Abstract: Implementations relate to asymmetric quantization of large language models (LLMs). Processor(s) of a system can: obtain a trained LLM, wherein the trained LLM includes a plurality of layers, each layer comprising a respective plurality of weights; for each layer of the plurality of layers: calculate an optimal clipping range for the respective plurality of weights, and clip one or more weights of the respective plurality of weights that lie outside of the optimal clipping range to produce a clipped layer; quantize the LLM to generate a quantized LLM, wherein the instructions to quantize include instructions to map weights of the plurality of clipped layers of the LLM from continuous values to discrete values; and provide the quantized LLM for downstream processing.

Abstract: Machine learning models are used to generate a plan that responds to a user request. The plan includes one or more skills selected from a list of available skills. The prompt may be written in natural language, enabling the user to express their intent without having to know which skills are available or their intricacies. In some configurations, a skill is included in the plan if an embedding representation of an example prompt associated with the skill is within a defined distance of an embedding representation of the user request. Additionally, or alternatively, the embedding distance computations are used to narrow the list of available skills, which is then used to construct a meta-prompt that selects a skill. Skills listed in the meta-prompt may include data types of parameters and return values. This allows the model that processes the meta-prompt to order skills based on data type compatibility.

Abstract: The present disclosure relates to methods and systems for using large language models to support research activities. The methods and systems include a copilot engine that creates input prompts to provide to the large language model to use in generating responses to input messages. The copilot engine infers an intent of the input messages and sends the intent with the input message in the input prompt to the large language model. The large language model generates different types of responses for different intents.

Abstract: A computer-implemented method, according to one embodiment, includes extracting predicates from a predetermined plurality of sentences, and causing an explainer component to analyze the sentences to determine attentions from the predicates of the sentences. The method further includes causing the extracted predicates to be input into a predetermined pruner model. The pruner model is trained to use the attentions to generate a pruned list of predicates from the extracted predicates. A logical neural network is caused to be trained using the pruned list of predicates. A computer program product, according to another embodiment, includes a computer readable storage medium having program instructions embodied therewith. The program instructions are readable and/or executable by a computer to cause the computer to perform the foregoing method.

Abstract: A computer-implemented method for generating a fixed-size N-dimensional vector representation for a given document is disclosed. The method comprises extracting text-portions from a plurality of documents, embedding the extracted text-portions into fixed-sized K-dimensional text-portion vectors, clustering the text-portion vectors into N clusters C_1, C_2, . . . , C_N, generating an N-dimensional document vector E(D) for a document D by (i) associating its nth coordinate value E(D)_n to the nth cluster C_n, and (ii) if not previously done for the document D extracting text-portions from the document D and embedding the extracted text-portions into K-dimensional text-portion vectors, and (iii) setting the values E(D)_n based on similarity matching score values between the text-portion vectors of the document D and text-portions vectors of the N clusters.

Abstract: A natural language processing (NLP) framework allows for receiving, recognizing, and performing actions in relation to multi-party communications (MPC) through interactive conversations with a user. In an embodiment, the NLP framework allows a user to conduct an interactive conversation with a conversation engine and provide it, through natural language inputs, with commands or requests related to MPCs.

Abstract: A method, a structure, and a computer system for diverse natural language generation. The exemplary embodiments may include training a data-to-text neural network (D2T NN) and training a text-to-text neural network (T2T NN), wherein the D2T NN and the T2T NN have identical transformer architectures, and wherein the training of the D2T NN and the T2T NN are indexed by time. The exemplary embodiments may further include interleaving weights between the D2T NN and the T2T NN, as well as generating a sentence based on the interleaved D2T NN and T2T NN.

Abstract: A method, computer program product, and computer system are provided for enhancing chatbot responses. Data corresponding to a user input to a chatbot is received. A prompt is generated for a large language model based on the received data and a conversation flowchart associated with the chatbot. The generated prompt is input to the large language model. A natural language response to the received data is generated based on an output from the large language model. The generated natural language response substantially corresponds to a response associated with the conversation flowchart.

Abstract: In one aspect, an example method includes (i) obtaining media content; (ii) extracting from the obtained media content, audio content representing speech; (iii) using the extracted audio content representing speech as a basis to generate corresponding speech text; (iv) replacing one or more words of the generated speech text with one or more corresponding replacement words, thereby generating modified speech text; (v) using the modified speech text as a basis to generate corresponding replacement audio content representing the modified speech; (vi) in the obtained media content, replacing the audio content representing speech with the generated replacement audio content representing speech, thereby generating modified media content; and (vii) outputting for presentation the generated modified media content.

