Abstract: The present disclosure relates to a method and an apparatus for answering questions in natural language. The method for answering questions in natural language, using a relational database storing image metadata obtained from an artificial intelligence, AI, program, comprises the steps of: receiving a natural language input question in text format; transforming, by a first large language model (LLM1) of a Retrieval-Augmented Generation, RAG, Architecture, the input question into queries with information independence for the database; executing each of the queries with information independence against the database; receiving data in response to the executed queries with information independence; and composing an answer in natural language to the input question from the received data, by a second large language model (LLM2) of a RAG Architecture.

Abstract: The present disclosure relates to an image processing method. The method may include: receiving an input image to be compressed; extracting features of the input image; encoding the extracted features of the input image into a latent feature, zinput, determining, from a background image database which contains one or more background images, a background image that has a latent feature, zbackground, best-matching with zinput; calculating a delta feature, ?z, which is the difference between zinput and zbackground, in accordance with the following formula: ?z=zinput?zbackground; and compressing ?z to obtain a bitstream, Ydifferences.

Abstract: The present disclosure relates to a method and an apparatus for generating images for AI training.

Abstract: A computer-implemented method for training an anomaly detection neural network system comprising an encoder and a decoder is described.

Abstract: The invention relates to a repository-level semantic graph system is used for repository-level code completion, the repository-level semantic graph system includes: a source code initialization module, a source code repository module, a Large Language Models (LLM) encoding module, a repository-level semantic graph, a search module, an expansion module, an update module, a re-ranking module, a node selection module, and a Large Language Models (LLM) decoding module.

Abstract: A computer-implemented method for generating and ranking source code for performing a task is described. The method includes: receiving input data comprising a task description, a code generation prompt and a test case generation prompt; processing the input data using at least one trained code generation neural network to generate a plurality of code solutions and a plurality of test cases; for each code solution, executing the set of candidate source code on the test inputs of the plurality of test cases to generate a plurality of execution outputs; clustering the plurality of code solutions into a plurality of clusters; computing an interaction matrix that specifies functional overlap between the plurality of clusters; determining, for each cluster, a score based the interaction matrix; and ranking the plurality of clusters based on the scores of the plurality of clusters.

Abstract: A neural network system for summarizing source code is described. The neural network system includes one or more computers and one or more non-transitory computer storage media storing instructions that, when executed by the one or more computers, cause the one or more computers to implement: a heterogeneous code representation (HCR) generator; a sub-system; a Transformer encoder neural network; a tree-based encoder neural network; a heterogeneous graph transformer (HGT) neural network; and a graph aggregator. In some implementations, the neural network system includes one or more of a token index selector, a hierarchy-aware cross attention neural network layer, a gating neural network layer, and a Transformer decoder neural network.

Abstract: A computer-implemented method for auto-splitting and classifying an input document into one or more sub-documents using a machine learning system is described. The machine learning system includes a visual splitting neural network, an optical character recognition sub-system, a title classifier, a document classifier, and a grouper sub-system. The method includes: receiving a visual input representing a plurality of pages of the input document; classifying, using the visual splitting neural network, each page of the input document into a respective template of a plurality of templates; for each page of the input document, determining a final document type that the page belongs to; and grouping, using a grouper sub-system, the plurality of pages of the input document into one or more sub-documents based on (i) the respective template of each page and (ii) the respective final document type that each page belongs to.

Abstract: A system including one or more computers and one or more storage devices storing instructions that, when executed by the one or more computers, cause the one or more computers to implement a group-equivariant convolutional neural network configured to process a network input including a set of 3D points to generate an output tensor representing transformed features of the network input is described. The group-equivariant convolutional neural network includes a grouping layer, a switching layer, a pre-processing layer, a group-equivariant layer, a first subnetwork, an average pooling layer, a second subnetwork, an output subnetwork, and a final output layer.