Complaint addressal with neural networks trained using quantum transfer learning

Abstract

A system and method for addressal of complaints. The method uses neural networks (NN) trained using quantum transfer learning wherein a training dataset includes labels and input features (e.g. natura

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of computer science. In particular, the present disclosure provides a system and a method for efficient addressal of complaints with neural networks trained using quantum transfer learning. BACKGROUND

[0002] Support services of many establishments receive significant numbers of queries and complaints. For instance, manufacturers, such as those of vehicles, may receive queries on customer warranty support, internal quality checks, complaints on one or more services, and the like. Addressing queries and issues associated with quality is often a priority for such establishments. Such queries and complaints may be received via quality reporting tools in natural language directly from customers. In many cases, the same complaint may be expressed in a plurality of ways, thereby requiring a human operator to manually classify and forward complaints to appropriate process units. For instance, if there is an oil leak due to a manufacturing defect in an engine of a vehicle, customers may report it as “the car is leaking oil,” “there is a puddle of oil under the car,” or “there is a strange smell coming from the engine.” In such cases, the human operator may have to infer that the defect lies with the engine, and accordingly intimate the manufacturing plant that is manufacturing the defective engines.

[0003] With ever increasing volume of complaints, the use of human operators becomes infeasible and error-prone, and necessitates automated systems therefor to receive, interpret, and appropriately forward the complaint to the appropriate process unit for resolution. Further, with rapid increase in the number of features introduced into products, it becomes difficult to accurately identify root causes of the complaints, and forward the complaints to the appropriate department for resolution. While advancements in natural language processing provide some respite to the aforementioned problems, they do not provide acceptable performance and consume excessive computational resources when handling large volumes of data, particularly when new classifications are added at increasing volumes. It is computationally expensive and time consuming to repeatedly retrain the models whenever new features are added to manufactured products. Furthermore, the complexity of the models increases as more features are added to the manufactured product, and as combinations of associations between the process units and the manufactured products increase. [00041 Some solutions for classifying complaint labels in the context of bank clients includes adding noise, and subsequently replacing one or more words from misclassified complaints to improve quality of sample generation.

[0005] Other solutions for complaint workorder identification in electric power customer service include formatting and vectorizing the complaint data, and identifying suspected complaints based on similarity with historical data of electrical power complaints and the prescribed workorders, and classifying the identified suspected complaints.

[0006] Another solution for classifying sentences discloses a method for generating pattern features from the sentences, and training a classifier to classify sentences therewith. The pattern feature is based on support of a generalized sentence of a sentence to be a generalized sentence pattern of sentences within training data.

[0007] Still another solution discloses a distributed register for data sharing in aviation that receives data from multiple sources, and cleans and clusters the received data for processing thereof using machine learning.

[0008] Other solutions disclose text inclusion recognition method based on Bidirectional Encoder Representations from Transformers (BERT) that pretrain and finetune BERT models in three stages so as to allow the model to better learn text implications and enhance task related knowledge thereof.

[0009] However, such solutions do not provide for a solution for computational time and cost-efficient means for training, both for pre-training and fine-tuning, neural networks in applications requiring processing of increased volume and variety of data.

[0010] Therefore, there is a need for a system and a method for processing data from a plurality of data sources. Particularly, there is a need for a solution for efficient addressal of complaints with neural networks trained using quantum transfer learning. SUMMARY

[0011] Aspects of the present disclosure relate to the field of computer science. In particular, the present disclosure provides a system and a method for efficient addressal of complaints with neural networks trained using quantum transfer learning.

[0012] An aspect of the present disclosure pertains to a system for efficient addressal of complaints with neural networks trained using quantum transfer learning. The system retrieves a training dataset having a plurality of entries, each entry including a plurality of input features having a corresponding set of labels. The plurality of entries in the training dataset are received from a plurality of data sources. The system pre-processes and embeds the plurality of input features into a variational circuit having a plurality of quantum gates corresponding to one or more weights of a neural network. The system determines one or more output values using the variational circuit, based on the input features, the weights, and the quantum gates. The system determines a loss value by comparing the corresponding set of labels of the input features with the one or more output values, and update the weights based on the loss value.

[0013] In an aspect, the variational circuit may include, a plurality of qubits corresponding to the plurality of input features, the plurality of input features being embedded into a corresponding qubit from the plurality of qubits. The variational circuit may include the plurality of quantum gates, may include, a plurality of two-qubit gates that each apply a first quantum transformation to a pair of qubits from the plurality of qubits, and one or more parameterized gates that each apply a second quantum transformation to the plurality of qubits based on a corresponding parameter, where the corresponding parameter corresponds to the one or more weights of the classical neural network.

[0014] In an aspect, the plurality of two-qubit gates of the variational circuit may be configured such that each pair of adjacent qubits from the plurality of qubits may be entangled.

[0015] In an aspect, the one or more classical processors may be configured to manipulate dimensions of the plurality of input features to correspond to a number of the plurality of qubits in the variational circuit by applying a linear transformation to the plurality of input features.

[0016] In an aspect, to embed the plurality of input features, the variational circuit may be configured to, apply a Hadamard Gate corresponding to each of the plurality of qubits to initialize input states of each of the plurality of qubits, and apply a corresponding parameterized gate from the one or more parameterized gates having the plurality of input features as the parameter therefor to the corresponding qubit.

[0017] In an aspect, to determine the one or more output values, the one or more classical processors may be configured to measure an expectation value of each of the plurality of qubits of the variational circuit.

[0018] In an aspect, to determine the one or more output values, the variational circuit may be configured to, apply a Pauli operator on the plurality of qubits, and determine an expectation value of each of the plurality of qubits, wherein the expectation values may be indicative of the one or more output values.

[0019] In an aspect, the plurality of input features and the corresponding set of labels may be indicative of natural language tokens. To pre-process the plurality of input features, the one or more classical processors may be configured to convert the plurality of input features into single sentences.

