14 results about "Sequence transformation" patented technology

In an example point cloud decoding method, at least a portion of a bitstream is entropy decoded to obtain an ordered sequence of transform coefficients. The transform coefficients are arranged such that a first set of the transform coefficients for a first block in a point cloud is followed consecutively by a second set of the transform coefficients for a second block in the point cloud; the transform coefficients in the first set are arranged in order of increasing frequency; and the transform coefficients in the second set are arranged in order of decreasing frequency. The first and second blocks in the point cloud are reconstructed using the entropy decoded transform coefficients.

The application provides a pulse repetition frequency type recognition method and device based on graph domain mapping and feature enhancement, and relates to the technical field of radar signal processing. The method comprises the following steps: determining a pulse repetition frequency sequence corresponding to a screened pulse repetition frequency; taking the reciprocal of the pulse repetition frequency sequence to obtain a pulse repetition interval sequence; dividing the pulse repetition interval sequence into multiple subsequences according to a preset length; transforming the multiple subsequences into a Gram difference angle field image according to a graph domain mapping algorithm, and dividing the image into training samples and test samples; constructing a pulse repetition frequency type recognition network, inputting the training samples into the pulse repetition frequency type recognition network, training the pulse repetition frequency type recognition network, and obtaining a trained pulse repetition frequency type recognition network; and inputting the test samples into the trained pulse repetition frequency type recognition network to recognize multiple pulse repetition frequency types. In this way, the recognition accuracy and robustness of the pulse repetition frequency type are improved in a complex non-ideal scene with a high proportion of missed pulses and false pulses.

The invention discloses a power grid line power return behavior identification and classification method based on a characteristic subsequence, and relates to the technical field of power grid operation state monitoring. The method comprises the following steps: performing preprocessing and sample construction on historical active power time sequence data to form an annotation data set; training an interpretable feature subsequence bottleneck model based on feature subsequence transformation, and providing a decision basis with clear physical significance through multi-scale feature subsequence similarity calculation and linear classification; training the deep neural network in parallel to capture complex features; designing a gating function based on prediction confidence to realize a dynamic routing decision; performing combined training to form a power grid line power return behavior network; the method effectively solves the black box problem of a traditional deep learning model and the defect that the physical significance of the features is not clear, achieves the unification of interpretability and high precision, can quickly and accurately classify and recognize the real-time power data, and provides reliable support for the safe and stable operation of a power grid.

Techniques for improving the training and prompt phase inferencing of a long sequence transformer are disclosed. A service shards an activation matrix and a weight matrix into chunks. The service distributes the activation matrix chunks and the weight matrix chunks to multiple computer systems. The activation matrix chunk remains stationary at each computer system. The weight matrix chunks, on the other hand, are subjected to a gathering operation in which each weight matrix chunk is used for a matrix multiplication operation against the activation matrix chunk and then replaced by a newly acquired weight matrix chunk. While the matrix multiplication operation is occurring, the service transmits the current weight matrix chunk to a new computer system and receives a new weight matrix chunk from another computer system.

This application belongs to the field of radar data processing technology, specifically disclosing a radar track filtering method, apparatus, electronic device, and storage medium. The method includes: acquiring current radar track observation data; inputting the radar track observation data into a radar track filtering model to obtain the current radar track filtering result output by the radar track filtering model; the radar track filtering model is trained based on radar track observation data samples and corresponding real track data labels, and is used to perform feature fusion based on sequence transformation features, time dependence features, and shape features obtained by feature extraction from the radar track observation data; and to determine the radar track filtering result based on the fused features. This application can accurately predict radar track data in complex scenarios, significantly improving the accuracy of radar track filtering in complex scenarios, and enhancing the precision and effectiveness of radar track filtering.

A service computes a self-attention of a long sequence transformer. The computation is two-dimensional, with a first dimension being along a Q-dimension and a second dimension being along a KV-dimension. The service determines that the Q-dimension does not carry any data dependencies but that the KV-dimension does carry one or more data dependencies. The service splits the Q-dimension and distributes those splits to a processor grid. The service splits the one or more data dependencies along the KV-dimension and distributes those splits to the processor grid. The service performs a reduction operation to obtain a final result. The service distributes the final result among the processors.

A computer-implemented method for aligning intercorrelated asynchronous time series datasets includes the steps of:(a) retrieving a first time series dataset (x) and a second time series dataset (y), the first and second time series dataset being intercorrelated,(b) segmenting each of the first and second time series dataset (x, y) into a plurality of consecutive smaller segments (xi, yi), all segments of the first and second time series dataset (x, y) having the same length,(c) determining pairs of corresponding segments by associating successive segments of the first time series dataset (x) with corresponding segments of the second time series dataset (y),(d) optimizing, for each pair of corresponding segments, a correlation function to obtain an approximation of a first times series transformation function (f1) and of a second time series transformation function (f2),(e) using the first and second time series transformation functions (f1,f2) to determine a vector of segment shifts (s) whose components contain approximations of the shifts between the first and the second segment in a pair of corresponding segments (xi, yi),(f) applying a multi-model fitting algorithm to the segment shift vector (s), said multi-model algorithm outputting a shift function (fopt) for aligning segments of each pair of corresponding segments (xi, yi), and(g) aligning the first time series dataset (x) with the second series dataset by applying said shift function (fopt) to all pairs of corresponding segments (xi, yi),wherein the first time series transformation function (f1) is parametrized by weights (w1) of a first neural network (N1) and outputs a first highly correlated time series dataset, and wherein the second time series transformation function (f2) is parametrized by weights (w2) of a second neural network (N2) and outputs a second highly correlated time series dataset.

