The Würzburg University satellite SONATE-2 was launched into orbit a year ago and has achieved all its mission objectives. The satellite, developed by Julius-Maximilians-Universität Würzburg, successfully demonstrated the training of AI autonomously on board for anomaly detection on Earth's surface. The satellite's neural networks were trained to recognize objects not typically found in the Sahara, such as the Nile and adjacent green regions. In addition to the AI payload, the satellite also tested other technologies like the MultiView star sensor and amateur radio services. The satellite remains fully functional and is used for educational purposes.
Recent #Neural Network news in the semiconductor industry
➀ The IEEE’s International Solid-State Circuit Conference in San Francisco featured an overview of next-generation high-performance CPUs.
➁ AMD revealed aspects of Zen 5, its x86-64 microprocessor for desktops and laptops.
➂ IBM described its 5.5GHz Telum II processor with enhanced AI acceleration and new data processing units.
➀ Raspberry Pi has released a 12.3 Megapixel AI camera based on Sony's IMX500 intelligent sensor; ➁ The camera supports on-board processing for various neural network models, freeing up the Raspberry Pi's processor for other tasks; ➂ The camera module measures 25 x 24 x 11.9mm and costs around £55.
➀ TDK has developed a spin-memristor for analog neural networks and built a network that can separate music, speech, and noise in real-time. The technology is based on a collaboration with CEA since 2020 and uses a 300mm-compatible spintronics semiconductor prototyping line at Tohoku University. ➁ The memristors are controllable, have low current leakage, and are robust against resistance drift, allowing for long-term data storage. ➂ TDK and CEA have implemented an analog algorithm that can separate the three types of sound, and the technology has a significantly lower power consumption compared to conventional digital neural network implementations.
➀ Huawei ADS 3.0 demonstrates impressive end-to-end intelligent driving capabilities, including autonomous parking and navigation from one parking spot to another. ➁ The system utilizes a biologically inspired neural network model to enhance its ability to understand and navigate complex road scenarios. ➂ The technology aims to simplify the driving experience by reducing the need for human intervention, making it more intuitive and efficient.
❶ A neural network has been trained to create spatial maps and predict subsequent video frames using a predictive coding algorithm and Minecraft gameplay, achieving a mean-squared error of 0.094%. ❷ The project, led by researchers at Caltech, demonstrates AI's ability to navigate conceptual spaces, moving beyond simple input-response memorization. ❸ The research, published in Nature Machine Intelligence, includes detailed methodology and the code is available on GitHub and Zenodo for public use.