Abstract: A decoder includes a feature extraction circuit for calculating one or more feature vectors. An acoustic model circuit is coupled to receive one or more feature vectors from and assign one or more likelihood values to the one or more feature vectors. A memory architecture that utilizes on-chip state lattices and an off-chip memory for storing states of transition of the decoder is used to reduce reading and writing to the off-chip memory. The on-chip state lattice is populated with at least one of the states of transition stored in the off-chip memory. An on-chip word is generated from a snapshot from the on-chip state lattice. The on-chip state lattice and the on-chip word lattice act as an on-chip cache to reduce reading and writing to the off-chip memory.

Abstract: A method for accelerating machine learning on a computing device is described. The method includes hosting a neural network in a first inference accelerator and a second inference accelerator. The neural network split between the first inference accelerator and the second inference accelerator. The method also includes routing intermediate inference request results directly between the first inference accelerator and the second inference accelerator. The method further includes generating a final inference request result from the intermediate inference request results.

Abstract: The disclosed computer-implemented method may include receiving input voice data synchronous with a visual state of a user interface of the third-party application, generating multiple sentence alternatives for the received input voice data, identifying a best sentence of the multiple sentence alternatives, executing a dialog script for the third-party application using the best sentence, the dialog script generating a response to the received voice data comprising output voice data and a corresponding visual response, and providing the visual response and the output voice data to the third-party application, the third-party application playing the output voice data synchronous with updating the user interface based on the visual response. Various other methods, systems, and computer-readable media are also disclosed.

Abstract: Embodiments provide an odor identification system including an operation array unit including at least two or more sensors which interact with odor causative substances included in an odor factor of a gas sample, a sensor data processing unit processing data obtained by interaction with the odor factor in the operation array unit, an odor factor information storing unit storing information of the odor factor and the interaction pattern information of the odor factor in advance, and a pattern identification unit identifying the odor factor on the basis of an interaction pattern while referring to the pattern processed by the sensor data processing unit and the information of the odor factor information storing unit, and collating the interaction pattern with the known odor information, wherein the odor of the object to be measured is contained.

Abstract: There is provided with an image identification apparatus. An extraction unit extracts a feature value of an image from image data using a Neural Network (NN). A processing unit identifies the image based on the feature value extracted by the extraction unit. The NN comprises a plurality of calculation layers connected hierarchically. The NN includes a plurality of sub-neural networks for performing processing of calculation layers after a specific calculation layer. Mutually different data from an output of the specific calculation layer are respectively inputted to the plurality of sub-neural networks.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for processing images using recurrent attention. One of the methods includes determining a location in the first image; extracting a glimpse from the first image using the location; generating a glimpse representation of the extracted glimpse; processing the glimpse representation using a recurrent neural network to update a current internal state of the recurrent neural network to generate a new internal state; processing the new internal state to select a location in a next image in the image sequence after the first image; and processing the new internal state to select an action from a predetermined set of possible actions.

Abstract: A model learning apparatus of the present invention has a question/answer-pair expansion unit and a translation-model learning unit. The question/answer-pair expansion unit generates expansion question/answer pairs by increasing the number of question/answer pairs associated with an index indicating that it sounds more like the character. The translation-model learning unit learns a translation model and a reverse translation model by using the expansion question/answer pairs. A response selecting apparatus of the present invention has a record unit, a document search unit, a score calculation unit, and a ranking unit. The record unit records question/answer pairs and the above described learned translation model. The score calculation unit obtains a translation likelihood which is a numerical value based on the probability of obtaining the answer from the input question and calculates a score of each of a search-result question/answer pair with respect to the input question.

Abstract: Systems and methods for classifying product feedback by an electronic device are described. According to certain aspects, an electronic device may receive consumer feedback entries associated with various products, where each entry may include an initial classification. The electronic device may analyze each entry using a machine learning model to determine a subsequent classification for the entry. When there is a mismatch between classifications, the electronic device may present information associated with the entry for review by a user, where the user may specify a final classification for the entry, and the electronic device may update the machine learning model for use in subsequent analyses.

Abstract: An entity learning recognition method and computer program product include learning training a model based on a combination of an original entity and an augmented entity in an augmented database, where the entity includes an image that is used for a training of the model and where the training is based on a visual element portion of the image with added noise.

