establishing a connection between the output and the input, within the reservoir. At any particular time step, the output of the quantum reservoir yt depends on: 25 • A memory until the previous time step • A current input xt • A current output yt . In a preferred embodiment, the quantum reservoir comprises three blocks. A first block preferably includes RY and CX gates and applies a transformation dependent on previously measured probability amplitudes h^. The second block preferably includes RY gates and applies a transformation dependent on 5 the current input xt. A third block may include for example an arbitrary assembly of H, X and CX gates and applies essentially a randomly-defined transformation. BRIEF DESCRIPTION OF THE DRAWINGS For a better understanding of the present invention, and to show more clearly 10 how it may be carried into effect, reference will now be made by way of example only to the accompanying drawings, in which: Figure 1 is a schematic showing the overall architecture of the generative system of the invention; Figure 2 is a block diagram showing the structure of the quantum reservoir 15 forming part of the generative system of Figure 1; and Figure 3 is a more detailed block diagram showing an example implementation of a quantum reservoir according to the structure of Figure 2. DESCRIPTION OF PREFERRED EMBODIMENTS Referring firstly to Figure 1, an interactive system for generating music is 20 indicated generally at 10. The system includes an input layer 12 which can accept input of a musical passage, for example played on a keyboard 15 by a user 17. The input in a simple case may be considered a linear sequence of notes, each note having a pitch and duration. The input layer 12 is a neural network layer which is designed to map discrete events in the input sequence 25 (e.g. a single note having a pitch and duration) into a multi-dimensional vector. The input layer preferably has A / weights (e.g. A / =32 in an embodiment) which are randomised between limits defined by an input scale coefficient (e.g. the weights may be randomised between -0.5 and 0.5). The input layer therefore maps a single note into an A / -dimensional vector. The input layer weights, once randomised, are fixed. In other words, the weights of the input layer are not updated during training. The input layer feeds into the quantum reservoir 14. The quantum reservoir 14 is described in more detail below, but for now can be considered as a recurrent 5 network which transforms the / V-dimensional vector from the input layer into a vector at the output of the quantum reservoir (the dimensionality of the vector being 2 to the power of the number of qubits in the system). The quantum reservoir 14, like the input layer 12, is fixed, in that the training process does not update the parameters or arrangement of the quantum reservoir 14. 10 However, parameters of the quantum reservoir can be changed during generation as will be seen. The quantum reservoir 14 feeds into output layers 16. In a simple embodiment there may be a just a single output layer (which would be 16b in Figure 1, with output layers 16a and 16c omitted). A great advantage of the architecture where 15 the hidden layers of the network are in effect provided by the quantum reservoir is that only the output layers need to have their weights updated during training. The output layers are pre-trained by applying training data (i.e. known musical sequences) to the whole network, in each case calculating a loss function at the output (i.e. a difference between the output and the known “correct” next 20 note in the training sequence), and then updating the weights in the output layer according to techniques which are familiar to the skilled person. In the embodiment shown, there is an output layer 16b which is trained to continue a melodic sequence provided at the input, but there are also further output layers 16a, 16c which are trained to learn accompaniment tracks. 25 Crucially, the output layers are connected in parallel to the output of the quantum reservoir, and the computational resource required to train the output layers scales linearly with the number of output layers, i.e. the number of voices in the polyphonic music being generated. In other words, to generate a three-track output as shown in Figure 1, the computational resource required for 30 training will be (ignoring constant overheads) three times that required to generate a one-track output. Likewise when the output layers are used in generation the resource requirement scales linearly. This compares very favourably with known deep learning techniques whereby the complexity of the network, the number of weights that need to be updated, and the computational resource required for both training and generation, scales exponentially where polyphonic output is required. 5 The output layer(s) are pre-trained. That is, the output layers are trained on a corpus of music to set weights and then remain fixed during use for generation. The weights are not updated further while being used to generate music. For continuous generation of musical sequences, a feedback loop 18 is provided so that the next note generated at the output is sent back to the input. 