7results about How to "Reduce training overhead" patented technology

The invention discloses an edge terminal-oriented memory efficient model fine tuning method and system, which realize decoupling of a backbone network and a fine tuning process by constructing a parallel side network module, and avoid video memory occupation caused by backbone gradient return. A double-adapter module is arranged, modeling is carried out on same-layer features and cross-layer context information, and the feature expression precision is improved. And constructing a feature fusion module, and performing adaptive weighting on different path outputs through learnable gating. A backbone grouping module is integrated, dynamic grouping is carried out on a backbone network based on interlayer similarity, and redundant modules and calculation overhead are reduced. And through the fine tuning execution module, only back propagation and parameter updating are carried out on the side network, and a trunk freezing state is kept, so that the memory and training cost is reduced. According to the method, the memory efficiency and the training speed of the edge fine adjustment process are improved, high-precision model self-adaption is achieved with extremely low calculation burden, and application in the fields of cloud edge collaborative intelligence and the like of unmanned aerial vehicles, robots, vehicle-mounted terminals and the like is effectively supported.

This invention belongs to the interdisciplinary field of artificial intelligence and cheminformatics, and provides a method, inference method, and apparatus for constructing a molecular multimodal large language model based on parameter space alignment. The method includes: constructing an instance set; constructing a molecular multimodal large language model, including an adaptive weight generator, a graph neural network encoder, and a large language model; training the molecular multimodal large language model using the instance set to update the parameters of the adaptive weight generator; the graph neural network encoder extracting node-level feature representations of the molecular structure diagram; the adaptive weight generator obtaining the low-rank weight update amount of the target weight component in the large language model based on the node-level feature representation, and injecting the low-rank weight update amount into the original weights of the target weight component; and the large language model after injecting the low-rank weight update amount generating an actual text response based on the input training text command. By integrating the molecular structure into the parameter layer of the large language model, deep perception of the molecular structure is achieved without changing the length of the input sequence.

The application discloses an improved ELM speech enhancement method and device for noise reduction, and belongs to the field of speech signal processing; specifically, first, a noisy speech signal and a clean speech signal are acquired to construct training samples; meanwhile, an improved ELM speech enhancement network is constructed, the training samples are used for training, and a network weight matrix is saved; then, a new noisy speech signal is acquired, input into the trained enhancement network to output a noise-reduced enhanced speech signal; the device comprises an audio acquisition module, an offline training module, an online running module and an audio playing module; the audio acquisition module acquires speech signal samples used for training and inputs the offline training module to train the improved ELM speech enhancement network and save the network weight matrix; a new noisy speech signal is acquired and input into the online running module to directly output a noise-reduced enhanced speech signal; and the audio playing module is used for playing. The application reduces training cost and has lower processing time delay.

The invention relates to a multi-mode millimeter wave beam prediction method based on multi-task learning, and belongs to the technical field of millimeter wave communication. The method comprises the following steps: converting a beam prediction task into a deep learning optimization task based on a geometric channel model in a dynamic communication scene; preprocessing the image, the three-dimensional point cloud and the user motion information acquired by the base station to complete region-of-interest extraction, point cloud downsampling and space coordinate conversion; through a cross-modal gating fusion module, adaptively extracting and carrying out weighted fusion on the multi-modal features; and constructing a multi-task learning framework, cooperatively training a beam prediction main task and blocking prediction and reflection intensity prediction auxiliary tasks, correcting an optimal beam probability by using physical constraint information output by the auxiliary tasks, and selecting an optimal beam. According to the method, the beam training overhead and the communication delay are remarkably reduced, the environmental limitation of single-mode sensing is effectively overcome, and the prediction robustness of the system in a complex dynamic scene is enhanced while the beam prediction precision is improved.

The application discloses a method and device for single-sample action recognition based on WiFi, which physically models the WiFi action recognition scene, generates virtual action data from basic actions, thereby enriching the data set and reducing the cost of collecting real data. Based on supervised learning and meta-learning, a single-sample action recognition framework is designed and implemented. The framework uses a traditional supervised learning mechanism for first-stage training on a virtual action data set, uses a single-sample meta-learning mechanism for second-stage training on a basic action data set, and uses a single-sample meta-learning mechanism for model fine-tuning on a single-sample data set of a new action, thereby obtaining a model capable of accurately recognizing the new action. The training of the two stages only needs to be performed at the initial deployment, and if there is a need to change the action type subsequently, only the model fine-tuning needs to be performed by using a single-sample data of a new action, thereby significantly reducing the model training cost, being highly scalable, and being applicable to actual application scenarios.

The application relates to a training sample screening method and device based on global semantic modeling. The method comprises the following steps: obtaining at least one training sample set, each training sample set comprising at least one training sample; for each training sample set, inputting each training sample into an artificial intelligence model to obtain a response result of each training sample; the response result is obtained by at least performing self-attention calculation and nonlinear transformation calculation on the training sample by the artificial intelligence model; comparing whether the response result of each training sample meets corresponding labeled data of each training sample to determine a response accuracy of the artificial intelligence model on the training sample set; if the response accuracy meets a preset condition, the training sample set is determined as a target post-training sample set; and the target post-training sample set is used for performing post-training on the artificial intelligence model. By using the method, high-quality and low-cost post-training can be realized.

The present application relates to the technical field of computer vision, in particular to a pig part zero-label recognition method based on discrete Morse theory. A video frame sequence containing pigs is obtained, and instance segmentation of the pigs is performed based on a text prompt through a visual basic model to obtain a binary mask and image embedding features; an Euclidean distance transform terrain field is predicted according to the image embedding features and the background area is removed; the Euclidean distance transform terrain field is taken as a discrete Morse function to determine topological critical points and persistent homology values; an initial skeleton is obtained by skeleton thinning the binary mask, and the terminal branches are pruned according to the persistent homology values to obtain a pig topological skeleton; the head, tail and leg are identified according to the pig center, end points in the pig topological skeleton, head direction and pig body axis to obtain a pig part recognition result. The present application does not require manual key point labeling, can reduce deployment cost and improve cross-scene applicability.