Abstract: This disclosure describes a framework for analyzing dubbed audio segments (audio translations converted into translated speech) of videos where the dubbed audio segments are generated in real time, including being generated locally on a client device. For instance, this disclosure describes a video dubbing system that utilizes various lightweight machine learning models to determine the dubbing quality (e.g., a dubbing quality score) of a real-time generated dubbed segment and identify the cause of low-quality dubbing segments (e.g., the root cause of a low-quality score). In addition, the video dubbing system provides proactive indications to a video player to signal poor-quality dubbing segments before or while they play. Furthermore, the video dubbing system can provide reasoning behind why a particular segment of a streaming video has low-quality dubbing before or when a dubbed audio segment begins playback.

Abstract: The present disclosure addresses energy overheads and latency challenges of architecture design of conventional gated recurrent unit (GRU) approaches by providing a Resistive Random-Access Memory In-Memory Computing-based gated recurrent unit (RRAM IMC-based GRU) network architecture for efficient implementation of GRUs. In the present disclosure, the RRAM IMC-based GRU network architecture is used which performs Multiply-Accumulate (MAC) operations using Ohm's law for multiplication and Kirchhoff's current law for accumulation. A plurality of GRU wight parameters are mapped as device conductance in a Resistive Random-Access Memory (RRAM) memristor array structure in a skewed arrangement. An input vector is applied as voltage pulses to wordlines of the RRAM memristor array structure corresponding to values which should be multiplied and accumulated.

Abstract: This disclosure relates generally to reducing earth-bound image volume with an efficient lossless compression technique. The embodiment thus provides a method and system for reducing earth-bound image volume based on a Spiking Neural Network (SNN) model. Moreover, the embodiments herein further provide a complete lossless compression framework comprises of a SNN-based Density Estimator (DE) followed by a classical Arithmetic Encoder (AE). The SNN model is used to obtain residual errors which are compressed by AE and thereafter transmitted to the receiving station. While reducing the power consumption during transmission by similar percentages, the system also saves in-situ computation power as it uses SNN based DE compared to its Deep Neural Network (DNN) counterpart. The SNN model has a lower memory footprint compared to a corresponding Arithmetic Neural Network (ANN) model and lower latency, which exactly fit the requirement for on-board computation in small satellite.

Abstract: A refrigeration system includes a compressor, a heat rejection heat exchanger comprising a condenser, and a receiver tank in fluid communication with the condenser. A heat exchanger is configured to receive a first refrigerant stream from the compressor discharge and a second refrigerant stream from the receiver tank. The heat exchanger transfers heat between the streams such that at least a portion of the receiver-derived liquid refrigerant vaporizes, producing vapor that is directable back to the receiver tank, while the compressor-derived refrigerant is directable toward the condenser. A flow control arrangement with first and second selectively actuatable valves regulates routing of the compressor discharge. In a first operating condition, the flow control arrangement opens the first valve to direct refrigerant from the compressor toward the condenser and restricts flow toward the heat exchanger via the second valve, enabling conventional condensation and liquid storage in the receiver.

Abstract: A text to speech (TTS) model is trained based on training data including text samples. The text samples are provided to a text embedding model for outputting text embeddings for the text samples. The text embeddings are clustered into several clusters of text embeddings. The several clusters are representative of variations in emotion. The TTS model is then trained based upon the several clusters of text embeddings. Upon being trained, the TTS model is configured to receive text input and output a spoken utterance that corresponds to the text input. The TTS model is configured to output the spoken utterance with emotion. The emotion is based upon the text input and the training of the TTS model.

Abstract: This disclosure relates generally to framework for simulation and cross-platform deployment of a Spiking Neural Network (SNN). models In conventional approach, SNN model created in hardware-agnostic framework suffer performance issues when deployed in a hardware-specific platform as the current standards for interoperability of the models among multiple frameworks maps only interfaces and does not consider difference in behavior due to different order of atomic computations, decay rates of current and voltage, and the reset mechanisms for membrane potential. The present disclosure overcomes this by creating a hardware abstract library having different flavors of middleware synaptic neuron corresponding to different hardware-specific software frameworks. The variation due to decay rates of current and voltage, and the reset mechanisms for membrane potential are determined and atomic instructions are parameterized using the variations.

