𝙩𝙮≃𝙛{𝕩}^A𝕀²·ℙarad𝕚g𝕞

统计数据

1161
文章
0
粉丝
0
获赞
17481
阅读
Max Bennet,智能简史的作者,在MLST的访谈中,提到的Tristan关于灵长类新皮质层layer4的解释和他自己的speculation,非常有启发。layer4接收从丘脑转发的感官信息,然后自己通过符号建模做行为预测,而哺乳动物从结构上也有layer4,但后来萎缩了。 让我接着speculate: 这也解释了具身与符号的强耦合,我们天生具备符号接地能力的,但显然我们的环境越来越被数字化后,我们的建模和行为预测也受到很大影响,未来如何真不好说,而且现在LLM又来了,会直接影响到我们的认知能力。 let’s about granular versus agranular because I think the best theory I’ve seen is Tristan’s theory on this. So what does layer 4 do in neocortex? Across the entire neocortex, layer 4 is where sensory input is received. The primary sensory input is received into the neocortical column. This comes from the thalamus so the canonical model is sensory input from sensors, eyes, ears, skin flows up through the brainstem to the thalamus and then from the thalamus propagates to layer 4 and then from layer 4 it goes within a variety of other layers of neocortex. And then other layers of neo- cortex project back to the thalamus and the rest of the brain. So why would it be the case that regions of neocortex would not have a layer 4? Well, if you actually watch an animal’s development, what’s interesting is mammals with an agranular prefrontal cortex is not always agranular. It actually starts having a layer 4 and the layer 4 atrophies over development. And so I think this is very, mirrors well for instance, of active inference where what’s happening is the neocortical column can kind of be in 2 phases. It can either be trying to match its model of the world to its sensory input. In other words, I see sensory input and I’m trying to infer what’s there and I’m gonna construct the idea of a triangle. But there’s another state of a neocortical column which is generation which is I’m gonna start the latent representation of a triangle and I’m gonna imagine and explore it. And so 1 idea is that what frontal cortex does is primarily try and fit the world to its model. In other words, it spends the vast majority of its time constructing intents and not trying to modify that intents to fit what it observes, but in fact try to change what an animal does to satify its intent. So layer 4 atrophies doesn’t actually go all the way away. If you go deep into a brain you see some basic layer 4 so it’s not completely gone but it atrophies because frontal cortex spends very little time trying to change what it perceives its intent to be to map what the animal’s doing, but in fact what it does is tries to change what an animal does to map it match to its intent. And what I think is so interesting and brilliant about this idea is it explains exactly why layer 4 doesn’t start not existing. Because at first an animal needs to build a model of itself, thus layer 4 is present, but over time it shifts towards once I have a model of what I want and who I am and the things I would do, I don’t need to spend as much time changing my model of self. I’m gonna spend most of my time trying to change my behavior. So this is a very speculative idea but it makes a lot of sense in the context of active inference and it’s the best to personally and all of my reading of, explanations of why, agranularity exists, the best explanation I’ve seen.