[0020] In an aspect, the plurality of input features and the corresponding set of labels may include one or more natural language tokens from different data sources.

[0021] In an aspect, the classical neural network may be a pre-trained neural network, and wherein the one or more classical processors may be configured to update a subset of the one or more weights in one or more layers of the pre-trained neural network.

[0022] In an aspect, a method for efficient addressal of complaints with neural networks trained using quantum transfer learning may include retrieving, by one or more classical processors, a training dataset having a plurality of entries, each entry may include a plurality of input features having a corresponding set of labels, where the plurality of entries in the training dataset may be received from a plurality of data sources. The method includes pre-processing, by the one or more classical processors, the plurality of input features, and embedding, by the one or more classical processors, the plurality of input features into a variational circuit having a plurality of quantum gates corresponding to one or more weights of a classical neural network. The method includes determining, by the one or more classical processors, one or more output values using the variational circuit, based on the plurality of input features, the one or more weights, and the plurality of quantum gates. The method includes determining, by the one or more classical processors, a loss value by comparing the corresponding set of labels of the plurality of input features and the one or more output values, and updating, by the one or more classical processors, the one or more weights based on the loss value.

[0023] hi an aspect, the plurality of input features and the corresponding labels may be indicative of natural language tokens. To pre-process the plurality of input features, the method may include converting, by the one or more classical processors, the plurality of input features into single sentences.

[0024] In an aspect, the plurality of input features and the corresponding set of labels may include one or more natural language tokens from different data sources.

[0025] In an aspect, the classical neural network may be a pre-trained neural network, and wherein the method may include updating, by the one or more classical processors, a subset of the one or more weights in one or more layers of the pre-trained neural network.

[0026] In an aspect, the present disclosure relates to a non-transitory computer-readable medium which may include processor-executable instructions that implement the system and the method of the present disclosure.

[0027] Various objects, features, aspects and advantages of the inventive subject matter will become more apparent from the following detailed description of preferred embodiments, along with the accompanying drawing figures in which like numerals represent like components. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings are included to provide a further understanding of the present disclosure, and are incorporated in and constitute a part of this specification. The drawings illustrate exemplary embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0029] FIG. 1 illustrates an example architecture diagram of a system for efficient addressal of complaints with neural networks trained using quantum transfer learning, according to embodiments of the present disclosure.

[0030] FIG. 2 illustrates an example block diagram of one or more classical processors of the system, according to embodiments of the present disclosure.

[0031] FIG. 3 illustrates an example representation of a variational circuit of the system, according to embodiments of the present disclosure.

[0032] FIG. 4 illustrates an example implementation of the system, according to embodiments of the present disclosure.

[0033] FIG. 5 illustrates an example flow chart of a method for efficient addressal of complaints with neural networks trained using quantum transfer learning, according to embodiments of the present disclosure.

[0034] FIG. 6 illustrates an example computer system in which or with which embodiments of the system may be implemented, according to embodiments of the present disclosure. DETAILED DESCRIPTION

[0035] The following is a detailed description of embodiments of the disclosure depicted in the accompanying drawings. The embodiments are in such details as to clearly communicate the disclosure. However, the amount of detail offered is not intended to limit the anticipated variations of embodiments; on the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the scope of the present disclosures as defined by the appended claims.

[0036] Embodiments explained herein relate to the of computer science. In particular, the present disclosure provides a system and a method for efficient addressal of complaints with neural networks trained using quantum transfer learning.

[0037] The present disclosure provides a system and a method for efficient addressal of complaints with neural networks trained using quantum transfer learning. The system retrieves a training dataset having a plurality of entries, each entry including a plurality of input features having a corresponding set of labels. The plurality of entries in the training dataset are received from a plurality of data sources. The system pre-processes and embeds the plurality of input features into a variational circuit having a plurality of quantum gates corresponding to one or more weights of a neural network. The system determines one or more output values using the variational circuit, based on the input features, the weights, and the quantum gates. The system determines a loss value by comparing the corresponding set of labels of the input features with the one or more output values, and update the weights based on the loss value.

[0038] Referring to FIG. 1, architecture diagram of a system 100 for efficient addressal of complaints with neural networks trained using quantum transfer learning, according to embodiments of the present disclosure is shown. As shown the system 100 may include a quantum processor 102, one or more classical processor(s) 104, a database 106, and one or more data sources, such as first and second data sources 108-1, 108-2 (collectively referred to as data sources 108). In some embodiments, the first data source 108-1 may be configured to receive a first set of data provided by a user 110, and the second data source 108-2 may be configured to receive a second set of data provided by an operator 112 of the system 100. However, it may be appreciated by those skilled in the art that the system 100 may also include other data sources 108 from which the system 100 receives data.

[0039] In some embodiments, the data sources 108 may be indicative of software and / or hardware elements that generate and transmit data to the classical processors 104. The classical processors 104 may be configured to generate a training dataset based on the first set of data received from the first data source 108-1, and a second set of data received from the second data source 108-2.

[0040] In some embodiments, the data sources 108 may be indicative of servers in communication with a plurality of including, but not limited to, user devices, Intemet-of-Things (loT) devices, automated process units, sensors, embedded hardware, smart devices or products such as vehicle, smart refrigerators, television, and like appliances. In some embodiments, at least one subset of data sources 108, such as the first data source 108-1, may be configured to receive the first set of data from the users 110. In such embodiments, the first set of data received from the first data source 108-1 may be indicative of queries and complaints raised by the users 110. In some embodiments, the users 110 may raise queries and complaints by transmitting a plurality of input features from using corresponding user devices. In some embodiments, the user devices may include, but not be limited to, phones, tablets, desktop computers, laptops, interfaces vehicles or hardware appliances such as refrigerators, washing machines, television, smart cooking appliances, air conditioners, lighting units, and the like, hi some embodiments, the queries and complaints may be transmitted through any one or combination of including, but not limited to, a set of natural language texts, images, videos, audio recordings, and the like. In an example, the user 110 may type their complaint in a text box provided on an interface associated with the user device, and submit the natural language text to cause the interface to transmit the complaint to the first data source 108-1. In other examples, the complaint may include a photo or a video recording describing a problem with the user 110’s vehicle. The first set of data may be transmitted as data packets to the first data source 108-1 using the user device.

