Fig. 5. The universal network
Thinking
I guess that the phenomenon of synesthesia is the result of multi-input and multi-output, because I don’t have other sensory data. Because the parameters of network shared by the multi-input and multi-output, so one input may affect other output when the two outputs are the attributes of some similar object, such as fire and sun which have red color and hot attributes. For example, When we see red(vision input), we may feel warm(warm sensation output), however When we see blue, we may feel cool. Because we can also use the fully connected network to process sequence data (The last paragraph is the proof)and nature language processing is also represented by functions, so nature language processing can use the same network as image recognition and speech recognition so long as the output node and the input node and more parameters add into the network. The network is universal so long as the network can process the different inputs. We can first send image or speech to the network and the output is language or other information, then send the language or information to the network’s input, process it or predict it. I guess that connection in mind can realize through the universal network and sending the output into input. For example, it’s cloudy now. This information is sent to the network’s input, the predict output is there will be a rain. This output is sent to network’s input, the predict output is that I need an umbrella. Because the information about it’s cloudy and there will be a rain is represented by image or audio or language, so only universal network can do it. The most likelyhood event can be seen as the effect of the cause. It’s causal reasoning. Connection in mind is the key of creativity, synesthesia is the assistant.
The Proof
The fully connected network can be expressed as that.
[1] (1)
(2)
(3)
has relations with because of . Meanwhile have relations with because of . So we don’t need recurrent neural network. Because fully connected network also can do what convolution neural network do, so fully connected network is the universal network.
Conclusions
1.The abstract network(prediction network) achieves the intended functionality of multi-attribute recognition through multi-dimension regression. The Concrete network can generate the output which the input rarely contained from the attributes the network learned. Its generalization ability is good because the network is a continuous function.
2.We can use one neural network to do image recognition, speech recognition, nature language processing and other things simultaneously so long as the output node and the input node and more parameters add into the network. The network is universal so long as it can process different inputs. I guess that the phenomenon of synesthesia is the result of multi-input and multi-output. I guess that connection in mind can realize through the universal network and sending the output into input.
References
  1. Ian Goodfellow, Yoshua Bengio, Aaron Courville, Deep Learning, MIT Press, (2016)
  2. Andrew L. Maas, Awni Y. Hannum and Andrew Y. Ng, Rectified Nonlinearities Improve Neural Network Acoustic Models, ICML, (2013)
  3. Tensorflow Tutorials and Apis, https://tensorflow.google.cn/learn last accessed 2020/3/20
  4. Jin-xin Wei, Qun-ying Ren, A Functionally Separate Autoencoder, unpublished