Reach Us +447482877761
Are Neural Networks Imitations of Mind? | OMICS International
ISSN: 0974-7230
Journal of Computer Science & Systems Biology

Like us on:

Make the best use of Scientific Research and information from our 700+ peer reviewed, Open Access Journals that operates with the help of 50,000+ Editorial Board Members and esteemed reviewers and 1000+ Scientific associations in Medical, Clinical, Pharmaceutical, Engineering, Technology and Management Fields.
Meet Inspiring Speakers and Experts at our 3000+ Global Conferenceseries Events with over 600+ Conferences, 1200+ Symposiums and 1200+ Workshops on Medical, Pharma, Engineering, Science, Technology and Business
All submissions of the EM system will be redirected to Online Manuscript Submission System. Authors are requested to submit articles directly to Online Manuscript Submission System of respective journal.

Are Neural Networks Imitations of Mind?

Gaetano Licata*

Gaetano Licata, Chair of Logic and Philosophy of Science, Dipartimento di Scienze Umanistiche, University of Palermo, Italy, Viale delle Scienze ed. 12 Palermo, Italy

*Corresponding Author:
Gaetano Licata
Chair of Logic and Philosophy of Science
Department of Humanities, University of Palermo
Via Catania 166, Palermo, 90141, Italy
Tel: 339-456-8136
E-mail: [email protected]

Received date: February 23, 2015; Accepted date: March 11, 2015; Published date: March 13, 2015

Citation: Licata G (2015) Are Neural Networks Imitations of Mind? J Comput Sci Syst Biol 8:124-126. doi:10.4172/jcsb.1000179

Copyright: ©2015 Licata G. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Visit for more related articles at Journal of Computer Science & Systems Biology


Artificial neural networks are often understood as a good way to imitate mind through the web structure of neurons in brain, but the very high complexity of human brain prevents to consider neural networks as good models for human mind;anyway neural networks are good devices for computation in parallel. The difference between feed-forward and feedback neural networks is introduced; the Hopfield network and the multi-layers Perceptron are discussed. In a very weak isomorphism (not similitude) between brain and neural networks, an artificial form of short term memory and of acknowledgement, in Elman neural networks, is proposed.


Artificial neural networks; Recurrent networks


Nowadays we are out of the illusion that computers can be good models for human mind. The human mind is the result of the biophysical structure of a nervous system in a body which evolved to survive in the environment, in communication with other individuals of same speciesand in relationship with other species of the ecosystem: its power is due to a very long and hard evolution and we are not able to understand its complexity [1].

(A)The goal that A.I. should attain is the emulation, through a computer, of some processes of mind in relationship with the environment (the world and other individuals).

With respect to this objective I want to underline two obstacles in neural networks strategy: 1) and 2).

1) Neural networks are a strategy to emulate directly the behavior of brainand not the behavior of mind. Thus an important problem that neural network strategy misses is the gap between brain and mind. This is the problem of the translation of states of neuronal activation in concrete mental activity. The mind/brain translation problem will not be overcome until we will not have a clear theory about thought, consciousness, perception and action as cerebral phenomena. Moreover, if this theory wants to be useful to neural network strategy it must be conceived following the neural network philosophy and language. A theory who speaks the language of neural networks should consider thought (i.e. mental representations, planning, consciousness, remember and so on), perception and action not as “states” but as fluxes of states which go through the network (ordered and structured sets of states which go through the network). About these fluxes that we, asthinking brains, perceive in ourselves, we haveunclear ideas on their beginning, on their developing and on their ending, but we know that perception can generate them.

2) Artificial neural networks are very poor imitations of brain. Human brain is a “network” of 100 milliard of neurons in which each neuron is connected to many thousands of other neurons, so, in a brain; there are millions milliards of connections. There are many kinds of structure of neural networks, but the architecture of the most common neural networks consists in a simple three layers structure of artificial neurons, like the three layers “perceptron” of Figure 1, that henceforth I will call TLP.


Figure 1: The TLP is an example of feed-forward neural network: the lower level is the “input layer”, the medium level is the “hidden layer” and the upper level is the “output layer”.


