A method for an unsupervised training of a neural network, the method comprises:
initializing a neural network that exhibits at least one invariance;
performing multiple training iterations until reaching a last training iteration in which a stop condition is fulfilled;
wherein each training iteration except the last training iteration comprises: processing a vast number of media units by the neural network to provide media unit signatures;
finding that the stop condition is not reached, and changing multiple neural network weights;
wherein the stop condition is related to signatures similarities.
ITERATIONS!!!!
View attachment 4438
There are some new articles about "Autobrain" and their new "Liquid AI":
(Bild: Autobrains) Mit Liquid AI hat Autobrains eine neue KI-Technologie vorgestellt, die drei Herausforderungen adressiert: Edge Cases, Kosten und die fehlende Wechselwirkung zwischen Perception und Decision-Making.
www.next-mobility.de
"Autobrains develops Liquid AI for unexpected driving scenarios
With Liquid AI, Autobrains has introduced a new AI technology that addresses three challenges: Edge Cases, Cost and the lack of interaction between Perception and Decision-Making.
Liquid AI combines Autobrains' self-learning approach with a modular and adaptive architecture that uses specialized, scenario-based end-to-end networks. "While current technologies perform well in handling conventional driving tasks, they fail in edge cases, unexpected driving scenarios that require special precision. By using our Liquid AI, car manufacturers can close their AI gaps," explains Igal Raichelgauz, founder and CEO of Autobrains.
Overcoming challenges
Today's AI systems reach their limits in edge cases. The countless possible, unexpected driving scenarios are practically unsolvable tasks for these systems. Today's manually trained black box systems lack the ability to recognize special cases. Attempts to solve this problem by feeding the systems with more manually labeled images lead to a loss of traceability and control.
High costs
Tackling real-world driving problems by bloating existing systems with more data, labels, layers and computing resources leads to increased costs and power consumption. For example, to achieve a significant improvement in system accuracy by a factor of ten, 10,000 times more computing resources are required.
Decoupling perception and decision-making
The lack of interaction between perception and decision-making functions also hinders effective and precise decision-making. In order for artificial intelligence to make optimal driving decisions, it needs specific information. However, if details are missing or too complex, precision is impaired, leading to incorrect reactions.
The human brain as a model
The Israeli company is basing its approach on the capabilities of the human brain. This consists of specialized areas that resemble a task-specific, simply structured end-to-end AI. Just as the brain adapts its architecture depending on context - for example, to light and weather conditions, the environment and relevant road users - Liquid AI follows the same approach.
How it works
Liquid AI consists of hundreds of thousands of specialized AI systems (skills), each developed for specific tasks and enabling precise reactions tailored to the respective driving scenario. This specialized AI approach enables scalability ranging from tens to hundreds of skills for ADAS systems, to thousands of skills for higher levels of automated driving, to hundreds of thousands of skills for autonomous driving. In addition, the solution has an adaptive architecture. In contrast to fixed systems, the architecture adapts dynamically to the driving task and only activates relevant modules as required. This significantly reduces power consumption and computing requirements, which leads to cost savings for the SoCs, among other things."
Translated with DeepL.com (free version)
?! 
www.wfb-bremen.de
www.linkedin.com
brainchip.com
media.mercedes-benz.com