Abstract: Aspects of the disclosure are directed to methods, systems, and computer readable media for adaptation with self-evaluation to improve selective prediction in large language models (LLMs), generally referred to as ASPIRE. ASPIRE includes training LLMs on a portion of training data from a question answering task to learn self-evaluation, e.g., learn to distinguish whether a generated answer is correct or not. ASPIRE further includes a selection score that combines a likelihood of that generated answer is correct with a self-evaluation score for selective prediction. ASPIRE demonstrates improved selective prediction performance with less computational cost.

Abstract: Disclosed is a method for generating a storyboard based on a script text, which is performed by a computing device. The method may include obtaining the script text. The method may include performing clustering for text parts included in script text based on location information. The method may include determining a prompt information for each cluster generated by clustering based on the location information. The method may include generating the storyboard for each cluster based on the determined prompt information.

Abstract: The presently disclosed embodiments may include a computer readable medium including instructions that when executed by one or more processing devices cause the one or more processing devices to perform a method. The method may include receiving a user input, wherein the user input includes a natural language question, using a trained language model to decompose the natural language question into two or more information requests, routing each of the two or more information requests to at least one information resource, receiving two or more information responses from the at least one information resource, wherein the two or more information responses correspond to the two or more information requests, generating a natural language response to the user input based on the two or more information responses and providing the natural language response to the user.

Abstract: A method and system for automated voice casting compares candidate voices samples from candidate speakers in a target language with a primary voice sample from a primary speaker in a primary language. Utterances in the audio samples of the candidates speakers and the primary speaker are identified and typed and voice samples generated that meet applicable utterance type criteria. A neural network is used to generate an embedding for the voice samples. A voice sample can include groups of different utterance types and embeddings generated for each utterance group in the voice sample and then combined in a weighted form wherein the resulting embedding emphasizes selected utterance types. Similarities between embeddings for the candidate voice samples relative to the primary voice sample are evaluated and used to select a candidate speaker that is a vocal match.

Abstract: A method of recommending, by a recommendation apparatus, an asset and content to a user, and including receiving learning data from a platform for using the content and an authorizing tool for producing the content; constructing a learning model based on the training data; generating 1/a first model for analyzing characteristics of the user, 2) a second model for recommending the content, and 3) a third model for recommending the element using the learning model; transferring recommended content information to the platform using the first model and the second model; and transferring recommended element information to the authorizing tool.

Abstract: Disclosed are apparatuses, systems, and techniques that may use machine learning for reversible translations of speech utterances. The techniques include training and using duplex neural networks (NNs) having a first subnetwork and a second subnetwork that are mirror images of each other. Training data for training the duplex NNs may include a target output that includes a first speech utterance in a first language, a first training input that includes the target output distorted by a noise, and a second training input that includes a second speech utterance in a second language. The duplex NNs may be trained to identify, using the first training input and the second training input, at least one of the target output or the first noise.

Abstract: A display apparatus includes a receiving unit to receive input of handwritten data on a screen, and a display control unit. The display control unit simultaneously displays, on a display, a plurality of character strings in different languages that are synonymous with a character string converted from the handwritten data.

Abstract: Systems and methods are disclosed for gradual and/or gradient-based pruning for language-specific multilingual neural machine translation. For example, a method may include determining a pruning mask for weights of a multilingual machine translation model based on a first pruning threshold, wherein the pruning mask includes at least one entry set to zero and at least one entry set to one; training the multilingual machine translation model, while applying the pruning mask to the multilingual machine translation model, for translation between a language pair based on training examples from a bilingual translation corpus; updating the pruning mask based on a second pruning threshold; and training the multilingual machine translation model, while applying the updated pruning mask to the multilingual machine translation model, for translation between the language pair based on training examples from the bilingual translation corpus. For example, the weights may be pruned based on gradient-based scores.

Abstract: Analog system and method for implementing an iterative neural network based model, the system comprising analog vector-by-matrix multiplication circuitry encoding a matrix of weights of an iterative neural network-based model, and analog nonlinearity circuitry encoding a non-linear function arranged in a feedback loop configured to return the output signals from the nonlinearity circuitry as inputs to the vector-by-matrix multiplication circuitry, wherein the system is configured to output a solution vector of values of the iterative neural network based model on convergence of the system.