The invention relates to the technical field of computers, and provides a security authentication method and device, equipment, a storage medium and a program product. The method comprises the following steps: in response to an access request of a user, performing sequence transformation based on a hyper-chaos transformation model to generate a current authentication requirement sequence; determining a target reverse behavior pattern based on the user level and the security level; updating the current authentication requirement sequence based on the target reverse behavior pattern to obtain a target authentication requirement sequence; and performing identity security authentication on the user based on the target authentication requirement sequence. Through the above mode, prediction attack and replay attack can be effectively resisted, an attacker is prevented from cracking through mode identification and mode simulation, the security authentication process can be dynamically updated according to the actual situation of a user, the flexibility and adaptability of security authentication are improved, different attack scenes can be effectively coped with, and the security authentication efficiency is improved. And the information security requirement is met.

Embodiments of the present specification provide a method for constructing a processing circuit for a target transform. The target transform is a discrete transform or its inverse transform that transforms an input coefficient sequence into an output coefficient sequence based on a rotation factor. The method includes, for a K-point input coefficient sequence to be processed, determining a plurality of candidate decomposition points corresponding to a k-bit coefficient index. Then, starting from a low-bit decomposition point, for each candidate decomposition point, determining the minimum storage cost under each candidate decomposition point according to a decomposition cost evaluation function through recursive index bit number decomposition for several levels, so as to determine a target decomposition mode of the k-bit coefficient index with the minimum storage cost. The decomposition cost evaluation function limits the storage cost to include the cost of a (n-p)th order first transform, the cost of a pth order second transform, and the local cost for inter-level rotation factor multiplication. According to the target decomposition mode, a corresponding memory for the rotation factor is allocated to form a processing circuit.

Systems and methods are provided for continuous, open-vocabulary decoding of imagined speech from non-invasive brain signals. Consistent with aspects of the disclosed technology, a user may silently compose sentences on diverse topics while high-density fNIRS, MEG, fMRI or similar sensors record neural activity. In one illustrative implementation, the recorded time-series are passed through a sequence-to-sequence transformer (or other encoder or processor) that projects the signals into the embedding space of a large language model. Such brain-derived embeddings may be combined with selectable textual context and injected as a prompt into a LLM, which may complete the sentence, yielding continuous text that reflects the user's internal speech. According to additional aspects, alignment across multiple participants plus rapid per-user fine-tuning improves accuracy and / or generalizability. In our example implementation, experimental results show statistically significant gains in BLEU and BERT similarity metrics. Further, the disclosed technology delivers a portable, adaptable path to non- invasive, thought-driven interaction with Al assistants, wearable controllers and other computing devices.

The application discloses a long-time sequence prediction method based on frequency domain cross enhancement, and relates to the field of time sequence prediction, which comprises the following steps: reversible instance normalization is performed on an input sequence, and the sequence is mapped to a high-dimensional semantic space through dimension expansion; then, the sequence is transformed to a frequency domain by using a discrete Fourier transform, and is decomposed into a real part sequence and an imaginary part sequence; a real part structure enhancement module and an imaginary part structure enhancement module are introduced into the real part and the imaginary part respectively, so that the structural perception ability and feature diversity of an attention matrix are improved; a bidirectional cross attention mechanism is further constructed to realize information interaction between the real part and the imaginary part; then, the sequence is recovered to a time domain through an inverse Fourier transform, and a prediction result is obtained by combining a residual connection, a linear mapping and inverse normalization. The application can effectively alleviate the problems of low rank, excessive focus and insufficient structure expression of traditional self-attention in frequency domain modeling, enhance the modeling ability of long-distance dependence, and improve the precision and generalization performance of multivariate long-time sequence prediction.

In an example point cloud decoding method, at least a portion of a bitstream is entropy decoded to obtain an ordered sequence of transform coefficients. The transform coefficients are arranged such that a first set of the transform coefficients for a first block in a point cloud is followed consecutively by a second set of the transform coefficients for a second block in the point cloud; the transform coefficients in the first set are arranged in order of increasing frequency; and the transform coefficients in the second set are arranged in order of decreasing frequency. The first and second blocks in the point cloud are reconstructed using the entropy decoded transform coefficients.

Embodiments of the present specification provide a method for storing or reading a rotation factor in a target transform. The target transform is a discrete transform or its inverse transform that transforms an input coefficient sequence into an output coefficient sequence based on a rotation factor. The method includes determining a target rotation factor to be stored, whose power is a result of a multiplication of a first factor of a first number of bits and a second factor of a second number of bits modulo a target value; and storing the target rotation factor by a two-level storage manner, the two-level storage manner including storing the target rotation factor at a first address in a rotation factor storage, wherein the rotation factor storage includes a first number of first storage units, the first number being a number of different modulo multiplication results generated by the first factor of the first number of bits and the second factor of the second number of bits; and storing an index value pointing to the first address at a second address in an index storage, wherein the index storage includes a target value of second storage units, the second address corresponding to the power of the target rotation factor; and the target value is greater than the first number.

In various examples, a technique for transforming an input sequence to an output sequence using a machine learning model is disclosed. The technique includes encoding a sequence of inputs in a representation of the sequence of inputs. The technique also includes causing the generation of a sequence of joint probabilities based on the representation of the sequence of inputs and no history of previously predicted output labels. The technique also includes causing the generation of a sequence of output labels based on the sequence of joint probabilities.