Abstract: State of the art techniques for reduced data transmission at edge devices while using lossless compression mostly rely on CNNs that are energy consuming networks. However, for edge devices such as LEO satellites the method and system discloses a unique combination of SNN running on a neuromorphic platform for low-power & low-latency. Relevant data of high-resolution images is filtered, and redundant data is discarded. The filtered data is then compressed using but packing based lossless compression before transmitting to a central station for further processing. At the central station a distributed Diagonal Parallel Fast Reconstruction (DDPFR) is applied to reconstruct the image. The system reduces data sent to the central station and reduces battery consumption of satellite for data transmission. SNN running on neuromorphic processor consumes up to 100 to 1000× less power compared to similar Artificial neural network running on GPU.

Abstract: A method of facilitating consumption of online content includes receiving source text for a source article to be translated, the source text being in a source language. The source language for the source text and the target language to which the source text is to be translated are each identified. The source text, the source language, and the target language are each provided to a machine translation model which automatically generates translated text in the target language from the source text. The translated text is provided as input to a generative language model which generates summary text in the target language from the translated text. The summary text is provided to a text-to-speech model which generates summary audio from the summary text. The summary text and summary audio are then sent to a user interface via which the summary text is displayed, and playback of the summary audio is enabled.

Abstract: This disclosure describes a framework for generating real-time audio translations of videos on a client device. Specifically, this disclosure describes a video dubbing system that utilizes a concurrent batch-processing architecture to provide real-time audio translations of videos on a client device. Additionally, in one or more implementations, the video dubbing system utilizes time-aware segmentation to prevent audio misalignment of the translated audio. As described below, the video dubbing system efficiently provides high-quality audio translations of videos that accurately align with the video content for the entire video, regardless of the video's length.

Abstract: State of art techniques such as ANN2SNN yield sub-optimal performance for Spiking Neural Networks (SNNs), while with SNN search space Neural Architecture Search (NAS) based approach many times the target SNN hardware constraints might not be met or there can be accuracy loss. A method and system for rapid automated generation of optimized SNNs for neuromorphic devices is disclosed. A Reinforcement Learning (RL)-NAS technique is utilized to obtain Deep Neural Networks (DNNs) using raw training dataset in non-spike data format and set of SNN constraints in accordance with the neuromorphic hardware of target neuromorphic devices along with general target hardware constraints, wherein the set of constraints are imbibed into the NAS space in form of constrained graph. An optimal DNN is model searched in an SNN contained NAS space and trained on the raw data. An optimal SNN model is obtained from the trained DNN model via neuromorphic OEM converter.

Abstract: A refrigeration system comprises a compressor, a condenser, a receiver tank, an evaporator and a heat exchanger. The heat exchanger comprises an inlet positioned downstream of the receiver tank and an inlet positioned downstream of the compressor. The heat exchanger is configured to transfer heat between refrigerant received from the compressor and refrigerant received from the receiver tank, wherein a transfer of heat causes at least a portion of the refrigerant received from the receiver tank to transition from liquid to vapor. This process propels the head pressure of the compressor to increase to compensate for low ambient conditions. The heat exchanger comprises a first outlet in fluid communication with the first inlet, the first outlet configured to dispense the vapor refrigerant toward the receiver tank, and a second outlet in fluid communication with the second inlet and configured to dispense the refrigerant received from the compressor toward the condenser.

Abstract: The disclosure generally relates to an FPGA-based online 3D bin packing. Online 3D bin packing is the process of packing boxes into larger bins-Long Distance Containers (LDCs) such that the space inside each LDC is used to the maximum extent. The use of deep reinforcement learning (Deep RL) for this process is effective and popular. However, since the existing processor-based implementations are limited by Von-Neumann architecture and take a long time to evaluate each alignment for a box, only a few potential alignments are considered, resulting in sub-optimal packing efficiency. This disclosure describes an architecture for bin packing which leverages pipelining and parallel processing on FPGA for faster and exhaustive evaluation of all alignments for each box resulting in increased efficiency. In addition, a suitable generic purpose processor is employed to train the neural network within the algorithm to make the disclosed techniques computationally light, faster and efficient.

Abstract: The present invention generally relates to the field of deep learning, and, more particularly, to a method and system for task agnostic distillation in foundation models. Conventional distillation methods are not scalable and also does not handle mismatches between embedding sizes of teacher and student models. Thus, embodiments of present disclosure first transforms the teacher model in such a way that its embedding size matches that of the student model. This is done by augmenting a linear layer having dimensions equal to the embedding size of the student model and a projector network to the teacher model. Then the augmented layers are trained using a self-supervised learning technique by freezing the teacher model. Finally, the projector network is discarded to obtain a transformed teacher model. The student model is then trained using transformed teacher model by performing knowledge distillation based on similarity loss.