Abstract: An apparatus and method are provided for controlling autonomous driving of a vehicle which may derive predicted paths of a pedestrian and a two-wheel vehicle during autonomous driving of the vehicle so as to minimize accidents. The method includes calculating first height information allocating a first gradient that descends in a proceeding direction of objects, including a vehicle and a pedestrian, from respective positions of the objects based on dynamic information of the objects, calculating second height information allocating a second gradient based on a probability that the pedestrian will occupy infrastructure, calculating final height information by fusing the first height information and the second height information, generating a predicted path of the pedestrian, determining a driving strategy of a host vehicle based on a predicted path of the host vehicle and the predicted path of the pedestrian.

Abstract: According to some embodiments, a method performed by a classification scanner comprises receiving an electronic message and determining a classification that applies to the electronic message. The classification is determined based on an express indication from a user. The method further comprises providing a machine learning trainer with the electronic message and an identification of the classification that applies to the electronic message. The machine learning trainer is adapted to determine a machine learning policy that associates attributes of the electronic message with the classification.

Abstract: A system and method for automated software testing that uses machine learning algorithms to automatically generate and implement software testing based on an automated analysis of the software. In an embodiment, a mobile software application comprising one or more screens is processed through a trained machine learning algorithm to identify screens and objects, understand the operational flow of the application, define priorities and dependencies within the application, define validation tests, and automatically generate one or more testing scenarios for the application. The testing scenarios may then be fed to an automated execution module which installs the application on one or more physical or virtual devices and performs testing on the application installed on those devices according to the testing scenario.

Abstract: A model-based engineering system (MBSE) tool includes functionality for sharing an MBSE model with multiple outside vendors using multiple versions of the MBSE model. A restricted version of the MBSE model is managed by a model maker (or other user with access rights) inside of an organization includes all of the model's engineering elements. The engineering elements are tagged with metadata that is analyzed to determine which elements are shareable and which are unshareable outside of the organization. An unrestricted version of the MBSE model is then created, either directly from the restricted version or through an intermediary version (referred to as the “stripped version”), to include only shareable engineering elements. This unrestricted version may then be shared with the outside vendors, and changes made by either in the restricted or unrestricted versions may be incorporated in the other versions, providing a truly collaborative MBSE experience.

Abstract: A method performed by an artificial neural network includes determining a conditional probability distribution representing a channel based on a data set of transmit and receive sequences. The method also includes determining a latent representation of the channel based on the conditional probability distribution. The method further includes performing a channel-based function based on the latent representation.

Abstract: A method and system that efficiently selects sensors without requiring advanced expertise or extensive experience even in a case of new machines and unknown failures. An abnormality detection system includes a storage unit for storing a latent variable model and a joint probability model, an acquisition unit for acquiring sensor data that is output by a sensor, a measurement unit for measuring the probability of the sensor data acquired by the acquisition unit based on the latent variable model and the joint probability model stored by the storage unit, a determination unit for determining whether the sensor data is normal or abnormal based on the probability of the sensor data measured by the measurement unit, and a learning unit for learning the latent variable model and the joint probability model based on the sensor data output by the sensor.

Abstract: Techniques for classifying and processing physical objects are disclosed. In an example, a computer system may receive first data indicating that a first machine learning model of a robotic system is incapable to classify a physical object according to at least one of a set of predetermined classifications. The computer system may also receive second data corresponding to one or more attributes associated with the physical object. A second machine learning model of the computer system may determine a cluster of physical objects that includes an identifier of the physical object, whereby the identifier is included in the cluster based at least in part on the first data and a common attribute with other physical objects of the cluster. The computer system may then determine data for processing subsequent physical objects that are determined to have the common attribute.

Abstract: The present invention relates to a breast image analysis method with four mammogram images which are input to a convolutional neural network as one input and a system therefor and the system includes an image receiving unit which receives four mammogram images; an image size adjusting unit which adjusts a size of a mammogram image received from the image receiving unit; a preprocessing unit which performs preprocessing on the mammogram image adjusted by the image size adjusting unit; a convolutional neural network (CNN)-based CNN learning unit which generates learning information by learning the mammogram image preprocessed by the preprocessing unit; and a CNN inference unit which receives the learning information learned from the CNN learning unit and a mammogram image to be classified from the image receiving unit to diagnose a breast abnormality.