Abstract: Broadcasting power limiting management responses in a processor-based system in an integrated circuit (IC) chip is disclosed herein. In one aspect, an IC chip comprises a processor-based system that includes a power estimation and limiting (PEL) circuit, a Limit Management Throughput Throttle (LMTT) source circuit, a plurality of activity management (AM) circuits, and an LMTT bus communicatively coupling the LMTT source circuit with each AM circuit of the plurality of AM circuits. The LMTT source circuit receives a power limiting management response from a PEL circuit via a communications network of the processor-based system, and generates an LMTT command based on the power limiting management response. The LMTT source circuit broadcasts the LMTT command to each AM circuit of the plurality of AM circuits via the LMTT bus.

Abstract: A transaction card may monitor an amount of electric current induced in a first near-field communication (NFC) component of a transaction card. Electric current may be induced in the first NFC component when the first NFC component is within an electromagnetic field generated by a second NFC component of a transaction terminal. The transaction card may dynamically activate one or more output components associated with the transaction card based on the amount of electric current induced in the first NFC component. The one or more output components may indicate whether the transaction card can communicate with the transaction terminal. The transaction card may perform an action related to completing a transaction after determining that the amount of electric current induced in the NFC component satisfies a first threshold. The first threshold may indicate that the transaction card can communicate with the transaction terminal.

Abstract: Systems and methods include a computer-implemented method for managing operations. A detailed sequence of drill site operations for each rig in a fleet is used to determine a time schedule for the detailed sequence of drill site operations, identifying routine operational steps and dependent prerequisite activities required to proceed to each next operational step. A critical path for drill site operations is determined based on the time schedules. A shortened critical path is created using machine learning on the drill site operations. Suggestions of possible next steps for an in-progress drilling operation using the shortened critical path are provided in real time. A sequence of operations is optimized using a user-selected suggestion. Materials and services requests are automated. Location information of personnel authorized to perform a task at a rig is collected. A determination is made that the personnel can perform the task. Notifications to perform the task are provided.

Abstract: A method includes: storing, within first and second computing devices, a first immutable token (IT) that incorporates at least a first data concerning a first characteristic of a registered virtual object; receiving, at the first and second computing devices, and from a reading device, a request to authenticate another virtual object as being the registered virtual object based on a second IT associated with the other virtual object; separately comparing, at the first and second computing devices, the second IT to the first IT; separately determining, at the first and second computing devices, whether to authenticate the other virtual object based on a degree to which the second IT matches the first IT; and separately transmitting, from the first and second computing devices, and to the reading device, separate indications of whether the other virtual object is able to be authenticated.

Abstract: A security platform includes a security tag having a wireless communication circuitry configured to receive a user identification (ID) from an ID tag. After receiving the user ID, the security tag tracks the distance the security tag moves away from a predetermined point. The security tag then compares the determined distance to a predetermined distance threshold to determine whether the security tag is within a range. In response to the determined distance being greater than the predetermined distance threshold, the security tag determines that the security tag is outside of the range and activates an alarm. The alarm of the security tag continues to be active until a second user ID having appropriate access rights is received.

Abstract: An induction system is disclosed for filtering the induction of objects to a plurality of object processing systems. The induction system includes an evaluation means for evaluating at least one characteristic of an object, and routing means for routing the object in one of a plurality of directions responsive to the evaluated characteristic, at least one of the plurality of directions leading to a selected one of the plurality of object processing systems.

Abstract: Described herein are barcode readers having a housing with upper and lower portions, and a weigh platter. Additionally, the barcode readers include a first imaging assembly having a first imaging sensor, the first imaging assembly having a first field of view (FOV) directed through at least one of the substantially horizontal window or the substantially upright window of the housing, and a second imaging assembly having a second imaging sensor, the second imaging assembly having a second FOV and being positioned near the distal edge of the weigh platter that is opposite the upper portion of the barcode reader housing.

Abstract: In an embodiment a method for reading a marking includes providing the marking on a substrate, the marking being formed with at least one marking material, providing at least one opaque covering layer covering the marking, briefly irradiating the at least one covering layer with electromagnetic radiation to which the at least one covering layer is impermeable; and contactless reading the marking by thermal imaging.

