Brainchip Klassengruppe WKN: A14Z7W ISIN: AU000000BRN8 Kürzel: BRN Forum: Aktien User: Rarosch

@7FÜR7. Mit ruhe vor dem Sturm meinst du, dass eine Ankündigung bevor steht? (Ob positiv oder negativ unbekannt)

@7FÜR7. Mit ruhe vor dem Sturm meinst du, dass eine Ankündigung bevor steht? (Ob positiv oder negativ unbekannt)

Und eine neue Demo von ihm https://www.linkedin.com/posts/kevin-d-johnson-42170b_i-built-a-system-that-watches-10-live-railfan-activity-7448134717607936000-tq9_ Kevin D. Johnson 3. Field CTO – HPC AI | Principal HPC Cloud Technical Specialist at IBM | Symphony • GPFS • LSF 36 Min. Bearbeitet I built a system that watches 10 live Railfan cameras across the United States, classifies every freight car in real time on neuromorphic hardware, and turns the results into economic intelligence. SymRail is the newest domain for SymWisdom, the Neuromorphic Hive Mind I have been building with a cluster of BrainChip AKD1000s. The same 10 chips that run defense ISR classifiers and whiskey barrel aging models now run a 9-class railcar classifier on video of major junctions from Rochelle, Illinois to Folkston, Georgia and beyond. Here is how it works. Livestreams from Railfan cameras feed into a frame extractor. Frames are forwarded to each of the 10 AKD1000 chips with their assigned camera location. The chips classify what they see: tank cars, intermodal containers, grain hoppers, coal hoppers, autoracks, boxcars, empty flats, locomotives, or no train at all. Spike records flow into shared memory then into GPFS archives. You can watch the Flask dashboard on the video below with a live US map. The rail car classifier started from an open dataset of 1,454 annotated rail car images. I trained an AkidaNet edge model using transfer learning from an ImageNet backbone, replacing the classification head with a 9-class output. Version 1 hit 85.6% accuracy on actual AKD1000 hardware. But I wanted better, so I fine-tuned a YOLOv8n detector on the same dataset, used it to auto-label 4,247 additional crops from real video footage, merged everything into an 8,295-image training set, and retrained. Version 2 reached 91% accuracy on the chips, quantized to 4-bit weights and activations, fitting in just 1 MB of SRAM. The point is not mere trainspotting. Tank car counts at strategic junctions are a leading indicator for petroleum logistics. When crude-by-rail volumes shift before EIA publishes weekly inventory data, that is signal. SymRail generates comparable freight visibility from publicly accessible cameras using neuromorphic chips for a lot less. There is actually Citizen Science academic work in this area. Manual rail monitors in Snohomish County organized 29 volunteers to count oil trains by hand in 2014. Oak Ridge National Lab researchers reconstructed crude-by-rail routes from geotagged Flickr photos and achieved 96% coverage of documented incident locations. SymRail automates what people have been doing manually for a decade, and it runs 24/7 without fatigue. Here's the economic question, though: do visual tank car counts at these junctions predict crude oil price movements before the official numbers come out? The answer is likely framed in line with the fact that satellite data on similar information is typically very expensive (>$50k/yr). Everything runs on the same IBM Spectrum Symphony cluster, the same GPFS filesystem, the same Palantir Foundry crystallization pipeline. No new infrastructure. Just a new model and a new way for the Neuromorphic Hive Mind to experience the world. All aboard!!! Where's my Crazy Train video?