Abstract: An image determination device includes: a training model which outputs, on the basis of an image to be examined, output data indicating a determination result about the image; a training part which trains the training model to output, by using training data including a training image and label data, output data indicating the label data associated with the training image, when the training image is input to the training model; a dividing part which divides the training data into a plurality of pieces of sub-training data; a measurement part which measures accuracy of determination when the training part trains the training model by using each of the plurality of pieces of sub-training data; and selection part which selects at least any one among the plurality of pieces of sub-training data on the basis of the accuracy of determination.

Abstract: A plurality of images are received from one or more social media platforms associated with a user. For a selected image of the plurality of images, a plurality of text descriptions are generated. The plurality of text descriptions are computer-generated captions that describe features of the selected image of the plurality of images. The plurality of text descriptions are processed through a natural language processing model. Based on processing, a plurality of interest contexts are derived from the plurality of text descriptions. A mapping of each of the plurality of interest contexts to one or more predefined categories associated with an online marketplace is generated. Based the mapping of each of the plurality of interest contexts to the one or more predefined categories, a user device associated with the user is caused to display an app page or web page associated with the one or more predefined categories.

Abstract: A respective set of features, including emotion-related features, are extracted from segments of a video for which a preview is to be generated. A subset of the segments is chosen using the features and filtering criteria including at least one emotion-based filtering criterion. Respective weighted preview-suitability scores are assigned to the segments of the subset using at least a metric of similarity between individual segments and a plot summary of the video. The scores are used to select and combine segments to form a preview for the video.

Abstract: An information processing device is configured to output work information related to work performed by a serving person, the information processing device including an image acquisition unit configured to acquire an original image including a served person and a plurality of served objects that the serving person serves, an image division unit configured to divide the original image into a served-person image, in which the served person is captured, and a plurality of served-object images, in which each served object is captured, a scene estimation unit configured to estimate a scene, which is the situation the serving person is in, by using a first trained model, a chunk estimation unit configured to estimate a chunk, which is information dividing or suggesting the work information, by using one of a plurality of second trained models, and an output unit configured to output the chunk.

Abstract: A computer-implemented method is presented for discovering new material candidates from a chemical database. The method includes extracting a feature vector from a chemical formula, learning a prediction model for predicting property values from the feature vector with a sparse kernel model employing the chemical database, selecting an existing material from a list of existing materials sorted in descending order based on the predicted property values by the prediction model learned in the learning step, selecting a basis material from a list of basis materials sorted in descending order of absolute reaction magnitudes to the selected existing material, and generating the new material candidates as variants of the selected existing material with consideration of the selected basis material.

Abstract: Methods, devices and processor-readable media for an integrated teacher-student machine learning system. One or more teacher-student modules are trained as part of the teacher neural network training. Each student sub-network uses a portion of the teacher neural network to generate an intermediate feature map, then provides the intermediate feature map to a student sub-network to generate inferences. The student sub-network may use a feature enhancement block to map the intermediate feature map to a subsequent feature map. A compression block may be used to compress intermediate feature map data for transmission in some embodiments.

Abstract: Systems, devices, and methods are disclosed for decision making based on plasticity rules of a neural network. A method may include obtaining a multilayered model. The multilayered model may include an input layer including one or more input units. The multilayered model may include one or more hidden layers including one or more hidden units. Each input unit may have a first connection with at least one hidden unit. The multilayered model may include an output layer including one or more output units. The method may also include receiving an input at a first input unit. The method may include sending a first signal from the first input unit to at least one hidden unit via a first connection comprising a first strength. The method may also include making a decision based on the model receiving the input.

Abstract: Techniques for securing displayed data on computing devices are disclosed. One example technique includes upon determining that the computing device is unlocked, capturing and analyzing an image in a field of view of the camera of the computing device to determine whether the image includes a human face. In response to determining that the image includes a human face, the technique includes determining facial attributes of the human face in the image via facial recognition and whether the human face is that of an authorized user of the computing device. In response to determining that the human face is not one of an authorized user of the computing device, the technique includes converting user data on the computing device from an original language to a new language to output on a display of the computing device, thereby securing the displayed user data even when the computing device is unlocked.