Abstract: A printing system and a method for printing a test chart by the printing system having a media input holder, a media output holder, a print engine and a print controller.

Abstract: According to one embodiment, an image forming apparatus has a printing unit to print images on a sheet of paper from a paper storage unit and a conveying unit configured to convey sheets from the storage unit to the printing unit. A tag reader/writer communicates with a wireless tag on a sheet when the sheet is at a communication position on the conveyance path. A control unit determines whether a writing of data to a tag on a first sheet has succeeded, then controls the printing unit to print on the first sheet if the writing of data succeeded but not if the writing of data to the wireless tag failed. When writing of data to the wireless tag has failed, the control unit executes a tag data writing retry process on another sheet from the paper storage unit.

Abstract: According to one embodiment, an image forming apparatus includes a tag communication unit configured to communicate with wireless tags, an image forming unit configured to form image on media carrying wireless tags, and a processor. The processor is configured to permit the image forming unit to form an image on a medium when the tag communication unit successfully writes data to a wireless tag carried by the medium, and indicate the image formed on the medium by the image forming unit is a use charge target after the image forming unit is permitted to form the image on the medium.

Abstract: A method includes generating a pattern, and projecting the pattern. A plurality of elements are arranged in the pattern. The method further includes selecting a color for each of the plurality of elements from a color candidate group including a plurality of colors with equal differences between the plurality of colors. The color of each of the plurality of elements is different from a neighboring color of a neighboring element adjacent to that element. A plurality of color sequences corresponding to a plurality of sub-arrays are different from each other. The plurality of sub-arrays includes at least one element from among the plurality of elements. Distances between pairs of color sequences of the plurality of color sequences within the pattern are greater than or equal to a predetermined threshold.

Abstract: Provided is a smart card including: first and second antennas performing wireless communication with a card reader; a smart card chip electrically connected to the first antenna and performing the wireless communication with the card reader through the first antenna; a power generation unit converting a frequency signal received through the second antenna to generate direct current power, and an application circuit unit receiving the direct current power from the power generation unit and having at least one passive element operating under the control of the smart card chip, wherein one end of the at least one passive element of the application circuit unit is electrically connected to the power generation unit and the other end thereof is electrically connected to any one of a plurality of GPIO terminals of the smart card chip.

Abstract: A carrier structure including a carrier, an antenna structure on the carrier and a reception region designed for removable reception of an RF chip. The reception region is arranged relative to the antenna structure in such a way that inductive coupling to the antenna structure is enabled when the RF chip is arranged in the reception region. An inductive blocking element on the carrier blocks communication with the RF chip using the antenna structure in a first position of the inductive blocking element. In a second position of the inductive blocking element, the inductive blocking element enables communication with the RF chip using the antenna structure.

Abstract: A memory card is provided with various pad layouts to prevent a data signal pad from contacting a power contact in a host during insertion and removal of the memory card. The memory card can have a form factor and features that accommodate a relatively-large memory with relatively-high performance and accompanying thermal conditions. An efficient card lock mechanism is also provided.

Abstract: The invention relates to an RFID arrangement comprising an RFID tag and a housing at least partially enclosing the RFID tag, wherein the RFID tag has a first contact and a second contact, and wherein the first contact is electrically connected to the housing. The invention relates further to a corresponding piece of jewelry, to corresponding methods, and to a corresponding local sending arrangement.

Abstract: Disclosed are various approaches to depositing an RFID tag or chip within a rubber substrate. A radio frequency identifier (RFID) chip encapsulated with an adhesive or adhesion promoter is attached to the rubber substrate. Then, a wire antenna is attached to the rubber substrate and inductively coupled to the RFID tag. The wire antenna can be encapsulated with an adhesive or adhesion promoter. The wire antenna can also be arranged longitudinally along the rubber substrate in a plurality of loops with at least one of the plurality of loops placed around the RFID tag.

Abstract: A reading device includes an antenna element, a substrate including a communication circuit and a ground plane, a connecting member protruding from the substrate in a thickness direction of the substrate and configured to electrically connect the communication circuit and the antenna element, and a casing configured to retain the antenna element, the substrate, and the connecting member. The antenna element is arranged separately from the substrate toward one side in the thickness direction. The casing includes a first surface portion that covers the substrate when viewed from the other side in the thickness direction. In a case where a distance in the thickness direction from a feed point to the ground plane is a first distance, and a distance in the thickness direction between the substrate and the first surface portion is a second distance, the first distance is greater than the second distance.

