80,000 Hours Podcast cover image

#107 – Chris Olah on what the hell is going on inside neural networks

80,000 Hours Podcast

NOTE

On the Issue of Police Multisemanticity and Neural Networks

Having diverse agendas and a portfolio of different concepts can mitigate worries and increase the chances of success. However, the issue of police menticity, where a neuron responds to multiple things, remains unclear. It is challenging to understand how this phenomenon occurs, especially when circuits seem to perform unrelated functions. This trade-off persists even in larger networks, as the number of neurons and concepts increases. It is unknown if polysomanticity eventually dissipates in very large models or if alternative approaches are necessary. The complexity of polysomanticity hinders full comprehension, as other hidden elements may exist within neurons. The distinction between features and neurons is important, as features can represent combinations or correspond to multiple neurons. Disentangling representations is a related problem in machine learning, but with different challenges and objectives.

00:00
Transcript
Play full episode

Remember Everything You Learn from Podcasts

Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
App store bannerPlay store banner