Abstract: Aspects of the present disclosure relate to automated content generation. Embodiments include receiving a prompt from a user. Embodiments further include providing the prompt as an input to a classification model, wherein the classification model has been trained to generate outputs that indicate levels of quantization for a generative machine learning model when provided with input prompts. Embodiments further include receiving, based on the prompt, an output from the classification model indicating a given level of quantization. Embodiments further include providing the prompt as input to a given generative machine learning model based on the output, wherein the given generative machine learning model has been quantized according to the given level of quantization. Embodiments further include generating, via the given generative machine learning model, a response to the prompt.

Abstract: Certain aspects of the disclosure provide techniques for model merging. A method may include, for each respective model of multiple models trained for multiple domains: processing a multiple questions to generate multiple predicted answers; for each respective predicted answer: generating a gradient vector indicating, for each weight of the respective model, a weight change that is needed to minimize a loss value, the loss value being based on the respective predicted answer and an incorrect answer to the respective question; and summing each gradient vector for each respective answer to generate a final gradient vector for the respective model; and combining, based on the final gradient vector generated for each respective model, at least one weight of the weight(s) associated with each respective model of the multiple models to obtain a single merged model associated with the plurality of domains.

Abstract: Certain aspects provide a method for merging multiple transformer models. The method includes obtaining a first weight matrix and a second weight matrix, the first weight matrix comprising a first layer of parameters of a first transformer model, the second weight matrix comprising a second layer of parameters of a second transformer model; mapping the first weight matrix to a first point on a curved manifold; mapping the second weight matrix to a second point on the curved manifold; generating a first optimized weight matrix based on a first manifold-constrained optimization of the first point and the second point on the curved manifold; and generating a first merged transformer model of the first transformer model and the second transformer model by mapping the first optimized weight matrix to a first merged layer of parameters of the first merged transformer model.

Abstract: Certain aspects of the disclosure provide techniques for training data generation for large language model (LLM) training and/or benchmarking. A method generally includes obtaining domain data associated with a domain; generating a prompt based on configuration parameter(s) included in a configuration file, the prompt comprising: a request to generate training data for the domain based on the domain data, wherein the training data comprises: a first plurality of question and answer pairs; a conversation comprising a first plurality of questions and a first plurality of answers corresponding to the first plurality of questions; a second plurality of questions; or a question and a second plurality of answers corresponding to the question; guideline(s) for generating the training data; example training data for the domain; and the domain data; prompting an LLM with the prompt to generate the training data; and receiving, from the LLM, the training data based on the prompt.

Abstract: An example system includes a processor to receive unlabeled data, few-shot training data, and a pre-trained model. The processor can split the unlabeled data into a number of groups corresponding to different perspectives. The processor can generate weakly labeled data for each of the number of groups using a respective associated heuristic. The processor can inter-train a model for each different perspective based on respective weakly labeled data. The processor can fine-tune each inter-trained model based on the few-shot training data for each different perspective to generate a final model for each different perspective.

Abstract: A specification of a property required to be upheld by a computerized machine learning system is obtained. A training data set corresponding to the property and inputs and outputs of the system is built. The system is trained on the training data set. Activity of the system is monitored before, during, and after the training. Based on the monitoring, performance of the system is evaluated to determine whether the system, once trained on the training data set, upholds the property.

Abstract: Detecting sequences of computer-executed operations, including training a BLSTM to determine forward and backward probabilities of encountering each computer-executed operations within a training set of consecutive computer-executed operations in forward and backward execution directions of the operations, and identifying reference sequences of operations within the training set where for each given one of the sequences the forward probability of encountering a first computer-executed operation in the given sequence is below a predefined lower threshold, the forward probability of encountering a last computer-executed operation in the given sequence is above a predefined upper threshold, the backward probability of encountering the last computer-executed operation in the given sequence is below the predefined lower threshold, and the backward probability of encountering the first computer-executed operation in the given sequence is above the predefined upper threshold, and where the predefined lower threshold

Abstract: During end-to-end training of a Deep Neural Network (DNN), a differentiable estimator subnetwork is operated to estimate a functionality of an external software application. Then, during inference by the trained DNN, the differentiable estimator subnetwork is replaced with the functionality of the external software application, by enabling API communication between the DNN and the external software application.

