One of the major challenges in applying LLM (Generative AI) to mission-critical applications is hallucination. We are researching the fundamental theory of applying Shannon Information Theory in the vector embedding space to eliminate an obvious ‘error’.”
A machine learning-based autonomous driving system is fundamentally challenging to assess in terms of safety metrics. Even with a model test result showing 99.9999% (6-sigma) sensitivity and specificity over a finite test dataset, it cannot guarantee that the system will handle untrained situations (corner cases) with the same level of safety as tested cases. The infinite number of corner cases (the long-tail problem) cannot be solved through iterative training alone. We envision that the next generation of autonomous driving systems will rely on an AI Error Correction model. This model quantifies the error rate by measuring the distance between drivable events and fatal events encoded in the vector embedding space.”
We believe that the future of AI lies at the edge, driven by reasons of privacy, safety, responsiveness, and cost. No one wants to send sensitive data to the cloud. For mission-critical applications, it is better not to rely solely on network connectivity. Cloud latency, especially for mobile voice chatbots, might be too long and challenging to improve. The cost of running AI applications to serve billions of prompts may simply prove infeasible. Therefore, there is a need to run a decent-sized language model (with over 70 billion parameters) on a smartphone in the near future. The solution could be an AI ASIC integrated into a System-in-Package (SiP) that offers 10X higher computation density (TOPs/mm²) compared to current NPUs/GPUs, while consuming 10X less power and being 10X cheaper.
The recent advancement in 4D radar combines traditional radar technology with additional capabilities, such as higher resolutions, elevation, and velocity measurements. 4D radar excels in long-range detection, making it effective for applications that require early detection of distant objects. Moreover, radar remains unaffected by adverse weather conditions or lighting variations, ensuring reliable performance in challenging environments. By combining stereo cameras with 4D radar, this system addresses an unmet need for a cost-effective, long-range, high-resolution, evidence-based detection and ranging solution. It excels in challenging lighting and adverse weather conditions, leveraging state-of-the-art artificial intelligence technologies to support advanced object detection, classification, and scene segmentation.
We offer consultation services to help your company transform into an AI-infused organization, boosting productivity and competitiveness.
3458 visits