3 ms·

Kind-of a bummer article for what I think is a very good point (that I'm not actually sure he is trying to make). Current models based on biological principals

Kind-of a bummer article for what I think is a very good point (that I'm not actually sure he is trying to make). Current models based on biological principals are starting at a level of complexity which is emergent from a more foundational level which is ignored in our currently best performing models. There is a strong argument for understanding 'mind' or 'intelligence' at a cellular level. Neural networks were originally inspired by how we thought brains work; a highly connected network. Within this framework, it was kind-of ignored that the fundamental units of this network themselves have 'mind' (intelligence? sense of self?) and at least the ability to change their behavior (outputs) based on what they sense from their environment (inputs). Currently, we train NN based on training data, but once trained, the network is 'fixed', in that we've come up with some optimal set of weights and biases that don't change. Biological networks don't work like this. Individual units within biological networks can individually adjust their behavior, based on external inputs both from other cells in the network and external to whole network (environmental) inputs. I would argue that there has been a strong bias in cell biology (especially human cell biology) against attributing individual cells with 'mind' or a sense of self; however, this obviously can not be the case, since most of the diversity of life is unicellular. I think there would be a strong benefit to considering the ideas of 'cells' and compartmentalization in regards to how we design and engineer intelligent systems: we know that fundamentally, this is at least part of how cells work.

AllegedAlec 8y ago
> Current models based on biological principals are starting at a level of complexity which is emergent from a more foundational level which is ignored in our currently best performing models. Not really. It's too complex, we understand too little, and there's very little reason for it, since intelligence is not a property which is encoded in a neuron, but an emergent property, emerging from the interactions between them and the development of the network. It's a bad idea for the same reason that if you want to simulate the aerodynamic properties of an airplane, you don't start with string theory. > Individual units within biological networks can individually adjust their behavior, based on external inputs both from other cells in the network and external to whole network (environmental) inputs. You don't need to start from the cellular level to achieve this though.

AllegedAlec 8y ago

> Current models based on biological principals are starting at a level of complexity which is emergent from a more foundational level which is ignored in our currently best performing models. Not really. It's too complex, we understand too little, and there's very little reason for it, since intelligence is not a property which is encoded in a neuron, but an emergent property, emerging from the interactions between them and the development of the network. It's a bad idea for the same reason that if you want to simulate the aerodynamic properties of an airplane, you don't start with string theory. > Individual units within biological networks can individually adjust their behavior, based on external inputs both from other cells in the network and external to whole network (environmental) inputs. You don't need to start from the cellular level to achieve this though.