Abstract: An active chatbot system with composite finite state machine and a method thereof are disclosed. In the active chatbot system, a rough question message having a natural language structure is generated based on a client behavior state and an on-demand conversation setting, and the rough question message is inputted to a question optimization circuit to generate a precise question message, and the precise question message is transmitted to an artificial intelligence platform to obtain a corresponding answer message, the answer message is inputted to a trained emotional AI model to generate an emotional answer message and the emotional answer message is stored in an answer list, so that the emotional answer message matching the on-demand conversation setting can be filtered out as an on-demand conversation message, the on-demand conversation message is transmitted to the client-end host for output. Therefore, the technical effect of improving conversational flexibility and realism of chatbot can be achieved.

Abstract: An active chatbot system with customized setting and updating download and a method thereof are disclosed. In the active chatbot system, a rough question message having a natural language structure is generated based on a client behavior state and an customized parameter, and the rough question message is inputted to a logic circuit to generate a precise question message, and the precise question message is transmitted to an artificial intelligence platform to obtain a corresponding answer message which is used as first training data, the customization is used as second training data, the first training data and the second training data are inputted to an AI model to perform a training, weighting values of the AI model are translated to correspond to different customized parameter, a translating result is provided for the client-end host to download for customization setting and update. Therefore, the technical effect of improving convenience and specificity in model customization can be achieved.

Abstract: Embodiments described herein provide a framework that integrates a retriever model and the LLM to feed retrieved passages to an LLM to generate an answer conditioned on the retrieved passages in response to a query. For example, in one embodiment, a single-round approach is implemented, which involves directly transmitting the retrieved passages to the LLM. For another example, a multi-round methodology is implemented, which involves initially presenting the retrieved passages to the Language Model, collecting its responses, and then adjusting our interaction with the Language Model based on this acquired feedback.

Abstract: A system captures and utilizes expert knowledge in artificial intelligence. The system includes a knowledge capture module for extracting expert knowledge from subject matter experts in a conversational format and a knowledge management module for cataloging and summarizing the extracted knowledge. The system also includes a digital subject matter expert (dSME) module for ingesting the cataloged knowledge and using it to guide users in building AI models. A chatbot interacts with a user and selects the appropriate dSME module that is relevant to the user request. The system attempts to answer the user request based on the dSME module. If the dSME module lacks knowledge to solve the problem, the system uses a set of tools, for example, internet based search engine to solve the problem.

Abstract: Artificial intelligence is an increasingly important sector of the computer industry. However, artificial intelligence is extremely computationally intensive field such that it can be expensive, time consuming, and energy consuming. Fortunately, many of the calculations required for artificial intelligence can be performed in parallel such that specialized processors can great increase computational performance. Specifically, artificial intelligence generally requires large numbers of matrix operations to implement neural networks such that specialized matrix processor circuits can improve performance. To perform all these matrix operations, the neural processing circuits must be quickly and efficiently supplied with data to process or else the matrix processor circuits end up idle or spending large amounts of time loading weight matrix data.

Abstract: Disclosed is a spiking neural network circuit, which includes an axon circuit that generates first and second input spike signals, a synapse circuit that generates a first current based on the first input spike signal and a weight and generates a second current based on the second input spike signal and the weight, a capacitor that forms a first membrane voltage based on the first current, and a neuron circuit including a comparator and that resets the first membrane voltage, and after the capacitor further forms a second membrane voltage based on the second current, and the comparator includes a first input terminal and a second input terminal, receives the first membrane voltage through the first input terminal and a reference voltage through the second input terminal, generates a first spike signal based on a first comparison operation of the first membrane voltage and the reference voltage.

Abstract: Technology providing in-memory neural network protection can include a memory to store a neural network, and a processor executing instructions to generate a neural network memory structure having a plurality of memory blocks in the memory, scatter the neural network among the plurality of memory blocks based on a randomized memory storage pattern, and reshuffle the neural network among the plurality of memory blocks based on a neural network memory access pattern. Scattering the neural network model can include dividing each layer of the neural network into a plurality of chunks, for each layer, selecting, for each chunk of the plurality of chunks, one of the plurality of memory blocks based on the randomized memory storage pattern, and storing each chunk in the respective selected memory block. The plurality of memory blocks can be organized into a groups of memory blocks and be divided between stack space and heap space.