Neural networks can be feed-forward or feedback networks. In feed-forward neural networks like TLP the information propagates in only one direction, from input layer to output layer through the hidden layer (that can be more than one), and there are no cycles. Each unit is connected with every unit of the following layer, there are no connections between units of the same layer or with a unit of previous layer, and there are no connections which jump one layer. A feedforward network simply calculates a function of input values which depend on the distribution of weights (w) of the incoming connections and on the activation function of the outgoing connection. It has not any internal state different with respect to the weights of connections.

In feedback networks (also called ‘recurrent networks’) the connections are arbitrary. The Hopfield network (Figure 2) is a fully connected graph, typically represented as a matrix of weights; it has bi-directional connections and symmetrical weights [2]. There is no input or output specific layers, all neurons are input and output units; activation levels are only +1 or -1. These kinds of network, with very high redundancy of connections, produce associative memory and permit the recovery of missing information.


Figure 2: An example of feedback network: the Hopfield network.

Sometimes human brain behaves as a feed-forward network with layers, but it has also many connections that lead information backward to neurons of “preceding layer”, i.e. the brain is a feedback network in which can be many cycles of neurons. Given that sometime the activation goes back to neurons which have caused it, feedback networks (and the brain) have an internal status memorized as activation levels of units. In recurrent networks the computation has much less order with respect to feed-forward networks.Artificial feedback networks can become unstable, chaotic or can fluctuate and it can be very hard to obtain a stable output from a given input; so it is a mystery how our brain, as feedback network, is able to produce its(so good) computation.

The learning process, in a neural network, is commonly understood as the transformation of the state of the network in direction of a specific goal: a neural network changes its state updating the weights of its connections. In this aspect neural network and brain are considered similar, but brain’s learning is much richer than neural network “learning”, because a lot of fluxes of modification are needed in the brain to learn and to stably change the structure of connections between neurons (memory). On the other hand, in neural network theory an output generated after a certain number of epochs confirms that the network “has learnt”. A neural network is an adaptive system which changes its structure on the basis of external information. With respect to the comparison between the Hopfield network (Figure 2) (or any other feedback network without input and output specific layers) and the TLP (Figure 1), it is clear that TLP, for the need of order in computation, for the absence of chaotic fluctuations and for the idea that learning is a process which starts from precise data and has a precise target, has attracted the interest of scholars. The way to obtain a good emulation of mind, indeed, is not the precise imitation of the structure of brain, even if we are speaking about neural networks, which were conceived as imitations of brain by W.S. McCulloch and W. Pitts in 1943 [3] and by F. Rosenblat in 1958 [4]. Actually, neural networks like TLP should be not conceived as models of brain but as good schemes of nonlinear computation, nonetheless neural network theory speaks about artificial learning.

A classical learning process for networks like TLP is the “supervised learning” with back propagation algorithm [5], which has today a lot of technological applications. In the case of “supervised learning” the network learns the unknown relationship between the input variables and the output variables, so the network, after the learning process, is able to “make previsions”, i.e. to give outputs on the basis of inputs similar to those of learning process. The “training set” to administrate to the network contains typical examples of inputs with relative outputs (in ordered pairs); when all the training set is administrated to the network, the network will be able to associate, to a new input, the desired output with an error that the network can correct. The error in the output can be corrected through the comparison with the “expected output” (supervision). Given that the output is expressed as neural activation of output layer, the error is a difference of activation between the output proposed by the network and the “expected output”. To delete this difference, according to the strategy of Rumelhart et al. [5], the information is propagated back from output layer to hidden layer (or to hidden layers), until the input layer. Step by step, backwards, the network modifies the weights and the activation functions which bind the units, to minimize the difference between the resulted activation of output layer and the activation desired in output. The network, in this kind of supervised learning, has power of generalization: it is able to deal with unknown cases knowing similar cases, as in natural logical induction.

As it is known the weight is the scalar parameter of synaptic “strength”of incoming connections: e.g., in Figure 1, the first arrow, from the first neuron on the left of input layer, to the first neuron on the left of hidden layer. An important incoming connection will have a high weight, while a less important connection will have a low weight. Thus the learning process is the creation of selected connections, between the units, from the input to the output: this is considered a similarity between this kind of neural networksand the brain. Supposing to draw, in TLP, some arrows fatter than others, we could have a representation, in a very smaller and simpler scale, of how the synapses of biological brains are reinforced or weakened by learning and memory. Is it the matter really so simple?

