9621 Agnes Crossing, Lake Suzanneview, New Mexico Island 84604-9295.
American AI company Anthropic has announced that it has made progress in understanding artificial neural networks. AI startup Anthropic has made a breakthrough in its understanding of how
American AI company Anthropic has announced that it has made progress in understanding artificial neural networks.
AI startup Anthropic has made a breakthrough in its understanding of how artificial neural networks work. The company announced this with an extensive blog.
Like most AI models, neural networks are trained with data. However, because they are not programmed to follow any rules, such models can behave in many ways. Anthropic understands the calculations behind training the networks, but not how they influence behavior. This limits diagnostics, repairs and safety.
Experiments can record the activation of each neuron in an artificial network. This allows researchers to stimulate or stop it and test the network. Unfortunately, neurons do not have a consistent influence on the behavior of networks. The activation of a neuron can mean something completely different depending on the context.
Anthroponically built machines to find entities (“features”) associated with patterns of neuron activation. This allows researchers to break down neural networks into more understandable parts. They also build on previously acquired information.
A layer of 512 neurons is divided into four thousand units with different meanings. These are usually invisible when neurons are activated individually. This happened in one Transform language model.
It turned out that the units were much more interpretable than individual neurons.
An LLM then created descriptions of the units. This was compared to the ability of another model to predict activation of a unit with this description. In this case, too, individual neurons performed significantly worse. Artificially activating a unit changed a model in a predictable way.
A number of learning units often proved to be universal between different models, so researchers now dare to generalize them.
After Google’s expected billion-dollar investment, the results for Anthropic are even more positive news. This is after Amazon already did the same thing once.
Source: IT Daily
As an experienced journalist and author, Mary has been reporting on the latest news and trends for over 5 years. With a passion for uncovering the stories behind the headlines, Mary has earned a reputation as a trusted voice in the world of journalism. Her writing style is insightful, engaging and thought-provoking, as she takes a deep dive into the most pressing issues of our time.