Abstract: Embodiments may provide techniques to generate training data for summarization of complex documents, such as scientific papers, articles, etc., that are scalable to provide large scale training data. For example, in an embodiment, a method may be implemented in a computer system and may comprise collecting a plurality of video and audio recordings of presentations of documents, collecting a plurality of documents corresponding to the video and audio recordings, converting the plurality of video and audio recordings of presentations of documents into transcripts of the plurality of presentations, generating a summary of each document by selecting a plurality of sentences from each document using the transcript of the that document, generating a dataset comprising a plurality of the generated summaries, and training a machine learning model using the generated dataset.

Abstract: Embodiments may provide techniques to generate training data for summarization of complex documents, such as scientific papers, articles, etc., that are scalable to provide large scale training data. For example, in an embodiment, a method may be implemented in a computer system and may comprise collecting a plurality of video and audio recordings of presentations of documents, collecting a plurality of documents corresponding to the video and audio recordings, converting the plurality of video and audio recordings of presentations of documents into transcripts of the plurality of presentations, generating a summary of each document by selecting a plurality of sentences from each document using the transcript of the that document, generating a dataset comprising a plurality of the generated summaries, and training a machine learning model using the generated dataset.

Abstract: Detecting sequences of computer-executed operations, including training a BLSTM to determine forward and backward probabilities of encountering each computer-executed operations within a training set of consecutive computer-executed operations in forward and backward execution directions of the operations, and identifying reference sequences of operations within the training set where for each given one of the sequences the forward probability of encountering a first computer-executed operation in the given sequence is below a predefined lower threshold, the forward probability of encountering a last computer-executed operation in the given sequence is above a predefined upper threshold, the backward probability of encountering the last computer-executed operation in the given sequence is below the predefined lower threshold, and the backward probability of encountering the first computer-executed operation in the given sequence is above the predefined upper threshold, and where the predefined lower threshold

Abstract: A specification of a property required to be upheld by a computerized machine learning system is obtained. A training data set corresponding to the property and inputs and outputs of the system is built. The system is trained on the training data set. Activity of the system is monitored before, during, and after the training. Based on the monitoring, performance of the system is evaluated to determine whether the system, once trained on the training data set, upholds the property.

Abstract: An example system includes a processor to receive a plurality of object aspects of an object to be evaluated using a process, a structure of the process, a plurality of extracted facts from documents, a tree related to the plurality of object aspects and the structure, and a thesis for each leaf in the tree. The processor is also to relate the extracted facts to the theses in the tree. The processor is to generate a score for each leaf corresponding to a fact in the tree. The processor is to generate a thesis score and a thesis summary for each thesis based on the scores and the summaries of related facts for each thesis. The processor is to further generate a final score for the object based on the thesis scores.

Abstract: During end-to-end training of a Deep Neural Network (DNN), a differentiable estimator subnetwork is operated to estimate a functionality of an external software application. Then, during inference by the trained DNN, the differentiable estimator subnetwork is replaced with the functionality of the external software application, by enabling API communication between the DNN and the external software application.

Abstract: A method, system and computer program product, the method comprising: mutually training, using feedback, a generator and a discriminator of a conditional adversarial generative adversarial networks using training item groups, each item group representing events in a time window, the generator comprises a generator Recurrent Neural Network (RNN), the discriminator comprises a discriminator RNN; receiving by the discriminator, discrete sequential data comprising a sequence of item groups comprising an item group representing events in a time window, and item groups representing events in preceding time windows; altering the sequence of item groups into collections of real numbers and providing them to the discriminator RNN; processing the collections by the discriminator RNN to obtain a probability for the item group to comprise an anomaly, in an unsupervised manner; and providing output to a user, the output based on the probability and indicative of a label for the discrete sequential data.

Abstract: An example system includes a processor to receive a plurality of object aspects of an object to be evaluated using a process, a structure of the process, a plurality of extracted facts from documents, a tree related to the plurality of object aspects and the structure, and a thesis for each leaf in the tree. The processor is also to relate the extracted facts to the theses in the tree. The processor is to generate a score for each leaf corresponding to a fact in the tree. The processor is to generate a thesis score and a thesis summary for each thesis based on the scores and the summaries of related facts for each thesis. The processor is to further generate a final score for the object based on the thesis scores.

Abstract: A method for challenge-based authentication of a communication entity to an access network. The access network uses a password-based communication protocol. The method comprises a) pre-supplying to the communication entity a challenge, thereby allowing the communication entity to provide a challenge response, b) supplying to the communication entity a password request, c) receiving the challenge response via the password request, and d) authenticating the communication entity if the challenge response is correct. Presupplying may be during a previous IP session, wherein communication entities are simply given challenges for next time they connect to the hotspot. Alternatively presupplying could be during a brief probationary connection that the access network gives to its users.