10 Hence the system 10 can recursively estimate the next note in a musical sequence, allowing for the continuous generation of music. In embodiments, the system can be set to generate a fixed amount of music (for example, a few bars or a whole tune), or could be set to just keep going until stopped. When the generation is stopped or interrupted, a user can introduce a new music 15 sequence with keyboard 15 and resume the generation from there. Throughout this process, the internal memory ht is preserved. An input means 20 is provided for controlling parameters of the quantum reservoir 14. In other words, characteristics of the quantum reservoir can be changed during generation of music. This allows the user further interaction and 20 creative input during generation. In particular, the number of shots executed in the quantum network can be adjusted in embodiments. Another possibility is to adjust qubit rotations, i.e. adjust the parameters of rotation gates within the quantum reservoir. A further input means 22 is provided for controlling parameters of the output 25 layer(s). In particular the level of optimisation of the output layer(s) can preferably be controlled during an interactive session. Since the output layer(s) are pre-trained, the main way in which it is envisaged that this can be done is to save multiple sets of weights as training progresses. With more training, the weights will converge to closely approximate the function implied by the training 30 data (i.e. the output layer will become more optimised). However, by saving the weights at intermediate stages, when the output layer is less optimised and more loosely approximates the function implied by the training data, the input means 22 (called an “abstractness control”) can be used to select between different sets of weights to adjust the level of optimisation of the output layer. For example, weights in a highly-optimised output layer can be changed out 5 during generation for weights which were saved at an earlier stage in the training. The output layer then becomes less optimised and the generated music will not so closely follow the patterns implied by the training data. It is found that the effects of controlling the quantum parameters via input means 20 is also enhanced with a less-optimised output layer. Therefore, the two input 10 means 20 and 22 together provide further creative input for the user 17. By providing input sequences at 15 and controlling the quantum parameters by input 20 and output layer optimisation by input 22, the user 17 can essentially use the system as a compositional aid or performance tool. It can be considered to be a musical instrument in its own right. 15 With reference to Figures 2 and 3, the arrangement of the quantum reservoir 14 in one embodiment will now be described in more detail. In Figure 2, the overall structure of the quantum reservoir is shown. The quantum reservoir is in the form of a quantum circuit which in this embodiment processes five qubits. It includes three blocks of unitary evolution of the qubits 20 through the circuit. The circuit is initialised with qubits set to state |0). In the first block a unitary evolution based on previously measured probability amplitudes is applied. In the second block U(xt), a unitary evolution based on current input xt is applied, and in the third block [ / ( / ?) there is a unitary evolution based on a set of random parameters (or based on an arbitrary 25 arrangement of non-parameterised gates). In each unitary block, the qubits can be rotated and entangled using RY and CX gates, for example. In the first block the rotation parameters depend on previously measured amplitudes ht^ whereas in the second block the rotation parameters depend on the current input xt. The third block may include H, X and CX gates, or in some 30 embodiments may include RY gates similar to the first and second blocks, but with random parameters p. Note that xt, at a given timestep t, is an A / -dimensional vector, recalling that the input layer transforms the input into an / V-dimensional vector. Likewise, ht at a given timestep is a vector of dimensionality determined by the number of qubits in the quantum circuit. In this example there are five qubits and therefore the 5 dimensionality of ht is 25 = 32. This happens to be the same as the chosen example value for N in this embodiment, but this does not have to be the case. As discussed above, the second input means 20 may control parameters of the quantum circuit. Preferably, some or all of the rotation parameters, i.e. the parameters of the RY gates, are adjusted by the second input means. 