 Veliko v Las Vegas in v artificial inteligenciju labu v CES. Mi je Danilo Pau, imam ingenir in v avancenju vsečenju labu Stimikroelectronikov in atripoli. Imam tudi tudi, da imamo všečenje na AI. Vseč, kako se vsečenje, arhitekči in neuraletorki, ne musili dobročati dvodnih tukov, tkje z Keras, Stavna, Kaffe in sovna. Prič tudi, vse vse je vzumil povrti resursi, povrti GPUs, klas resursi in začilo komputativne pačje in memoji. Zelo, da bi nekaj všeč resursi in Tukaj slad conductiv have limited computing capabilities and memory. To kot, da je a gap. Now the tool that we just released, since January the 3rd called STM32 cube dot AI, closes this gap by allowing the programmers to quickly generate automatically code from pre-training neural networks to be executed efficiently v termu memori in komputationalne kapabilitje na sdm32 mikrocontrolleru, basite na Arm, Cortex, M4 in M7 istračnji set. Včešč je izgleda izgleda tukaj, kaj sem izstancijila multipolje netvore, tudi na klasifikacijstvu, na rekonicijstvu in tukaj, To je poslednje, da je poslednje pretenjera in neuronetor, kako je poslednja in karast in tensovla spoder in se površal se v sekund, vse površal se komplexitivne komputationalne kapabilitje in memori. In poveč, karast, v njeg tudi v X86 in voka v takratiku mikrokontrollera, da se je nezavrša in akuracija versus the one that the customer achieved during the training phases. Thanks for your attention and now I leave the word to my colleague Viviana. Hello, my name is Viviana Dalto. I am responsible of the AI applications and tools activity in system research and application in ST. I am going to introduce you the ready to use software function pack AI sensing one. As Danilo was introducing before, STM32QBI is providing in an easy, quick way a neural network library that can be used in an STM32 project and run on the STM32 platform of choice. The software pack AI sensing one is going to provide you examples of neural network running on inertial and audio signals and running on our reference platform sensor tile. Sensor tile is providing an STM32L4, 880 megahertz, inertial sensors, one microphone and environmental sensors. And we are providing various examples of human activity recognition neural network with different topologies that are running in firmware on this platform. The results of the classification can be shown using a mobile application for smartphone that can be downloaded from Google Play or iTunes. For instance, we have in this case an application of human activity recognition where we can discriminate among five different activities, stationery, walking, running, cycling and driving. In particular, I can pretend to run and the result is displayed here. The second example of neural network is for audio sync classification. It is possible with this neural network to discriminate among three different classes indoor, like home and office, outdoor, city center and parks and in vehicle, in the car, train and bus. In particular, I can show here that with an infidelity loudspeaker and a YouTube video that a train contest is easily recognized by the application. Also, the application is able to provide the possibility to log data with ground truth in the SD card that is present in the sensor tile. We can select the sensors of choice, we can select also the sampling rate and also we can create the classes we want to identify so that as soon as we start logging the data and streaming them on the SD card, we are able also to label them. We have seen what is available already on STM32 QBI. This is the next step. This is a prototype of what will be available in the next generation. As you know, image classification using neural network is a computational expensive task. We need to enable the possibility in order to run this in software on STM32 microcontroller unit to perform neural networks using fixed point. In this case, we have an example of neural network classifying among 18 different classes of food. There is a camera here, which is streaming data in VGA to the STM32 H7 at 400 MHz and the image is scaled to 224 by 244 and then the microcontroller is performing completely in software on a neural network using 8-bit fixed point neural network layers library that will be available soon by STM32 QBI. In particular, this shows that we are able to perform one classification every 150 millisecond purely in software. Here we are going to show an example of a neural network running on STM32 L5, which is extending the ultra low power STM32 family with Cortex M33 core at 110 MHz. Here we are showing a character recognition neural network that is able to grab characters that are handwritten on the screen like call in order to enable a specific action. So now I am leaving the world to my colleague Mathieu. Hello, my name is Mathieu Dronohan. I am the head of AI application within the microcontroller division. So let me introduce you a product developed by our partner Sienna System. It's a sport and activity tracker, and it's allowing not only to do some advanced human activity recognition, but also to support selectable sport centering model like basketball or tennis. The device is based on the STM32 L4 microcontroller and is embedding also accelerometer, gyroscope and pressure sensor from ST. Sienna System use the sensor type to collect data, to prototype and test, and they also use our STM32 cube.ai tool to port their neural network inside the STM32 with a huge gain in term of productivity and efficiency. They are keeping on adding some new sport to the device and new feature. Sienna System is a part of our ST AI partner program. I invite you to visit our website, ST.com slash STM32 cube AI. Thank you.