Abstract: Software-based solutions may mitigate physical damage to multi-layer networks, such as neural networks having shortcut (residual) connections. An example includes: providing a multi-layer network comprising a plurality of nodes; for each of a plurality of training cases: determining a set of dropout nodes, based at least on a damage model having a probability of a node being selected for dropout that is based at least on a target operating environment of the multi-layer network, wherein the probability of a node being selected is spatially correlated; and training the multi-layer network with the determined set of dropout nodes disabled (with a different set of dropout nodes for different training cases). In some examples the damage model involves expected physical radiation damage to a computing device hosting the multi-layer network, such as on board an aircraft or an earth-orbiting satellite. Thus, multiple degrees of expected damage may be addressed.

Abstract: A method may include obtaining a plurality of training samples with a plurality of classifications that include a first classification and a second classification, training an initial tree with an initial set of training samples selected from the plurality of training samples using an initial set of feature values extracted from the set of training samples, and, in response to determining that the initial tree incorrectly classified the initial set of training samples at an output node of the initial tree, training a subsequent tree using a subsequent set of feature values extracted from a subsequent set of training samples.

Abstract: A stylus includes an elongated housing, a tip extending from a first end of the elongated housing and a tri-axial force sensor mounted on a second end. A first wireless transmitter transmits a signal via the tip based on which the tip interacts with a digitizer sensor of a touch screen. The tri-axial force sensor senses contact force applied by a user pressing against the tri-axial sensor. A second wireless transmitter transmits output sensed by the tri-axial force sensor. The stylus further includes a controller that controls transmission of the first wireless transmitter and the second wireless transmitter.

Abstract: An electronic device for reconstructing an artificial intelligence model, and a control method thereof are provided. The control method includes inputting at least one input data to a first artificial intelligence (AI) model, to acquire at least one output data, acquiring first usage information, based on the acquired at least one output data, acquiring first reconstruction information for reconstructing the first AI model, based on the acquired first usage information, and reconstructing the first AI model, based on the acquired first reconstruction information, to acquire a second AI model.

Abstract: Systems and methods for aggregating data. The system is configured to receive metadata from an interactive graphical user interface (GUI) of a user device, aggregate field values from the data stored on one or more databases based on the received metadata and generate filter instructions based on the received metadata. The system is further configured to transmit the aggregated field values and the filter instructions to the user device, receive a user-customized filter set and subscription request for a synthetic symbol associated with the user-customized filter set from the user device, and create the synthetic symbol responsive to the subscription request. Moreover, the system aggregates one or more data values from the data stored on the databases associated with the created synthetic symbol and generates instructions to display the data values on the interactive GUI in accordance with the user-customized filter set associated with the created synthetic symbol.

Abstract: Disclosed herein is an apparatus for analyzing a video shot. The apparatus includes at least one program, memory in which the program is recorded, and a processor for executing the program. The program may include a frame extraction unit for extracting at least one frame from a video shot, a shot composition and camera position recognition unit for predicting shot composition and a camera position for the extracted at least one frame based on a previously trained shot composition recognition model, a place and time information extraction unit for predicting a shot location and a shot time for the extracted at least one frame based on previously trained shot location recognition model and shot time recognition model, and an information combination unit for combining pieces of information, respectively predicted for the at least one frame, for each video shot and tagging the video shot with the combined pieces of information.

Abstract: A system for detecting a conflict of interest between entities includes a server and a database. The server selects a category to be evaluated for detecting whether or not two entities have a conflict of interest. A tree structure is created including the category and child and descendent categories of the category. The category and the child and descendent categories are hierarchically arranged as nodes of the tree structure. For the two entities, a conflict potential value is iteratively computed for each node based upon a previous conflict potential value of that node, previous conflict potential values of neighbouring nodes and distances between that node and neighbouring nodes. A conflict index value is computed for each node based upon the conflict potential value for that node for each entity. A conflict of interest between the entities is detected if the conflict index value of a node matches a predefined criteria.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a generator neural network and an encoder neural network. The generator neural network generates, based on a set of latent values, data items which are samples of a distribution. The encoder neural network generates a set of latent values for a respective data item. The training method comprises jointly training the generator neural network, the encoder neural network and a discriminator neural network configured to distinguish between samples generated by the generator network and samples of the distribution which are not generated by the generator network. The discriminator neural network is configured to distinguish by processing, by the discriminator neural network, an input pair comprising a sample part and a latent part.

