the last neural cell

LLM_interpretability_AI _summary.pdf

71.2 KB

Over the past time I collected some interesting mechanistic interpretability papers - trying to understand the information processing in AI systems.

Since I'm not an expert, I delegated some of the work to AI itself 🙃

Why I find it interesting? Well, the field of "AI neuroscience" is emerging. This, firstly, brings two fields closer together (neuroAI doesn't count). Second, I believe both understanding of AI and the brain can be advanced by testing theories on the systems where everything can be measured and controlled.

Of course, LLMs are far from natural biological nervous systems, but I believe there are fundamental information processing principles that are shared.

💡Some insights:
> Activations in RNNs mirror the statistics of their input data
> Apparently chunking of computations is present akin to brain's hierarchical processing
> "Grandmother cell" are real but sparse, distributed processing is more likely to be happening

Have a look, it's fun 🙂

www.group-telegram.com/us/neural_cell.com/253

633 viewsAleksejs Timčenko, edited  Feb 18 at 11:17

Over the past time I collected some interesting mechanistic interpretability papers - trying to understand the information processing in AI systems.

Since I'm not an expert, I delegated some of the work to AI itself 🙃

Why I find it interesting? Well, the field of "AI neuroscience" is emerging. This, firstly, brings two fields closer together (neuroAI doesn't count). Second, I believe both understanding of AI and the brain can be advanced by testing theories on the systems where everything can be measured and controlled.

Of course, LLMs are far from natural biological nervous systems, but I believe there are fundamental information processing principles that are shared.

💡Some insights:
> Activations in RNNs mirror the statistics of their input data
> Apparently chunking of computations is present akin to brain's hierarchical processing
> "Grandmother cell" are real but sparse, distributed processing is more likely to be happening

Have a look, it's fun 🙂

BY the last neural cell