Oral Sessions

Brain-inspired Hyperdimensional (HD) computing is an emerging technique for cognitive tasks in the field of low-power design. As an energy-efficient and fast learning computational paradigm, HD computing has shown great success in many real-world applications. However, an HD model incrementally trained on multiple tasks suffers from the negative impacts of catastrophic forgetting. The model forgets the knowledge learned from previous tasks and only focuses on the current one. To the best of our knowledge, no study has been conducted to investigate the feasibility of applying multi-task learning to HD computing. In this paper, we propose Task-Projected Hyperdimensional Computing (TP-HDC) to make the HD model simultaneously support multiple tasks by exploiting the redundant dimensionality in the hyperspace. To mitigate the interferences between different tasks, we project each task into a separate subspace for learning. Compared with the baseline method, our approach efficiently utilizes the unused capacity in the hyperspace and shows a 12.8% improvement in averaged accuracy with negligible memory overhead.

 

Multimedia displays are generally used nowadays. With different display panels and wide range of input source, image scaling techniques are indispensable in order to convert image sources from input resolution to output resolution. Super-resolution (SR) was used to improve the scaling quality through different computations. However, in some special cases, such as overexposed or underexposed images, the image details still can’t be recovered well through complex SR scaling. In this paper, we discuss the extremely exposed image processing and propose a lightweight deep learning architecture to enhance the extremely exposed image scaling quality.

 

Object detection obtains huge improvement after adopting deep learning technique. However, deep learning technique requires extremely high computation complexity and heavy DRAM (Dynamic Random Access Memory) bandwidth requirements, which blocks the deployment over various kinds of platforms. This paper provides a distance-aware technique for object detection that can adjust the required computation complexity and DRAM access amount according to several different searching distances. According to our analysis, the perception system can save up to 34.2% and 20.1% of computation complexity and DRAM access amount, respectively, for detecting near field objects when compared to the full range detection. The proposed distance-aware technique can let the deep learning neural networks adjust the required computation complexity and DRAM access amount according to several different searching distances. Unlike other works that emphasized on removing dummy computations and DRAM access from the network by trading-off limited detection accuracy, we propose a human natural way for machines to follow. By following this way, the machines search the surrounded environment in a priority scheme, from near to far, just like humans.

 

This work presents an intelligent system platform design for hybrid mixer and digital equalizer using speech recognition technology instead of manual turning the knobs. The proposed platform provides both touch control and voice control for adjusting sound sources, volume, and the effects of equalization and reverberation using a smartphone APP. For speech recognition, a Google cloud service releases a strong function library, i.e. “SpeechRecongnizer class”, which helps convert short-time speech into a string. Then, a useful control message extracted from a string segmentation and comparison processing is transmitted to a data transfer controller (DTC) by a Bluetooth module. To connect the smartphone and the proposed DTC, we especially define a self-defined Bluetooth Packet Format (SBPF) into the Bluetooth5.0 protocol. The proposed SBPF also adopts Cyclic Redundancy Check (CRC, CRC-16-CCITT), and acknowledgment mechanism to improve the transmission correctness and data integrity. For the proposed hybrid mixer and digital equalizer, a Digital Signal Processor (DSP KT0707) is employed to be the kernel of this intelligent system platform and is connected to the proposed DTC by I2C protocol. Compared to traditional audio mixer designs, a man who is responsible for public address system (PA) can more quickly predominate the effects of the ambient sound field by using voice/touch control via the proposed APP, and enhances listeners’ feelings about live band performance.

 

* Abstract is not available.

 

* Abstract is not available.