I could imagine two ways how to create universal mind using NNs.
One is to use NNs at low level, like image recognition, and something like rule based inference engine on the upper level, which will used data from NNs for decision making. I think that NN using self-driving cars are built this way.
Another is to create large scale functional model of a human brain, where each block will consist of NN, but the blocks will be rather independent of each other and only exchange information. There are at least 50 Brodmann areas in the human brain with different anatomy.
I am sure we will see an attempt to create universal robotic brains using similar approaches.
Added: The third way is to create very large NN able to predict the whole behaviour of a person based on the whole recording of his previous activities. It will behave as if it is reasoning and as if it understands. As a result, we will have something like rather primitive and noisy upload, very resource hungry. It could be useful to create self-driving cars, which behave exactly like human drivers. But they will fail in non-standard situations. The size of human learning dataset is around 90 000 hours, or something like 100 TB of video, which is 100 000 times more than the size of Imagenet labelled database. https://www.fastcompany.com/1733627/mit-scientist-captures-90000-hours-video-his-sons-first-words-graphs-it The difficulty in training NN is growing non-lineary with the dataset size.
The first one you mention IS the OpenCog model ;) The second isn’t really an architecture.
There are ideas for AGI based on pure NN primitives—such as what DeepMind is working towards—but so far they are just ideas and napkin sketches. The only working general intelligence codebases are GOFAI to varying degrees at this time.
Personal phenomenological observation: when I write a text, I feel like some generative network creates a text stream similar to everything I read before, so it works like RNN described by Karpathy. But above it is reasoning engine, which checks if there is some meaning in this generated stream.
I could imagine two ways how to create universal mind using NNs.
One is to use NNs at low level, like image recognition, and something like rule based inference engine on the upper level, which will used data from NNs for decision making. I think that NN using self-driving cars are built this way.
Another is to create large scale functional model of a human brain, where each block will consist of NN, but the blocks will be rather independent of each other and only exchange information. There are at least 50 Brodmann areas in the human brain with different anatomy.
I am sure we will see an attempt to create universal robotic brains using similar approaches.
Added: The third way is to create very large NN able to predict the whole behaviour of a person based on the whole recording of his previous activities. It will behave as if it is reasoning and as if it understands. As a result, we will have something like rather primitive and noisy upload, very resource hungry. It could be useful to create self-driving cars, which behave exactly like human drivers. But they will fail in non-standard situations. The size of human learning dataset is around 90 000 hours, or something like 100 TB of video, which is 100 000 times more than the size of Imagenet labelled database. https://www.fastcompany.com/1733627/mit-scientist-captures-90000-hours-video-his-sons-first-words-graphs-it The difficulty in training NN is growing non-lineary with the dataset size.
The first one you mention IS the OpenCog model ;) The second isn’t really an architecture.
There are ideas for AGI based on pure NN primitives—such as what DeepMind is working towards—but so far they are just ideas and napkin sketches. The only working general intelligence codebases are GOFAI to varying degrees at this time.
Personal phenomenological observation: when I write a text, I feel like some generative network creates a text stream similar to everything I read before, so it works like RNN described by Karpathy. But above it is reasoning engine, which checks if there is some meaning in this generated stream.