Usually, the characteristics of the output signals are treated as derivative values from the characteristics of the input signals in the most models of neuron. But what is the meaning and value of such derivative information? Already on the second step of such information processing its meaning is not obvious and is completely lost on the next steps. The value of all the internal network signals is reduced only to the formation of paths of input signals to a specific output. In this regard, there is a representation of an artificial neural network as a kind of "black box", where the semantic interpretation can be given only its input and output signals.
Yuri,
Consider for a moment a very simple neural network composed of 4 neurons. Two neurons are for output, one is for input, and the last is for strengthening or weakening the pathways between the I/O neurons. Based on stimulus to the input, and feedback from that 4th neuron, the output will be different, and the resulting neural network and future outputs will be different from the previous (non-linear).
This may seem like a black box since future states are largely unknown, however the logic behind the system is quite simple to understand.
Now consider massive neural networks or even the human brain. Cells and connections between cells are constantly changing based on input and feedback, meaning it is very difficult to capture any single logic state. In a sense, it actually pays to think of this kind of dynamic system as a sort of black box since it is constantly and dynamically changing.
I do think however, that the methods used to provide feedback, groupings of neurons, and number of connections between networks can be regulated, and fall outside the "black box" paradigm. By measuring and regulating feedback or monitoring tendencies of certain states, we don't have to ignore the "second step" you mention above.
Perhaps this answer is a bit abstract, but I hope it communicates the general idea effectively.
.
you might be interested in the following :
Artur S. d Avila Garcez, Luis C. Lamb, Dov M. Gabbay
Neural-Symbolic Cognitive Reasoning
Springer
http://ip144.qb.fcen.uba.ar/libroslfp/Biologia/Neurociencia/Garcez%20-%20Neural-Symbolic%20Cognitive%20Reasoning%20%28Springer,%202009%29.pdf
(warning : the above link seems a deep link which was not necessarily meant not be publicly available ...)
.
I read about a revival of parallel computing regarding the study of neural network. It seems to refer to the holographic paradigm...
In traditional multilevel feed-forward sigmoid NNs, each neuron can be loosely treated as a hyperplane that split the input hyperspace into two subspaces. Hence a single-layer network can describe only continuous areas that are the intersections of hyperplane-separated subspaces. Two- and three-layer networks can describe more complicated objects. Perhaps, some inner-layer signals can tell you something, but generally multilayer NNs are used when we do not know a simplier model to describe the patterns of the input space. Since backpropagation is a deliberately black-box approach, there is no guarantee that any internal neurons have "meaning", which one can easily verbalize. However, there are some NN architectures that are not trained by backpropagation, but are constructed from input space features. Radial-basis networks are a good example.
- Badges
- Science topic