7 results about "Descent direction" patented technology

The invention discloses an intelligent design method and system for a machining clamp based on Zhongwishi 3D. The core innovation of the method is that a'dynamic time variant prime field 'and'reverse differential evolution' mechanism is constructed; in multi-physics field simulation of virtual manufacturing, interference potential energy of a tool time sequence envelope and a constraint unit is calculated in real time; a Jacobian matrix of a parameter space and a potential energy space is established, the steepest descent direction of the interference gradient is analyzed, and a Newton-Raphson iteration method is used for driving a tool model to conduct self-adaptive geometric deformation or pose transformation. According to the method, full-process automation from geometric analysis, rule reasoning and physical field simulation to parameter closed-loop optimization is achieved, the problems of CAD / CAM data fault and open-loop iteration in traditional tool design are solved, and mathematical-level global convergence and machining safety of technological equipment are ensured.

This invention belongs to the field of image generation technology, specifically relating to a semantic alignment method and system for image generation based on a diffusion model with dual-loss optimization. The method includes: extracting cross-attention maps corresponding to each topic word in the text prompt during denoising; calculating aggregation loss and isolation loss based on the cross-attention maps, whereby the aggregation loss aggregates attention activations corresponding to the same topic word, and the isolation loss separates attention activations corresponding to different topic words; constructing a joint loss function based on the aggregation loss and isolation loss, and updating the latent representation along the gradient descent direction; inputting the updated latent representation into the denoising process at the next time step, until a preset number of denoising steps is reached, generating the output image. This invention optimizes the aggregation loss and isolation loss, simultaneously solving the problems of internal scattering and overlap in the attention maps, and can be applied to downstream tasks such as controllable layout generation and personalized generation.

This application provides a point cloud processing system and method. The system includes: a PL (Plug-in) terminal and a PS (Power Supply) terminal. The PS terminal sends point cloud data calculation commands to the PL terminal. A hash filtering module generates point cloud data read commands based on the point cloud data calculation commands, sends the point cloud data read commands to a data transfer module, and processes the initial point cloud data in real time using a pipeline to obtain processed point cloud data. A gradient calculation module performs gradient calculations on the processed point cloud data in parallel using a data stream through multiple calculation sub-modules connected by FIFO or ping-pong buffers to obtain multiple gradient vectors. An application processor on the PS terminal calculates the descent direction of the multiple gradient vectors and determines whether the descent direction converges. If the descent direction converges, the point cloud computing is terminated. Thus, the provided point cloud processing scheme can improve the throughput of point cloud data processing and reduce latency.

A path planning method based on analytical gradients of a neural network includes: acquiring sample datasets of a rocket under multiple typical flight conditions; training the neural network with flight control variables as input and aerodynamic coefficients as output until convergence is met, resulting in a surrogate model for fitting the rocket's aerodynamic characteristics; collecting the rocket's current state vector and control vector, inputting the control vector into the surrogate model to obtain the current aerodynamic coefficients, and constructing the rocket's flight state equation; based on the flight state equation, using an automatic differentiation method, analytically calculating the gradient with respect to the control vector to form a gradient sensitivity matrix; using the gradient sensitivity matrix to determine the descent direction; and iteratively updating the control vector using a numerical optimization algorithm until the termination condition is met, outputting the optimal flight path. The optimal flight path obtained by this method more closely matches the rocket's actual aerodynamic environment, exhibiting higher terminal accuracy.

This invention relates to the field of surveying and mapping technology, and proposes a contour gradient checking method, system, and device, comprising: acquiring all contour line entities and their corresponding contour line information; determining whether any contour line entity and its information have been stored in a cache unit; if not, searching for adjacent contour lines along the gradient descent direction of the contour line entity, and storing the matched adjacent contour lines and their information in the cache unit; searching for the upper and lower contour lines of the contour line entity according to the cache unit storing the contour line entity and its information, or the corresponding cache unit of the contour line entity; if upper and lower contour lines exist, comparing the contour interval of the adjacent upper and lower contour lines with a set value, and storing the comparison result; if no upper and lower contour lines exist, checking the next contour line entity; repeating the above steps until all contour line entities have been checked; and displaying the comparison result of the upper and lower contour lines of the target contour line entity.

This disclosure provides an integrated clock gate placement method, apparatus, device, and storage medium based on gradient descent. The method includes: determining the anchor point of the integrated clock gate based on the register position of the register controlled by the integrated clock gate; determining a corresponding line length loss function based on a first distance between the integrated clock gate and the anchor point, and a second distance between the integrated clock gate and the register; a first weight of the first distance being greater than a second weight of the second distance; determining a total loss function based on the line length loss function and a placement loss function; the placement loss function including at least one of a density loss function, a congestion loss function, or a timing loss function; and updating the position of the integrated clock gate based on the gradient descent direction of the total loss function to obtain a placement position that satisfies the convergence condition of the total loss function. This method achieves multi-dimensional improvements in placement quality while ensuring clock tree performance.

The invention discloses a large language model zero-order fine tuning method and system based on low-rank projection matrix learning, and belongs to the technical field of model fine tuning, and the method comprises the steps: during low-rank projection matrix learning, freezing large language model parameters to be used for text processing, utilizing small-batch training sample data, updating a low-rank projection matrix through zero-order optimization, and carrying out zero-order fine tuning on the low-rank projection matrix; the low-rank projection matrix can be multiplied with the low-dimensional random matrix when the model parameters are updated, disturbance with the same dimension as the model parameters is constructed, and the disturbance direction in a low-rank disturbance space formed by the low-rank projection matrix can be close to the descending direction of the loss function by adjusting the low-rank disturbance space; and constructing a low-rank disturbance matrix by utilizing the low-rank projection matrix obtained by learning when model parameters are updated, and performing zero-order gradient estimation and updating on the model parameters of the large language model by combining all training sample data. Therefore, the text processing capability of the large language model is improved while the optimization efficiency in the high-dimensional parameter space is improved and the gradient variance is reduced.