[0041] hi some embodiments, the at least one of the data sources 108, such as the second data source 108-2, may be configured to receive the second set of data from the operators 112. In some embodiments, the second set of data received from the operators 112 may be indicative of data associated with including, but not limited to, development data, manufacturing data, and testing data of a product. In some examples, the operators 112 may provide development data including, but not limited to, digitized blueprints, prototype design data, 3D models of the manufactured product, documentation of the manufacturing process of the products, transcripts of design events / meetings, and the like. In some examples, the operators 112 may also provide the manufacturing data including, but not limited to, data collected by loT devices associated with one or more process units in manufacturing plants, such as duration of annealing of a metal component of a vehicle, and the like. In further examples, the operators 112 may provide the testing data associated with quality control tests performed on the manufactured product, such as performance metrics of the vehicle’s engine, stress test of the vehicle body, safety tests, and the like. In some embodiments, the second data source 108-2 may be provided with second set of data manually by the operator 112. In other embodiments, the second data source 108-2 may be configured to receive the second set of data from other devices that generate the development, manufacturing, and testing data.

[0042] In some embodiments, the classical processors 104 may be configured to generate the training dataset based on the first set of data and the second set of data from the first and second data sources 108-1,108-2. In some embodiments, the classical processors 104 may be configured to convert the first set of data and the second set of data into a plurality of entries having the input features assigned with a corresponding set of labels. In some embodiments, the training dataset may be curated by the operators 112 manually. In other embodiments, the classical processors 104 may be configured to assign the set of labels to the input features based on a predefined set of rules. In some embodiments, the labels may be indicative of a binary classification. In other embodiments, the labels may be indicative of a multi-class classification. In yet other embodiments, the labels may be indicative of one or more natural language tokens extracted from the input features. In some embodiments, the labels may correspond to natural language replies or responses to queries or complaints raised by the user 110. In further embodiments, the labels may be indicative of natural language tokens associated with the second set of data, with said natural language tokens being used to classify on the input features. In some examples, the clustered natural language tokens of the input features may allow the operator 112 to associate the complaint text with the appropriate process unit from the manufacturing plant, thereby allowing the operator 112 to analyse and resolve the user 110’s complaint. The training dataset may be used for training a neural network.

[0043] hi some embodiments, the neural network may be a multi-model network adapted to process input features in a plurality of form factors. In such embodiments, the input features may have any one or combination of including, but not limited to, images, natural language texts, videos, audio recording, and the like. In other embodiments, the neural network may be adapted to perform any natural language processing task, including, but not limited to, text classification, named entity recognition, mask filling, translating, text generation, summarization, and the like.

[0044] In some embodiments, the neural network may have a set of weights associated therewith. The weights may be indicative of integers or float point numbers. The weights of the neural network may be arranged in one or more layers, each layer being adapted to receive and process outputs of a preceding layer. In some embodiments, the neural network may be indicative of a pre-trained model, where the pre-trained model is configured to perform a predetermined functionality. In an example, a pre-trained encoder-only transformer model may be trained to classify natural language inputs into any one of a predefined set of labels. In an example, the pretrained neural network may be a Bidirectional Encoder Representations from Transformers (BERT) model. The BERT model may be finetuned to extract the one or more keywords from the training dataset. Pre-trained models may have the layers thereof trained and updated to perform the predetermined function. In some embodiments, pre-trained models may be finetuned for a specific use case, by updating weights of a subset of the layers of said pre-trained models. In some embodiments, the pre-trained models may be finetuned using the training dataset generated from the data sources 108.

[0045] Given the increasing volume of complaints and the increased rate of adding new features to manufactured products, the neural networks may be finetuned periodically. In some embodiments, the neural networks may be fine-tuned in real-time when the neural network is adapted to cluster the input features using unsupervised techniques, hi other embodiments, the neural network may be finetuned periodically at predefined intervals. In yet other embodiments, the neural network may be fine-tuned when a performance metric, such as accuracy, precision, recall, or a combination thereof, but not limited thereto, falls below a predetermined threshold.

[0046] hi some embodiments, the neural network may be trained, including but not limited to through pretraining and / or fine-tuning, by the classical processors 104. However, training the neural network using only classical processors 104, such as in conventional solutions, may be computationally expensive and infeasible, when processing the volume and variety of data received from the first and second data sources 108-1, 108-2. Alternatively, the system 100 may use a variational circuit, such as the variational circuit 300 of FIG. 3, to partly or fully train the neural network. The variational circuit 300 may allow the system 100 to process the variety and volume of data received from the data sources 108, thereby reducing training time, computational expense, and performance.

[0047] In some embodiments, the variational circuit 300 may be simulated on the classical processors 104. In other embodiments, the variational circuit 300 may be implemented on the quantum processor 102. The variational circuit 300 may be configured to initialize, manipulate, measure, or change states of a plurality of qubits. In some embodiments, the quantum processor 102 may implement the variational circuit 300 to manipulate the plurality of qubits. In some embodiments, the quantum processor 102 may be any one or combination of including, but not limited to, an adiabatic quantum computer, gate model quantum computer (GMQC), cluster state quantum computer (CSQC), measurement-based quantum computer (MBQC), a topological quantum computer, or the like.