10 The output of the quantum reservoir ht is a probability distribution over the of possible states of the set of qubits (the 25 possible states of five qubits in this case). This distribution is measured by executing the quantum circuit multiple times and recording the measured output qubits each time. The number of measurements is referred to as the number of “shots”. The number of shots is 15 another parameter of the quantum reservoir which may be controlled in real time by the second input means 20. Figure 3 shows the quantum reservoir of Figure 2 in more detail, in a simple embodiment where input is encoded by the input layers into a 32-dimensional vector, and in which a 5-qubit quantum circuit is used. It is seen that in the first 20 block, a series of rotation (RY) gates are provided in the circuit, in an arrangement together with CX gates (CX gates are shown with the symbol in the Figure). The arrangement is shown in Figure 3, and is a repeating sequence of RY gates with an ht parameter across the qubits. So RY(h0) rotates qubit qQ from its initial state which is then entangled with q± in a CX gate, 25 which then is rotates by RY^J and entangled with q2, and so on. After q^ is rotated with RY(h4) it is entangled with q3, and then the sequence starts again with q0 being rotated with RY(h5), and so on as shown in Figure 3. Each of the RY gates has a previously measured probability amplitude h as a parameter. All the measured probability amplitudes are at the previous timestep 30 t - 1 and the subscripts in Figure 3 are the different components of the output vector at that timestep, i.e. hi represents the probability of state i at timestep t - 1. In this embodiment there are 5 qubits and so 32 possible states, each possible state at a timestep having a probability amplitude represented by h0 through h31. In the second block rotation (RY) gates are arranged, in this 5 embodiment without CX gates, in a somewhat orderly way, each RY gate in the second block having an element of the input vector xt as a parameter. Since the input layer in this embodiment projects the input event into 32-dimensional space, there are 32 elements of the input vector x at timestep t, shown in Figure 3 by the subscripts x0 through x31. Note that the choice of arrangement 10 of the RY gates in the second block is somewhat arbitrary. In principle the main requirement is that the setup must create a non-linear transformation of the input. CX gates could be included in the second block in some embodiments. The arrangement of gates may be optimised by experimentation, but the arrangement shown in Figure 3 is found to be workable. 15 Note that, although not specifically shown in Figure 3, any of the parameters may be adjusted by the second input means 20. The third block is a fixed arrangement of X, H, and CX gates. In other embodiments the third block may include parameterised rotation gates RY with random parameters , and variations thereof. The output is a probability 20 distribution over possible states of the five qubits in the circuit, as explained above. The system of the invention provides a generative network which can be used to create music. The network is “trained” using techniques familiar in the field of machine learning, using a corpus of training data (typically, existing music 25 composed by humans). The training phase aims to adjust the weights of the network so that it can accurately estimate the next note in a sequence to create a coherent musical passage. In other words, the network is trained so that it learns the rules of musical sequencing from examples in a given corpus. It can then use these rules to generate new music. However, compared to known 30 machine learning techniques which make use of large scale neural networks with large numbers of nodes and trainable parameters, the quantum reservoir approach brings a significant reduction in trainable parameters for equivalent learning. Furthermore, the ability to adjust the non-trained parameters in the quantum reservoir in real time, while music is being continuously generated, provides a creative tool for musicians and in effect a novel musical instrument. Also, the architecture of multiple parallel output layers provides for polyphonic 5 music generation without the exponential increase in computational complexity which has been associated with achieving polyphony in known neural-network based generative systems. The measured states at the output of the quantum reservoir are re-used multiple times with different output layers connected to the quantum reservoir in parallel. 10 Further, by saving weights at multiple points during training, the level of optimisation of the output layers can be adjusted during generation. Again, the relatively small number of trained parameters makes it practical to do this in real time, during continuous generation of musical sequences. The quantum reservoir used in the system is feasible (and has been 15 demonstrated) with current technology, i.e. NISQ devices. It does not rely on quantum algorithms which fail by qubit decoherence or noise. Rather, the system of the invention harnesses quantum dynamics to function essentially as a fixed (non-trained) layer in a neural network. The embodiments described above are provided by way of example only, and 20 various changes and modifications will be apparent to persons skilled in the art without departing from the scope of the present invention as defined by the appended claims. . >