Abstract: Datasets are available from different dataset servers and often lack well-defined metadata. Thus, comparing datasets is difficult. Additionally, there might be different versions of the same dataset which makes the search even more difficult. Using systems and methods described herein, quality scores, dataset versioning, topic identification, and semantic relatedness metadata is stored about datasets stored on dataset servers. A user interface is provided to allow a user to search for datasets by specifying search criteria (e.g., a topic and a minimum quality score) and to be informed of responsive datasets. The user interface may further inform the user of the quality scores of the responsive datasets, the versions of the responsive datasets, or other metadata. From the search results, the user may select and download one or more of the responsive datasets.

Abstract: An image determination device includes: a training model which outputs, on the basis of an image to be examined, output data indicating a determination result about the image; a training part which trains the training model to output, by using training data including a training image and label data, output data indicating the label data associated with the training image, when the training image is input to the training model; a dividing part which divides the training data into a plurality of pieces of sub-training data; a measurement part which measures accuracy of determination when the training part trains the training model by using each of the plurality of pieces of sub-training data; and selection part which selects at least any one among the plurality of pieces of sub-training data on the basis of the accuracy of determination.

Abstract: A method for detecting blurriness of a human face in an image includes: performing a face detection in a target image; when a human face is detected in the target image, cropping the human face from the target image to obtain a face image and inputting the face image to a first neural network model to perform preliminary detection on a blurriness of the human face in the face image to obtain a preliminary detection result; and when the preliminary detection result meets a deep detection condition, inputting the face image to a second neural network model to perform deep detection on the blurriness of the human face in the face image to obtain a deep detection result.

Abstract: A method and system that efficiently selects sensors without requiring advanced expertise or extensive experience even in a case of new machines and unknown failures. An abnormality detection system includes a storage unit for storing a latent variable model and a joint probability model, an acquisition unit for acquiring sensor data that is output by a sensor, a measurement unit for measuring the probability of the sensor data acquired by the acquisition unit based on the latent variable model and the joint probability model stored by the storage unit, a determination unit for determining whether the sensor data is normal or abnormal based on the probability of the sensor data measured by the measurement unit, and a learning unit for learning the latent variable model and the joint probability model based on the sensor data output by the sensor.

Abstract: A neural network is trained to process input data and generate a classification value that characterizes the input with respect to an ordered continuum of classes. For example, the input data may comprise an image and the classification value may be indicative of a quality of the image. The ordered continuum of classes may represent classes of quality of the image ranging from “worst”, “bad”, “normal”, “good”, to “best”. During training, loss values are determined using an ordered classification loss function. The ordered classification loss function maintains monotonicity in the loss values that corresponds to placement in the continuum. For example, the classification value for a “bad” image will be less than the classification value indicative of a “best” image. The classification value may be used for subsequent processing. For example, biometric input data may be required to have a minimum classification value for further processing.

Abstract: A method for diagnosing one or more diseases of the respiratory tract for a patient including the steps of: acquiring cough sounds from the patient; processing the cough sounds to produce cough sound feature signals representing one or more cough sound features from the cough segments; obtaining one or more disease signatures based on the cough sound feature signals; and classifying the one or more disease signatures to deem the cough segments as indicative of one or more of said diseases; wherein the step of obtaining the one or more disease signatures based on the cough sound feature signals includes applying the cough sound features to each of one or more pre-trained disease signature decision machines, each said decision machine having been pre-trained to classify the cough sound features as corresponding to either a particular disease or to a non-disease state or as corresponding to first particular disease or a second particular disease different from the first particular disease.

Abstract: State of art techniques that utilize spatial association based Table structure Recognition (TSR) have limitation in selecting minimal but most informative word pairs to generate digital table representation. Embodiments herein provide a method and system for TSR from an table image via deep spatial association of words using optimal number of word pairs, analyzed by a single classifier to determine word association. The optimal number of word pairs are identified by utilizing immediate left neighbors and immediate top neighbors approach followed redundant word pair elimination, thus enabling accurate capture of structural feature of even complex table images via minimal word pairs.

Abstract: Systems, methods, and computer-readable storage media for using bots. One method includes receiving, from a user device, an input indicative of a selection of a bot, and retrieving data associated with a user of the user device. Further, the method includes identifying a selection of bot options based on analyzing the user data, and presenting, via the user device, the selection of bot options. Further, the method includes receiving, from the user device, a selection of an option from the selection of bot options, and retrieving user payment information and shipping information based on the user data and populate one or more fields associated with the selection of the option. Further, the method includes presenting, via the user device, a confirmation page to confirm the populated one or more fields.

