Telefonaktiebolaget Lm Ericsson (publ)

Zi Decuma Ab

Nuance Communications, Inc.

Cerence Operating Company

Jakob Sternby

Luis Barriga

Michael Liljenstam

Erik Thormarker

Björn Johansson

Jonas Andersson

Christer Friberg

Lars Jonas Morwing

Jonas Morwing

Anders Holtsberg

Vasileios Giannokostas

Seonghyun Kim

Training a neural network and embedding a watermark in the network to prove ownership. The network includes a plurality of trainable parameters associated with network nodes in which the plurality of trainable parameters is split into a first set of trainable parameters and a second set of trainable parameters. A first set of training samples is input to the network and the network is trained by iterating the first set of samples through the network to update the first set of parameters and hindering the second set of parameters to be updated during iteration of the first set of samples. A second set of samples is input and the watermark is embedded by iterating the second set of samples through the network to update the second set of parameters and hindering the first set of parameters to be updated during iteration of the second set of samples.

Neural network watermarking

The present disclosure relates to a computer-implemented method and an apparatus for classifying anomalies of one or more feature-associated anomalies in network data traffic between devices in a first part of a network and devices in a second part of the network. The method comprises retrieving at least one network data traffic sample and determining one or more feature-associated anomaly scores for the retrieved at least one network data traffic sample. The method further comprises determining feature importance of each feature of a feature-associated anomaly score and classifying one or more anomalies based on the determined one or more feature-associated anomaly scores and the determined feature importance.

Computer-implemented method and arrangement for classifying anomalies

A neural network having one or more public parts and one or more confidential parts is trained to perform a primary task. A deployment instantiation of the neural network is trained based on optimal performance of the primary task, and based on sub-optimal performance of the primary task conditioned on the confidential parts of the deployment instantiation being inaccessible. An adversary instantiation of the neural network is trained based on optimal performance of the primary task conditioned on the public parts being identical for the deployment instantiation and for the adversary instantiation, and conditioned on the confidential parts of the deployment instantiation being inaccessible. The training of the deployment instantiation and the training of the adversary instantiation are based on a plurality of training data samples, and are performed iteratively by alternating between the training of the deployment instantiation and the training of the adversary instantiation.

Neural network confidentiality

The present disclosure relates to a method and an apparatus for training a model for detecting anomalies in network data traffic between devices in a first part of a network and devices in a second part of the network. The method comprises collecting feature samples of network data traffic at a monitoring point between a first and a second part of the network, and training the model for detecting anomalies on the collected feature samples using a plurality of anomaly detection, AD, trees. The training comprises creating the plurality of AD trees using respective subsets of the collected feature samples, at least some of the AD tree comprising subspace selection nodes and anomaly-catching nodes to a predetermined AD tree depth limit. Each subspace selection node is arranged to bisect a set of feature samples reaching the subspace selection node to at least one anomaly-catching node when a number of feature samples leaving the subspace selection node for the at least one anomaly-catching node is below a predetermined threshold. The disclosure also relates to a method and an apparatus for anomalies in network data traffic using said model.

Method and arrangement for detecting anomalies in network data traffic

A system for inputting and processing handwritten, multi-character text may comprise a handwriting recognition subsystem, a word completion subsystem, and an audio feedback system. The handwriting recognition system may be configured to capture a series of handwritten characters formed by a user and to convert the handwritten characters into a set of candidate partial text strings. The word completion subsystem may be configured to identify if a candidate partial text string constitutes a word segment and if so, generate one or both of (i) at least one clarifying word and (ii) at least one clarifying phrase that includes the clarifying word. The word segment may be an arbitrary string and not correspond to a valid complete word in a language associated with the system. The audio feedback subsystem may be configured to produce an audio representation of the word segment(s), the clarifying word(s), and the clarifying phrase(s).

Multi-character text input system with audio feedback and word completion

A system and method for receiving character input from a user includes a programmed processor that receives inputs from the user and disambiguates the inputs to present character sequence choices corresponding to the input characters. In one embodiment, a first character input is received and a corresponding first recognized character is stored in a temporary storage buffer and displayed to the user for editing. After a predetermined number of subsequent input characters and/or predetermined amount of time without being edited, the system determines that the first recognized character is the intended character input by the user and removes the first recognized character from the buffer, thereby inhibiting future editing.

Text/character input system, such as for use with touch screens on mobile phones

The present invention generally describes a method for classifying a line segment of a handwritten line into a reference feature set, wherein said handwritten line comprises one or several curves representing a plurality of symbols. First, sample data representing said handwritten line is received. Next, a sample line segment in said received sample data is identified by detecting a sample line segment start point (SLSSP) and a sample line segment end point (SLSEP). Then, a sample feature set of said identified sample line segment is determined. Finally, the determined sample feature set is matched to a reference feature set among a plurality of reference feature sets.

Method for character recognition

A method for recognition of a handwritten pattern comprising one or more curves is presented. The method comprises a step of receiving sample data representing the handwritten pattern. The method further comprises a step of segmenting the handwritten pattern by detecting segmentation points on each curve, and by dividing the handwritten pattern into segments. Further, the method comprises a step of comparing the handwritten pattern to templates wherein the comparing comprises a step of normalizing said segments according to a scheme which is independent of the templates to which the segments are to be compared, and a step of determining matching measures for selecting at least one sequence of templates representing a recognintion candidate of the handwritten pattern.

Method for additive character recognition and an apparatus thereof

A method for determining at least one recognition candidate for a handwritten pattern comprises selecting possible segmentation points in the handwritten pattern for use in segmenting and recognizing the handwritten pattern. The method further may comprise comparing segments of the handwritten pattern to templates. The comparison may return segment candidates forming possible recognition results of the segments of the handwritten pattern. The method further comprises forming a representation of sequences of segment candidates, said representation comprising data blocks corresponding to segmentation points, wherein a data block comprises references to data blocks corresponding to subsequent segmentation points. The reference may comprise information of segment candidates. The method further may comprise comparing the representation of the sequences of segment candidates to a dictionary, finding sequences of segment candidates that correspond to allowed sequences of symbols in the dictionary, and returning at least one of these allowed sequences of symbols as a recognition candidate for the handwritten pattern.

Handwriting recognition using a graph of segmentation candidates and dictionary search

A method for recognition of a handwritten pattern comprises selecting core points among a sequence of detected points of the handwritten pattern. The core points are selected for use in segmenting and recognizing the handwritten pattern. The method may further comprise identifying potential diacritics in the sequence of points of the handwritten pattern, determining features of each core point, assigning at least one feature associated with identified potential diacritics to each core point of a subset of core points, and comparing the handwritten pattern to templates.

Handling of diacritic points

A method for recognition of a handwritten pattern comprises selecting core points among a sequence of detected points of the handwritten pattern. The core points are selected for use in segmenting and recognizing the handwritten pattern. The method further comprises determining features of each core point, and comparing the handwritten pattern to templates. The comparing comprises stepwise analyzing the core points in sequence by matching the features of sequences of core points that either start with the first core point or the last core point of a previous sequence to said templates and calculating a distance value, and assigning a cumulative distance value to the last core point in the matched sequence of core points, whereby a smallest cumulative distance value is assigned to the last core point and corresponds to a sequence of matched templates which represent a possible recognition result of the handwritten pattern.

Growing community of inventors

Lund, Sweden

Jakob Sternby