[0048] In some embodiments, the variational circuit 300 may include a plurality of quantum gates. The quantum gates may be indicative of quantum transformations performed on the qubits. In some embodiments, the quantum processor 102 may translate the quantum gates and the arrangements thereof by inducing including, but not limited to, voltages, signals, frequencies, and the like, to manipulate the states of the qubits based on form factor thereof. In some embodiments, the quantum processors 102 may physically realize the qubits as any one or more of: an electron in one of the two orbits around the nucleus of an atom (either in ground state or excited state), a photon in either of the two polarized states, a subatomic particle having either a clockwise or an anticlockwise spin, quantum particle in a potential well (such as in a particle in a box model), quantum dots indicative of electrons trapped in quantum boxes in miniaturized semiconductors, Josephson junction having any one of superconducting charge, flux, or phase qubits, Nitrogen-vacancy centres in crystals having charge and spin states, or the like, in other embodiments, the qubits may be physically realized using any quantum mechanical model having two or more identifiable states, said states being denoted by |0> and |1>, and capable of being in a superposition of states. While the present disclosure has been described in the context of CSQC, it may be appreciated by those skilled in the art that the embodiments of the present disclosure may be suitably adapted for implementation in other quantum computers.

[0049] In some embodiments, the quantum gates may be configured to manipulate the qubits. The quantum gates may perform quantum transformations on the qubits to change the states thereof. The quantum transformations may be performed once the input features are embedded into the qubits. In some embodiments, quantum transformations may include, but not limited to, performing rotations, entanglements, or other operations to manipulate the state of the qubits. The quantum gates of the variation circuit 300 indicate the quantum transformations to be applied to the qubits. Physical realizations of the quantum gates for performing the quantum transformations may be suitably adapted based on the form factor of the qubits. In some embodiments, the qubits, and the states thereof, may be represented as a vector having a set of values, such as in embodiments simulating the variational circuit 300. In such embodiments, the quantum transformations may be indicative of matrix transformations performed on the vectors corresponding to the qubits.

[0050] hi some embodiments, the quantum gates may include a plurality of two-qubit gates that each apply a first quantum transformation to a pair of qubits from the plurality of qubits, hi some embodiments, the first quantum transformation may be indicative of quantum entanglement operation. In some embodiments, the plurality of two-qubit gates of the variational circuit 300 may be configured to entangle each pair of adjacent qubits from the plurality of qubits. In some embodiments, each qubit may be alternatively entangled with either a succeeding qubit or a preceding qubit from the plurality of qubits at each depth / time step. In some embodiments, the quantum transformations may be performed by applying unitary matrices to the vector representing the state of the qubits. In an example, the qubits may be entangled by applying a controlled-NOT (CNOT) gate, which entangles two qubits by flipping the state of a second qubit depending on the state of a first qubit.

[0051] In some embodiments, the quantum gates may include one or more parameterized gates that each apply a second quantum transformation to the plurality of qubits based on a corresponding parameter. In some embodiments, the corresponding parameter may correspond to the one or more weights of the neural network. In some embodiments, the parameterized gates may be indicative of logical operators applied to the qubits. In some embodiments, the second quantum transformation may be indicative of single qubit rotation operations. The single qubit rotation operations may include, but not be limited to, rotations around the x, y, or z axis. The single qubit rotation operations may be performed by any one or more of including, but not limited to, RZ, RX, or RY gates. In such embodiments, the RZ, RX, or RY gates may receive the corresponding parameter indicative of one or more angles, such as, but not limited to, a rotation angle. The rotation angle may be any angle between 0 and 2%. The rotation angle may be determined based on the desired rotation of the qubit state. In some embodiments, the weights of the neural network may be provided as the rotation angle for the parameterized gates. In such embodiments, the weights may be standardized to have a value between 0 and 2k.

[0052] In some embodiments, the depth, and the number of parameterized gates of the variational circuit 300 may be determined based on the number of qubits. In some embodiments, the depth of the variational circuit 300 may be equivalent to the number of weights divided by the number of qubits in the variational circuit 300.

[0053] For training the neural network, in some embodiments, the classical processors 104 may retrieve the training dataset having a plurality of entries, each entry including the plurality of the input features having the corresponding set of labels. In some embodiments, the plurality of entries in the training dataset may be received from the data sources 108. In some embodiments, the input features and the corresponding labels may include one or more natural language tokens from different data sources 108. In some examples, the input features may be indicative natural language complaints from the users 110, and the labels may be indicative of identifiers associated with the process units, the complaints being mapped to the process units, thereby allowing the operators 112 to identify causes of and resolve the complaints.

[0054] In some embodiments, the classical processors 104 may be configured to pre-process the input features. In some embodiments, pre-processing the input features may include, but not limited to, stop-word removal, punctuation removal, symbol removal, number removal, removal of a predetermined list of words, spell correction, translation, cleaning, imputation, tokenization, embeddings, and the like. In some embodiments, the input features may be indicative of natural language texts. The classical processor 104 may preprocess the natural language texts. In some embodiments, to pre-process the input features, the classical processors 104 may be configured to convert the plurality of input features into single sentences. In such embodiments, preprocessing may include tokenizing the natural language texts. In some embodiments, tokenization may include converting the natural language texts into vector representations. In some embodiments, tokenization may be performed using techniques including, but not limited to, word embedding, one-hot encoding, bag-of-words representation, term-frequency inverse-document-frequency (TF-IDF) representation, and the like. In some examples, the BERT model may convert the input features into tensors. In other embodiments, the input features may be multi-modal. In such embodiments, the input features may be tokenized using multi-modal tokenization techniques. [00551 In some embodiments, the classical processors 102 may be configured to manipulate dimensions of the input features. In such embodiments, the dimensions of the input features may be manipulated such that the dimensions are equivalent to the number of qubits in the variational circuit 300. The dimensions of the input features may be either increased or decreased to match the number of qubits in the variational circuit 300. hi some embodiments, the dimensions for the input features may be manipulated by applying a linear transformation to the input features.