Claims

1. An interactive system for generating music, the system comprising: . >. . ■first input means for inputting a passage of music;i ■i ■an input layer of a neural network, the input layer taking input from the5 first input means and outputting a projection into multi-dimensionalvector space;a quantum reservoir in the form of a quantum circuit comprising a plurality of quantum gates, the quantum reservoir being connected at its input to the output of the input layer;10 an output layer of a neural network, the output layer taking input fromthe output of the quantum reservoir and outputting a prediction of the next event in the passage of music,in which a second input means is provided, the second input means . allowing interactive control of at least one parameter of the quantum15 reservoir,and in which the output from the output layer is fed back to the input layer to continuously generate music.

2. An interactive system for generating music as claimed in claim 1, in which the input layer is a neural network layer with fixed weights.20 3. An interactive system for generating music as claimed in claim 1 or claim 2, inwhich the output layer is a pre-trained neural network layer, having weights I . •determined by training on training data.

4. An interactive system as claimed in claim 3, in which a third input means is provided for interactively adjusting the weights in the output layer to change the25 level of optimisation of the output layer.

5. An interactive system as claimed in any of the preceding claims, in which a parameter of the quantum reservoir which can be adjusted by the second input means includes the number of shots.

6. An interactive system as claimed in any of the preceding claims, in which a parameter of the quantum reservoir which can be adjusted by the second input means includes a parameter associated with a qubit rotation in the quantum reservoir.5 7. An interactive system as claimed in any of the preceding claims, in whichmultiple output layers are provided, each output layer being fed with the same vector from the quantum reservoir and the output layers being configured to map the quantum reservoir output in parallel, the output of each output layer being associated with a different voice to generate polyphonic music.10 8. An interactive system as claimed in claim 7, in which each of the output layersis pre-trained to generate a different voice in polyphonic music.

9. An interactive system as claimed in any of the preceding claims, in which the quantum reservoir is a noisy quantum reservoir.

10. An interactive system as claimed in any of the preceding claims, in which the 15 quantum reservoir is a recurrent quantum reservoir.

11. An interactive system as claimed in any of the preceding claims, in which the quantum reservoir includes a first block which applies a unitary transformation dependent on previously measured probability amplitudes at the output of the reservoir, and a second block which applies a unitary transformation dependent20 on the current input.

12. An interactive system as claimed in claim 11, in which the first block includes quantum rotation gates parameterised by the previously measured probability amplitudes.

13. An interactive system as claimed in claim 11 or claim 12, in which the second 25 block includes quantum rotation gates parameterised by the current input.

14. An interactive system as claimed in claim 12 or claim 13, in which the first block further includes CX gates.Application No: GB2409081.3Examiner:Dr Mark LewneyClaims searched: 1-14Date of search: 4 December 2024Patents Act 1977: Search Report under Section 17Documents considered to be relevant:Category Relevant to claims Identity of document and passage or figure of particular relevance Y Y 1-14 1-14 https: / / physicsworld.com / a / can-we-use-quantum-computers-to-make-music / 28th February 2023 PHILIP BALL: Can we use quantum computers to make music? April 2023 Phys. World 36(4) 26 DOI 10.1088 / 2058-7058 / 36 / 04 / 27 https: / / www.nature.com / articles / s41534-023-00734-4 DUDAS et al: "Quantum reservoir computing implementation on coherently coupled quantum oscillators", npj Quantum Info 9, 64 (2023). https: / / doi.org / 10.1038 / s41534-023-00734-4Categories: X Document indicating lack of novelty or inventive step A Document indicating technological background and / or state of the art. Y Document indicating lack of inventive step if combined with one or more other documents of same category. P Document published on or after the declared priority date but before the filing date of this invention. & Member of the same patent family E Patent document published on or after, but with priority date earlier than, the filing date of this application.Field of Search:Search of GB, EP, WO &US patent documents classified in the following areas of the UKCX :Worldwide search of patent documents classified in the following areas of the IPC_____________G06N; G10H_______________________________________________The following online and other databases have been used in the preparation of this search report WPI, EPODOC, INTERNET, INSPECInternational Classification:Subclass Subgroup Valid From G10H 0001 / 00 01 / 01 / 2006 G06N 0010 / 20 01 / 01 / 2022

Images