Brain Waves and Bots: The Unlikely Convergence

Large language models, or LLMs, have been making waves lately. They seem to be aligning more closely with human cognition. At least, that's what some researchers are claiming as they draw parallels between these AI models and human brain activity.

LLMs and Emotional Valence

In a quirky experiment, scientists constructed a one-dimensional valence direction using sentences that evoke emotion. They called this the V-axis. It’s like a mood ring for text. With just nine sentences, these researchers were able to validate this axis by applying it to sentiment benchmarks and comparing its outcomes across fourteen different LLMs. Then, they took it a step further. They found that this same direction mapped onto human neural activity in 123 subjects watching emotionally charged videos. Sounds impressive, right?

Sure, but here's the kicker. Despite the apparent alignment between AI and human brainwaves, this convergence didn't improve the effectiveness of training signals. Out of twenty-five attempted strategies to enhance decoding accuracy, none worked. In fact, sixteen attempts reduced accuracy. The data's already whispering failure.

The Saturation Regularity

This persistent failure led to the formalization of what researchers call the saturation regularity. Once a brain-decoding network hits its target, extra supervision only distorts it. It's like trying to fix a car by painting over the dents. The real improvements lie in the residual subspace that supervision can’t reach. So, instead of doubling down on a distorted model, they opted for diversity. By ensembling across different residuals, they upped balanced accuracy by 10.5% on the FACED dataset, with similar results on SEED-V. A small win, but a win nonetheless.

What's Next?

What does this all mean? Should we care about this alignment between AI and brainwaves that seems to go nowhere? It's a bit like discovering a new planet only to find out it's uninhabitable. These breakthroughs in alignment sound promising, but unless they translate into practical improvements, they’re just noise. Everyone has a plan until liquidation hits, or in this case, until accuracy doesn't budge.

So, the next time someone hypes up the potential of AI to decode human emotions, ask them where the real improvements are. Will ensembling across residual diversity be the answer? Or is this another case of being bullish on hopium and bearish on math?