[0056] In some embodiments, the classical processors 104 may embed the plurality of input features into the variational circuit 300. hi some embodiments, the classical processors 104 may provide input features to the variational circuits 300 a corresponding parameterized gate from the parameterized gates. In some embodiments, to embed the input features, the variational circuit 300 may be configured to apply a Hadamard Gate corresponding to each of the qubits to initialize input states thereof. In such embodiments, the variational circuits 300 may embed the value of input feature by applying the corresponding parameterized gate to the corresponding qubit.

[0057] In some embodiments, the classical processors 104 may determine one or more output values using the variational circuit 300, based on the input features, the weights, and the quantum gates. In some embodiments, to determine the output values, the classical processors 104 may measure an expectation value of each of the qubits of the variational circuit 300. In some embodiments, to determine the output values, the variational circuit 300 may be configured to, apply a Pauli operator on the qubits, and determine the expectation value of each of the qubits. In such embodiments, the expectation values are indicative of the output values. In some examples, the Pauli operator may be indicative of a Pauli Z gate.

[0058] In some embodiments, the classical processors 104 determine a loss value by comparing the corresponding set of labels of the plurality of input features and the one or more output values. In some embodiments, the loss value may indicate the performance of the neural network. The loss value may be determined by a predetermined loss function. The loss function may be selected based on the task for which the neural network is being trained. In an example, the loss function selected for a text classification task may include, but not be limited to, binary crossentropy loss, multi-class cross-entropy loss, hinge loss, and the like.

[0059] In some embodiments, the classical processors 104 may update the one or more weights based on the loss value. In such embodiments, the classical processors 104 may update the parameters of the parameterized gates based on the loss value. In some embodiments, each weight may be updated based on a gradient of the loss value with respect to quantum transformation performed by the corresponding parameterized gate. In some embodiments, the gradient may be determined using parameter shift rule. In some embodiments, the gradient may be determined using eigenvalue dependent shifts of parameters for that corresponding qubit.

[0060] In some embodiments, the classical processor 104 may be configured to iteratively determine the output values from the input features, determine the loss value, and update the weights until a stopping criterion is met. In some embodiments, the classical processor 104 may stop iterating when the loss value at the last iteration is below a predetermined threshold. In other embodiments, the classical processor 104 may stop iterating when the percentage change in loss value over a predetermined number of previous iterations is below the predetermined threshold.

[0061] In some embodiments, the classical processors 104, the quantum processor 102, the data sources 108, and the database 106 may communicate through a communication means. The communication means may be wired communication means, or wireless communication means, or a combination thereof. In some embodiments, the wired communication means may include, but not limited to, wires, cables, data buses, optical fibre cables, and the like. In some embodiments, the wireless communication means may include, but not be limited to, telecommunication networks, Near Field Communication (NFC), Bluetooth, Internet, Local Area Networks (LAN), Wide Area Networks (WAN), Light Fidelity (Li-FI) networks, a carrier network, and the like. In some embodiments, the form factor of the data transmitted through the communication means may be any one or combination of including, but not limited to, analogue signals, electrical signals, digital signals, radio signals, infrared signals, data packets, and the like.

[0062] In some embodiments, the present disclosure relates to a non-transitory computer-readable medium which may include processor-executable instructions that implement the system and the method of the present disclosure.

[0063] FIG. 2 illustrates an example block diagram 200 of one or more classical processors of the system, according to embodiments of the present disclosure.

[0064] Referring to FIG. 2, a block diagram 200 of the classical processor 104 may include one or more processor(s) 202. The one or more processor(s) 202 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, logic circuitries, and / or any devices that manipulate data based on operational instructions. Among other capabilities, the one or more processor(s) 202 may be configured to fetch and execute computer-readable instructions stored in a memory 204 of the classical processor 104. The memory 204 may store one or more computer-readable instructions or routines, which may be fetched and executed to create or share the data units over a network service. The memory 204 may include any non-transitory storage device including, for example, volatile memory such as Random Access Memory (RAM), or non-volatile memory such as an Erasable Programmable Read-Only Memory (EPROM), flash memory, and the like.

[0065] In an embodiment, the classical processor 104 may also include an interface(s) 206. The interface(s) 206 may include a variety of interfaces, for example, interfaces for data input and output devices, referred to as I / O devices, storage devices, and the like. The interface(s) 206 may facilitate communication between the classical processor 104, the quantum processor 102, the data sources 108, and the database 106. The interface(s) 206 may also provide a communication pathway for one or more components of the classical processor 104. Examples of such components include, but are not limited to, processing module(s) 208 and the database 106.

[0066] In an embodiment, the processing module(s) 208 may be implemented as a combination of hardware and programming (for example, programmable instructions) to implement one or more functionalities of the processing module(s) 208. In examples described herein, such combinations of hardware and programming may be implemented in several different ways. For example, the programming for the processing module(s) 208 may be processorexecutable instructions stored on a non-transitory machine-readable storage medium and the hardware for the processing module(s) 208 may include a processing resource (for example, one or more processors), to execute such instructions.

[0067] In the present examples, the machine-readable storage medium may store instructions that, when executed by the processing resource, implement the processing module(s) 208. In such examples, the classical processor 104 may include the machine-readable storage medium storing the instructions and the processing resource to execute the instructions, or the machine-readable storage medium may be separate but accessible to the classical processor 104 and the processing resource. In other examples, the processing module(s) 208 may be implemented by electronic circuitry. The database 106 may include data that is either stored or generated as a result of functionalities implemented by any of the components of the processing module(s) 208. In some embodiments, the database 106 may store the weights of the neural network.

[0068] In some embodiments, the processing module(s) 208 may include a retrieving module 210, a preprocessing module 212, a loss determination module 214, an inference module 216, and other module(s) 218. The other module(s) 218 may implement functionalities that supplement applications / functions performed by the classical processor 104.