Abstract: Anomaly detection in neural networks is provided. The method comprises extracting, from different layers of a recurrent neural network (RNN) for a specified time interval, a number of on-the-fly node activations produced by sequences of event data from a number of data sources and sensors. The method references known node activations of the RNN produced by normal sequences of event data, and for each layer of the RNN, calculates a maximum nonparametric divergence of the on-the-fly node activations from the known node activations. For each layer of the RNN, the method determines a subset of nodes that most contribute to the maximum nonparametric divergence for that layer for a given time window and identifies data sources or sensors from among the number of data sources and sensors responsible for activating the subset of nodes that most contribute to the maximum nonparametric divergence for each layer.

Abstract: The systems and methods may use one or more artificial intelligence models that predict an effect of a predicted event on a current state of the system. For example, the model may predict how a rate of change in time-series data may be altered throughout the first time period based on the predicted event. However, as noted above, correctly predicting the occurrence of outlier events (e.g., the predicted event), and in particular characteristics about the outlier events (e.g., when an outlier may occur, what may be a source of the outlier, what rate of change the outlier may cause, etc.), in data-sparse environments and based on time-series data presents a technical challenge.

Abstract: A method for analyzing an operation of a robot includes performing a training phase by obtaining a first dataset containing at least one temporal characteristic of at least one state parameter of a first robot and training an artificial neural network. The artificial neural network includes a first autoencoder having an encoder that maps the first dataset onto temporal characteristic patterns and the activation thereof, and a decoder that uses these temporal characteristic patterns to reconstruct the first dataset; and a second autoencoder having an encoder that maps the temporal characteristic patterns and the activation thereof onto pattern groups, and a decoder that uses these pattern groups to reconstruct the temporal characteristic patterns and the activation thereof.

Abstract: A generative machine-learning model may be trained using communication session data to tailor a generative model on particular classes of communications. A set of communications sessions may be received with each communication session including, for example, communications between a doctor and a patient. A set of features may be generated by a natural language model from communications extracted from the communication sessions. A subset of the set of communication sessions may be defined based on the set of features. A training dataset may be defined using the subset of the set of communication sessions and used to train a machine-learning model. The machine-learning model can be configured to generate one or more contexts associated with a feature of the set of features. Upon receiving a request associated with a particular, a response may be generated using the machine-learning model and a feature vector derived from the request.

Abstract: Disclosed herein are techniques and architectures for enhancing the efficiency of deep neural networks through the implementation of a pattern clustering system. A pattern clustering system can enforce shared clustering topologies on filters, thereby leading to a significant reduction in memory usage through the reuse of index information. Some embodiments of the present disclosure relate to techniques for determining and assigning clustering patterns, as well as for training a network to adhere to these target patterns. Some embodiments of the present disclosure relate to an efficient accelerator based on the patterned filters. The pattern clustering system can reduce both the memory footprint and the operation count, while maintaining accuracy comparable to that of baseline models. Furthermore, the accelerator for the pattern clustering system can significantly enhance energy efficiency, surpassing the performance of conventional technologies and setting a new benchmark in the field.

Abstract: The present invention relates to validating a linear model. Input data is received (102) and the linear model to be validated is provided (104). Predicted data is determined based on processing input data by the linear model (106). Residual data is determined based on a difference between the predicted data and the input data (108). A set of validation data including homoscedasticity validation data or normality validation data is generated based on the residual data (110). A binary classifier is provided and used for determining whether the set of validation data fulfills a validation condition (112), namely a homoscedasticity condition or a normality condition. The binary classifier is a trained data driven model that outputs that the validation condition is fulfilled or not fulfilled depending on the set of validation data. Finally, it is determined whether the linear model is valid based on the output of the binary classifier (114).

Abstract: The present application discloses a method, apparatus, device, medium and convolutional neural network for quantization aware training. The method is performed by electronic equipment and includes: performing sample training on a first convolutional neural network, wherein at least one convolution layer of the first convolutional neural network includes a mergeable branch structure and at least one first shortcut; reserving the at least one first shortcut in the at least one convolution layer of trained first convolutional neural network; merging the mergeable branch structure except the at least one first shortcut in the at least one convolution layer, and obtaining a second convolutional neural network having a first shortcut structure; and, performing quantization aware training based on the second convolutional neural network to improve accuracy of quantization of the convolutional neural network model.