Abstract: A refrigeration system comprises a compressor, a condenser, a receiver tank, an evaporator and a heat exchanger. The heat exchanger comprises an inlet positioned downstream of the receiver tank and an inlet positioned downstream of the compressor. The heat exchanger is configured to transfer heat between refrigerant received from the compressor and refrigerant received from the receiver tank, wherein a transfer of heat causes at least a portion of the refrigerant received from the receiver tank to transition from liquid to vapor. This process propels the head pressure of the compressor to increase to compensate for low ambient conditions. The heat exchanger comprises a first outlet in fluid communication with the first inlet, the first outlet configured to dispense the vapor refrigerant toward the receiver tank, and a second outlet in fluid communication with the second inlet and configured to dispense the refrigerant received from the compressor toward the condenser.

Abstract: This disclosure relates generally to time series forecasting, and, more particularly, to a system and method for online time series forecasting using spiking reservoir. Existing systems do not cater for efficient online time-series analysis and forecasting due to their memory and computation power requirements. System and method of the present disclosure convert a time series value F(t) at time ‘t’ to an encoded multivariate spike train and extracts temporal features from the encoded multivariate spike train by the excitatory neurons of a reservoir, predict a time series value Y(t+k) at time ‘t’ by performing a linear combination of extracted temporal features with read-out weights, compute an error for predicted time series value Y(t+k) with input time series value F(t+k), employs a FORCE learning on read-out weights using the error to reduce error in future forecasting. Feeding a feedback value back to the reservoir to optimize memory of the reservoir.

Abstract: Small and compact Deep Learning models are required for embedded AI in several domains. In many industrial use-cases, there are requirements to transform already trained models to ensemble embedded systems or re-train those for a given deployment scenario, with limited data for transfer learning. Moreover, the hardware platforms used in embedded application include FPGAs, AI hardware accelerators, System-on-Chips and on-premises computing elements (Fog/Network Edge). These are interconnected through heterogenous bus/network with different capacities. Method of the present disclosure finds how to automatically partition a given DNN into ensemble devices, considering the effect of accuracy—latency power—tradeoff, due to intermediate compression and effect of quantization due to conversion to AI accelerator SDKs.

Abstract: A technique translates speech in a first language to text in a second language manner in a manner that is appropriate for the gender of the speaker. In some implementations, the technique receives an input setting that specifies one of three translation modes: masculine mode, feminine mode, and auto mode. The first two modes produce translations in masculine and feminine modes by default, respectively, while the auto mode produces translations in forms that are based on the detected characteristics of audio signals. According to some implementations, the technique uses a training framework that automatically converts a corpus of training examples that exhibit gender bias (e.g., a male gender bias) to training examples having a reducing incidence of gender bias. In some implementations, the training framework updates weights of the machine-trained model based on a combination of two loss components: translation loss and gender loss.

Abstract: A method for operating a refrigeration system includes cooling a refrigerant using a gas cooler, decreasing the pressure of the refrigerant from the gas cooler using an expansion valve, and separating the refrigerant provided by the expansion valve into a vapor refrigerant and a liquid refrigerant using a flash tank. The method proceeds by determining whether to cool a medium temperature (MT) space or a low temperature (LT) space, and determining if heat exchange should occur for the refrigerant compressed by the MT compressor unit and the LT compressor unit, among other operations.

Abstract: A refrigeration system includes a low temperature (LT) evaporator unit, a medium temperature (MT) evaporator unit, or both. The system further includes a LT compressor unit, an MT compressor unit, or both. The system further includes a heat exchanger, a gas cooler, an expansion valve, and a flash tank.

Abstract: This disclosure relates generally to action recognition and more particularly to system and method for real-time radar-based action recognition. The classical machine learning techniques used for learning and inferring human actions from radar images are compute intensive, and require volumes of training data, making them unsuitable for deployment on network edge. The disclosed system utilizes neuromorphic computing and Spiking Neural Networks (SNN) to learn human actions from radar data captured by radar sensor(s). In an embodiment, the disclosed system includes a SNN model having a data pre-processing layer, Convolutional SNN layers and a Classifier layer. The preprocessing layer receives radar data including doppler frequencies reflected from the target and determines a binarized matrix. The CSNN layers extracts features (spatial and temporal) associated with the target's actions based on the binarized matrix.