[0069] In some embodiments, the retrieving module 210 may be configured to retrieve the first and second set of data from the data sources 108, and generate the training dataset therefrom. In some embodiments, when the classical processor 104 is used for real-time inference, the retrieving module 210 may receive the first and second set of data from the data sources 108, and transmit the first and second set of data to the preprocessing module 212.

[0070] In some embodiments, the preprocessing module 212 may be configured to preprocess the input features such that the input features may be embedded into the variational circuit 300.

[0071] In some embodiments, the inference module 216 may be configured to execute the variational circuit 300, and determine the output values therefrom, hi some embodiments, the inference module 216 may be configured to simulate the execution of the variational circuit 300 within the classical processor 104. In other embodiments, the inference module 216 may be configured to transmit a signal to the quantum processor 102 external to the classical processor 104 to execute the variational circuit 300. The inference module 216 may receive the output values determined by the quantum processor 102 on execution of the variational circuit 300.

[0072] In some embodiments, the loss determination module 214 may be configured to determine the loss value. In some embodiments, the loss determination module 214 may determine the loss value based on the predetermined loss function. The loss value may be used for updating the weights of the neural network.

[0073] Referring to FIG. 3, an example representation of the variational circuit 300 of the system 100 is illustrated. In the example shown in FIG. 3, the variational circuit 300 may have the 4 qubits, represented by wires 301-1, 301-2, 301-3, and 301-4 (collectively referred to as wires 301). The wires 301 shown in FIG. 3 have a depth of 3. In such examples, the variational circuit 300 may include the plurality of quantum gates that transform the states of the qubits. However, it may be appreciated by those skilled in the art that the number of quantum gates and the arrangement thereof may be suitably adapted based on the number of weights of the neural network, and the distribution of weights across one or more layers of the neural network.

[0074] As shown, the quantum gates of the variational circuit 300 may include one or more Hadamard Gates 302 that initialize the input state of the corresponding qubits.

[0075] Further, the quantum gates may include the one or more parameterized gates. The parameterized gates may be any one of single qubit rotation gates including, but not limited to, RX gates, RZ gates, or RY gates such as RY gates 304-1, 304-2, and 304-3 (collectively referred to RY gates 304) shown in FIG. 3. The parameterized gates may receive a corresponding parameter, and transform the qubit based on said parameter. In such examples, the RY gates 304 may cause the qubits to change direction of rotation of the corresponding qubits along the Y-axis by the parameter value provided to the RY gates 304. In some embodiments, the first parameterized gate, such as RY gate 304-1, may be configured to embed the input features to the initial state of the qubits. Further, the subsequent parameterized gates, such as RY gates 304-2, 304-3, may be configured to apply quantum transformations to the qubits embedded with the input features. The parameters of the subsequent parameterized gates be indicative of weights of the neural network.

[0076] Additionally, the variational circuits 300 may include the two-qubit gates, such as C-NOT gates 306-1, 306-2, and 306-3 (collectively referred to as C-NOT gates 306), that entangle the qubits. In some embodiments, the C-NOT gates 306 may be arranged to alternatively entangle each qubit with either the succeeding qubit or the preceding qubit at each successive depth / time step.

[0077] Applying the parameterized gates to the qubits may correspond to multiplying the input features with the weights of the neural network to obtain the output values. The use of the variational circuit 300 may allow the output values of the neural network to be determined with reduced time and computational cost.

[0078] Referring to FIG. 4, an example implementation 400 of the system 100 is illustrated. As shown, the system 100 may be configured to classify and cluster one or more natural language tokens indicative of complaint data 402 received from the first data source 108-1. The system 100 may be configured to process the natural language tokens to provide any of or combination of tasks including, but not limited to, classification, clustering, summarization, translation, next-token completion, chatting, and the like. The classical neural network 404 may be pretrained to provide any of the aforementioned tasks. The complaint data 402 and weights of the classical neural network 404 may be embedded into the variational circuit 300 implemented in the quantum processor 102. The complaint data 402 may be converted into quantum embeddings 406 using the quantum gates of the variational circuit 300, and transformed by the parameterized gates thereof. Further, the variational circuit 300 may be measured to determine the variational measurements 408 indicative of the output values. The variational measurements 408 may provide the same output values as the classical processor 104 may have provided on multiplying the input features with the weights in the one or more layers of the classical neural network 404. The use of the variational circuit 300 reduces computational time and cost by allowing for embarrassingly parallel computation.

[0079] The variational measurements 408 may be decoded to obtain keyword labels 410 determined by the system 100. In some embodiments, the keyword labels 410 may be indicative of classification of the complaint data 4O2.In some examples, when the complaint data 402 is indicative of “oil leak below gear box”, the keyword label 410 outputted by the system 100 may be indicative of “gear oil leak”. In other embodiments, the keyword labels 410 may be indicative of cluster of natural language tokens used in the complaint data 402. The operators 112 may analyse the cluster of natural language tokens to identify cause of the complaint and intimate the appropriate process unit for resolution thereof. In some examples, the keyword labels 410 may be indicative of “door, front, window, noise, left” for a given complaint data 402. In such examples, the operator 112 may forward the complaint to process units associated with manufacturing and assembling of front door windows of vehicles.