Abstract: A data processing method and device are provided. The method includes: extracting a plurality of data sets from unlabeled data; and for each data set, creating a plurality of sample sets by assigning labels to data samples in the data set, respectively training, for each sample set created from the data set, a classifier by using the sample set and labeled data, obtaining a sample set that corresponds to a trained classifier with the highest performance, and adding the obtained sample set to a candidate training set. Each sample set includes the first preset number of data samples with respective labels, the labels of the data samples in each sample set constitutes a label combination, and label combinations corresponding to different sample sets are different from each other. The method also includes adding a second preset number of sample sets in the candidate training set to the labeled data.

Abstract: The present disclosure is generally directed to an apparatus for outputting at least a recipe to a user is described. The apparatus may include at least a processor, and a memory communicatively connected to the processor. The memory may contain instructions for configuring the at least a processor to receive user data from a user. The processor may be configured to classify the user data to a one or more phenotypic clusters and assign the user one or more cohort labels as a function of the one or more phenotypic clusters. Further, the processor may be configured to generate alimentary data as a function of the one or more phenotypic cluster and the one or more cohort labels. Moreover, the processor may be configured to output an alimentary program to the user.

Abstract: A computer that identifies the video. The computer annotates the video using a deep learning tool. The computer analyzes the annotated video to highlight a dangerous condition. The computer identifies a video from a repository with the dangerous condition. The computer analyzes the video and the video from the repository using a similarity analysis. The computer determines a score based on the annotated video and based on comparing the video to the video from the repository and based on determining the score is above a threshold value, the computer generates an action.

Abstract: A method and apparatus for data analysis according to various aspects of the present invention is configured to test a set of components and generate test data for the components. A diagnostic system automatically analyzes the test data to identify a characteristic of a component fabrication process by recognizing a pattern in the test data and classifying the pattern using a neural network.

Abstract: Some embodiments employ a novel procedure of training an artificial intelligence system (e.g., set of deep neural networks) for anomaly detection in applications such as natural language processing and computer security. Token sequences selected from a training corpus are distorted according to at least one of a plurality of pre-determined sequence transformations, before being fed to a sequence analyzer. In turn, the sequence analyzer is trained to correctly guess which transformation was used to generate the respective input token sequence.

Abstract: Disclosed is a method that includes training, at a client, a part of a deep learning network up to a split layer of the client. Based on an output of the split layer, the method includes completing, at a server, training of the deep learning network by forward propagating the output received at a split layer of the server to a last layer of the server. The server calculates a weighted loss function for the client at the last layer and stores the calculated loss function. After each respective client of a plurality of clients has a respective loss function stored, the server averages the plurality of respective weighted client loss functions and back propagates gradients based on the average loss value from the last layer of the server to the split layer of the server and transmits just the server split layer gradients to the respective clients.

Abstract: Aspects of the subject matter disclosed herein include methods, systems, and other techniques for training, in a first phase, an object classifier neural network with a first set of training data, the first set of training data including a first plurality of training examples, each training example in the first set of training data being labeled with a coarse-object classification; and training, in a second phase after completion of the first phase, the object classifier neural network with a second set of training data, the second set of training data including a second plurality of training examples, each training example in the second set of training data being labeled with a fine-object classification.

Abstract: Generally discussed herein are devices, systems, and methods for generating an embedding that is both local string dependent and global string dependent. The generated embedding can improve machine learning (ML) model performance. A method can include converting a string of words to a series of tokens, generating a local string-dependent embedding of each token of the series of tokens, generating a global string-dependent embedding of each token of the series of tokens, combining the local string dependent embedding the global string dependent embedding to generate an n-gram induced embedding of each token of the series of tokens, obtaining a masked language model (MLM) previously trained to generate a masked word prediction, and executing the MLM based on the n-based induced embedding of each token to generate the masked word prediction.

Abstract: Systems and methods for determining neural network brittleness are disclosed. For example, the system may include one or more memory units storing instructions and one or more processors configured to execute the instructions to perform operations. The operations may include receiving a modeling request comprising a preliminary model and a dataset. The operations may include determining a preliminary brittleness score of the preliminary model. The operations may include identifying a reference model and determining a reference brittleness score of the reference model. The operations may include comparing the preliminary brittleness score to the reference brittleness score and generating a preferred model based on the comparison. The operations may include providing the preferred model.