Although TLP is a neural network and it always works in parallel as a nonlinear function, its simplicity permits a linear explication of its state between a complete computation (“epoch”) and another. Let us call the neurons of input layer, from left to right, I1, I2, I3; the neurons of hidden layersH1, H2, H3, H4; and the output units O1 and O2. Calling Φ the outgoing activation function of the units and Σw the weighted sum of incoming connections, we can write that, after the exposition of the network to an input, TLP will give the output O1^O2 having the state STLP:



In this conjunction the difference between connections is given by the context of functions, e.g. the connection I1, in the context of equation, will have a different weight with respect to the I1 contained in equation so the whole function equation, in the context of ΦO1{…}, will have a different value with respect to equation contained in ΦO2{…}. This way to explicate linearly the state of TLP show its computational order, the internal relationships between the connections and their values, and contains the idea that the activation function (Φ) of a neuron is like a point of view on the weighted sum (Σw) of its incoming connections.


To conclude this discussion about feedback and feed-forward networks, I want to introduce a hybrid class of neural networks which has interesting properties: the Elman networks (Figure 3). In 1990 Elman [6] proposed a recurrent form of feed-forward networks similar to TLP creating bidirectional connections between the hidden layer and a layer which is“contextual” to inputlayer. The“contextual” layer has the same number of units of hidden layer and to its neurons is assigned 1 as constant weight. The contextual layer has the function to register the state of the hidden layer during the computation. Therefore the function learnt by the network will be based on new inputs and on the state registered in the contextual layer, so the network could learn which the state to remember is. The computational order of usual feed-forward networks is respected by the Elman networks, and a certain kind of artificial “short term memory” is realized by the specific and controlled feedback cycle. Moreover, given that the copy layer is like an alternative input layer, the back-propagation algorithm can be employed in this kind of networks, and this is a great advantage because the back-propagation algorithm is a good learning technique and it cannot be usually employed in recurrent networks [7].


Figure 3: The Elman network: we call the context units cluster or layer C.

Suppose that a self-driving car, equipped with an Elman network, receives as inputs not only the current perception but also recently past states, registered in the contextual cluster; it is possible to hypothesize that a simple “comparison mechanism” allows the network to identify a precise set X of “similar input states” (which are similar, say, when they are identical at least for the 75% of the state of the cluster) and to change its behavior as consequence of the increase (over a fixed threshold TX(f)) of the frequency, in the temporal steps (t1, t2, and so on), of states of X in input. If we want to transform our TLP in an Elman network, we should only create a bidirectional connection between the hidden layer H and a contextual cluster C, which will register the state SH of the hidden layer H, perceived at the temporal step t0, that now we can write as:


As we have told, the “acknowledgement” will be caused by the simple increase,over the threshold TX(f), of the frequency, in the temporal steps, of input states of type X; indeed the perception alone does not cause the increaseof the frequency over the fixed threshold TX(f) nor the memory alone (memory is represented by registered inputs which are activated from layer C in direction of hidden layer H at a regular frequency, which is lower than the threshold TX(f). In this case, it is clear that the flux of perception inputs plus the flux of registered inputs will cause the increaseof the frequency of input states of X typeover the threshold TX(f). Therefore the network will change its behavior as effect of the coupling of perception and short term memory, in a “resonance” which can be considered the homologue of biological acknowledgement. In this way it is possible to project a self-driving car which “stops” or“escapes” or “follows” an X type object moving in the neighborhoods, but only when the movement of the X type object isacknowledged by the network. The X type objects, which cause to the network the X states in input, can be selected with many characters, so we can design the network to react only to a very precise class of objects, and to react differently to many different classes of objects (X, Y, …, N), if we correspondingly increase the number of registration clusters.

Thus, in the theoretical frame of Elman networks, employing a very poor form of programming, it is possible to give an“intelligent” behavior to a system in which forms of dynamic short-term memory and acknowledgement are at work, a system in which perception, memory and action are due not to states of the system but to flux of states which go through the system.

I thank Giuseppe Nicolaci and Marco Buzzoni for their irreplaceable help in my research.


Select your language of interest to view the total content in your interested language
Post your comment

Share This Article

Article Usage

  • Total views: 13814
  • [From(publication date):
    May-2015 - Nov 15, 2019]
  • Breakdown by view type
  • HTML page views : 9916
  • PDF downloads : 3898