[0080] In yet other embodiments, the key work labels 410 may be indicative of natural language responses containing instructions to resolve the problem identified in the complaint. In some embodiments, the system 100 may be configured to identify and retrieve a predefined resolution plan for resolving the problem identified in the complaint. In some examples, the database 106 may include the predefined list of resolution plans for commonly raised complaints. In such examples, the classical neural network 404 may be trained to associate one or more of the predefined resolution plans for the complaint data 402. In other embodiments, the system 100 may provide a chat interface for the operator 112. The chat interface may be an autonomous agent built using a large language model (implemented using the classical neural network 404) having access to the first and second sets of data. The chat interface may process natural language questions provided by the operator 112, and process and generate natural language responses for the operator 112. In the foregoing example, the operator 112 may request the chat interface to retrieve development data associated with the gear box of the user’s vehicle to understand the design of thereof. Further, the operator 112 may request the manufacturing and testing data of the user’s vehicle to analyse the complaint data 402 in the context of its development, manufacturing, and testing data. In some embodiments, the chat interface may be trained to generate resolution plans to resolve the complaint of the user 110. In such embodiments, the keyword labels 410 may be indicative of natural language text providing instructions or suggestions for resolving the complaint. In some embodiments, the system 100 may also be trained to identify the appropriate process unit and transmit the natural language responses generated by the variational circuit 300 thereto for resolution of the complaint.

[0081] In some embodiments, the system 100 may be used for analysing and processing each complaint data 402 individually for real-time classification and resolution of the complaints. In other embodiments, the system 100 may be used to analyse complaint data over a predetermined interval of time. In such embodiments, the complaint data 402 may be analysed for systemic problems in the development, manufacturing or testing process.

[0082] In some embodiments, the output values of the variational circuit 300 may be used for analysing the input features. The system 100 may be used to associate one or more natural language tokens from the first set of data to one or more natural language tokens from the second set of data using multi-class classification, in an example. The classifications may be graphical visualized in the dashboard for ease of analysis for the operator 112.

[0083] In a first example, the system 100 may classify each complaint raised by the user 110 based on the parts of a vehicle referred to by the user 110 in the complaint. In a second example, the system 100 may classify each complaint based on the location of manufacturing the vehicle. In a third example, the system 100 may classify the complaint based on model or series of associated with the vehicle, and in a fourth example, the system 100 may classify the complaint based on a predefined set of damage codes. In a fifth example, the system 100 may classify the complaint based on the manufacturing plant associated with the vehicle. In such examples, the system 100 may classify the natural language tokens in the first set of data, i.e. the complaint data, with the natural language tokens in the second set of data, i.e. development, manufacturing, or testing data, thereby allowing the operator 112 forward the complaint to the appropriate process unit for resolution. In some examples, the system 100 may further provide chatting functionality to the operator 112.

[0084] Referring to FIG. 5, an example flowchart of a method 500 for efficient addressal of complaints with neural networks trained using quantum transfer learning is illustrated.

[0085] At step 502, the method 500 includes retrieving, by one or more classical processors, a training dataset having a plurality of entries, each entry comprising a plurality of input features having a corresponding set of labels, wherein the plurality of entries in the training dataset are received from a plurality of data sources.

[0086] At step 504, the method 500 includes pre-processing, by the one or more classical processors, the plurality of input features.

[0087] At step 506, the method 500 includes embedding, by the one or more classical processors, the plurality of input features into a variational circuit having a plurality of quantum gates corresponding to one or more weights of a neural network.

[0088] At step 508, the method 500 includes determining, by the one or more classical processors, one or more output values using the variational circuit, based on the plurality of input features, the one or more weights, and the plurality of quantum gates.

[0089] At step 510, the method 500 includes determining, by the one or more classical processors, a loss value by comparing the corresponding set of labels of the plurality of input features and the one or more output values.

[0090] At step 512, the method 500 includes updating, by the one or more classical processors, the one or more weights based on the loss value.

[0091] In some embodiments, the plurality of input features and the corresponding labels are indicative of natural language tokens. In some embodiments, for preprocessing the plurality of input features, the method 500 may include converting, by the one or more classical processors, the input features into single sentences.

[0092] In some embodiments, the plurality of input features and the corresponding set of labels may include one or more natural language tokens from different data sources.

[0093] hi some embodiments, the neural network may be a pre-trained neural network, and wherein the method 500 may include updating, by the one or more classical processors, a subset of the one or more weights in one or more layers of the pre-trained neural network.

[0094] Referring to FIG. 6, the block diagram represents a computer system 500 that includes an external storage device 610, a bus 620, a main memory 630, a read only memory 640, a mass storage device 650, a communication port 660, and a processor 670. A person skilled in the art will appreciate that the computer system 600 may include more than one processor 670 and communication ports 660. The processor 670 may include various modules associated with embodiments of the present disclosure. The communication port 660 can be any of a Recommended Standard 232 port for use with a modem-based dialup connection, a 10 / 100 Ethernet port, a Gigabit or 10 Gigabit port using copper or fiber, a serial port, a parallel port, or other existing or future ports. The communication port 660 may be chosen depending on a network, such as a Local Area Network (LAN), a Wide Area Network (WAN), or any network to which computer system 600 connects. [00951 In an embodiment, the memory 630 can be a RAM, or any other dynamic storage device commonly known in the art. The Read-Only Memory (ROM) 640 may be any static storage device(s) e.g., but not limited to, a Programmable Read-Only Memory (PROM) chip for storing static information. The mass storage 660 may be any current or future mass storage solution, which may be used to store information and / or instructions. Exemplary mass storage solutions may include, but are not limited to, Parallel Advanced Technology Attachment (PATA) or Serial Advanced Technology Attachment (SATA) hard disk drives or solid-state drives (internal or external, e.g., having Universal Serial Bus (USB) and / or Firewire interfaces), one or more optical discs, Redundant Array of Independent Disks (RAID) storage, e.g., an array of disks (e.g., SATA arrays). [00961 In an embodiment, the bus 620 communicatively couples the processor(s) 670 with the other memory, storage, and communication blocks. The bus 620 may be, e.g., a Peripheral Component Interconnect (PCI) / PCI Extended (PCLX) bus, Small Computer System Interface (SCSI), USB, or the like, for connecting expansion cards, drives, and other subsystems as well as other buses, such a front side bus (FSB), which connects the processor 670 to the computer system 600.

[0097] In another embodiment, operator and administrative interfaces, e.g., a display, keyboard, and a cursor control device, may also be coupled to the bus 620 to support direct operator interaction with computer system 600. Other operator and administrative interfaces may be provided through network connections connected through communication port 660. In some embodiments, the external storage device 610 can be any kind of external hard-drives, floppy drives, Compact Disc - Read Only Memory (CD-ROM), Compact Disc - Re-Writable (CD-RW), Digital Video Disk - Read Only Memory (DVD-ROM). Components described above are meant only to exemplify various possibilities. In no way should the 5 aforementioned exemplary computer system 600 limit the scope of the present disclosure.

[0098] While the foregoing describes various embodiments of the present disclosure, other and further embodiments of the present disclosure may be devised without departing from the basic scope thereof. The scope of the present disclosure 10 is determined by the claims that follow. The present disclosure is not limited to the described embodiments, versions or examples, which are included to enable a person having ordinary skill in the art to make and use the present disclosure when combined with information and knowledge available to the person having ordinary skill in the art.

Claims

1. A system for efficient addressal of complaints with neural networks trained using quantum transfer learning, comprising:one or more classical processors;a memory coupled to the one or more classical processors, the memory comprising one or more processor-executable instructions which, when executed, cause the one or more classical processors to:retrieve a training dataset having a plurality of entries, each entry comprising a plurality of input features having a corresponding set of labels, wherein the plurality of entries in the training dataset are received from a plurality of data sources;pre-process the plurality of input features;embed the plurality of input features into a variational circuit having a plurality of quantum gates corresponding to one or more weights of a neural network;determine one or more output values using the variational circuit, based on the plurality of input features, the one or more weights, and the plurality of quantum gates;determine a loss value by comparing the corresponding set of labels of the plurality of input features and the one or more output values; andupdate the one or more weights based on the loss value.

2. The system of claim 1, wherein the variational circuit comprises:a plurality of qubits corresponding to the plurality of input features, the plurality of input features being embedded into a corresponding qubit from the plurality of qubits; andthe plurality of quantum gates, comprising:a plurality of two-qubit gates that each apply a first quantum transformation to a pair of qubits from the plurality of qubits; andone or more parameterized gates that each apply a second quantum transformation to the plurality of qubits based on a corresponding parameter, wherein the corresponding parameter corresponds to the one or more weights of the neural network.

3. The system of claim 2, wherein the plurality of two-qubit gates of the variational circuit is configured such that each pair of adjacent qubits from the plurality of qubits are entangled.

4. The system of claim 2, wherein the one or more classical processors are configured to manipulate dimensions of the plurality of input features to correspond to a number of the plurality of qubits in the variational circuit by applying a linear transformation to the plurality of input features.

5. The system of claim 2, wherein to embed the plurality of input features, the variational circuit is configured to:apply a Hadamard Gate corresponding to each of the plurality of qubits to initialize input states of each of the plurality of qubits; andapply a corresponding parameterized gate from the one or more parameterized gates having the plurality of input features as the parameter therefor to the corresponding qubit.

6. The system of claim 2, wherein to determine the one or more output values, the one or more classical processors are configured to measure an expectation value of each of the plurality of qubits of the variational circuit.

7. The system of claim 2, wherein to determine the one or more output values, the variational circuit is configured to:apply a Pauli operator on the plurality of qubits; anddetermine an expectation value of each of the plurality of qubits, wherein the expectation values are indicative of the one or more output values.

8. The system of claim 1, wherein the plurality of input features and the corresponding set of labels are indicative of natural language tokens, andwherein to pre-process the plurality of input features, the one or more classical processors are configured to convert the plurality of input features into single sentences.

9. The system of claim 1, wherein the plurality of input features and the corresponding set of labels comprises one or more natural language tokens from different data sources.

10. The system of claim 1, wherein the neural network is a pre-trained neural network, and wherein the one or more classical processors are configured to update a subset of the one or more weights in one or more layers of the pretrained neural network.

11. A method for efficient addressal of complaints with neural networks trained using quantum transfer learning, comprising:retrieving, by one or more classical processors, a training dataset having a plurality of entries, each entry comprising a plurality of input features having a corresponding set of labels, wherein the plurality of entries in the training dataset are received from a plurality of data sources;pre-processing, by the one or more classical processors, the plurality of input features;embedding, by the one or more classical processors, the plurality of input features into a variational circuit having a plurality of quantum gates corresponding to one or more weights of a neural network;determining, by the one or more classical processors, one or more output values using the variational circuit, based on the plurality of input features, the one or more weights, and the plurality of quantum gates;determining, by the one or more classical processors, a loss value by comparing the corresponding set of labels of the plurality of input features and the one or more output values; andupdating, by the one or more classical processors, the one or more weights based on the loss value.

12. The method of claim 11,wherein the plurality of input features and the corresponding labels are indicative of natural language tokens, andwherein to preprocess the plurality of input features, the method comprises converting, by the one or more classical processors, the plurality of input features into single sentences.

13. The method of claim 11, wherein the plurality of input features and the corresponding set of labels comprises one or more natural language tokens from different data sources.

14. The method of claim 11, wherein the neural network is a pre-trained neural network, and wherein the method comprises updating, by the one or more classical processors, a subset of the one or more weights in one or more layers of the pre-trained neural network.

15. A non-transitory computer-readable medium comprising processorexecutable instructions that cause a processor to:retrieve a training dataset having a plurality of entries, each entry comprising a plurality of input features having a corresponding set of labels, wherein the plurality of entries in the training dataset are received from a plurality of data sources;pre-process the plurality of input features;embed the plurality of input features into a variational circuit having a plurality of quantum gates corresponding to one or more weights of a neural network;5 determine one or more output values using the variational circuit,based on the plurality of input features, the one or more weights, and the plurality of quantum gates;determine a loss value by comparing the corresponding set of labels of the plurality of input features and the one or more output values; and10 update the one or more weights based on the loss value.

Images