Thanks to science evolution, the functioning of our brain is becoming increasingly clear, but we are far from fully understanding it. We know which areas are activated in response to certain external stimuli, we know the signal waves emitted during the execution of a certain operation. There is still a long way to go before we understand some of the micro-aspects, but the ones that make all the difference. In the cerebral cortex, there are about 100 billion neurons interconnected to form a gigantic biological neural network. Some areas of the cortex are concerned with reasoning, others with moving a part of the body, others with mnemonic activities, and so on. These neurons define us, by memory and behaviour. It is easy to understand that it is not possible, at least with the tools available to science today, to trace every signal in this mega-network to understand every micro-activity in this network. It is, therefore, necessary to reason by areas of brain activation, obtaining a less in-depth analysis, but one that is essential to initiate its activities.
Neural networks, the structure
Based on this model offered to us by nature, mathematicians and computer scientists have tried to use components that, at least approximately, can replicate the behaviour of a neural network. Artificial neural networks were therefore created, which are nothing more than mathematical models composed in this way:
- Artificial neurons: a mathematical model that receives input data that simulates the electrical signals received inside a brain and processes them according to the characteristics of the neuron, expelling a final signal as output.
- Interconnections: there are various types, depending on the needs. Some neurons are all connected, neurons connected in different layers, up to one or more neurons that provide the final output of the neural network. Some neurons can be connected to others in previous layers to provide a kind of feedback on the operations carried out up to that point, to correct the activities that have not been carried out in the best way and to stimulate those that are working.
But as the neurons receive input they process it and receive increasingly refined outputs, simulating machine learning. In particular:
- The network can become ‘good’ at replicating a behaviour it deduces from the data it receives (supervised learning).
- The network does not have examples of past behaviour, but “learns” to find patterns (unsupervised learning).
- The network understands whether it is doing well or wrong based on feedback it receives during execution (reinforcement learning).
Neural networks, the phases
All three modes, which depend on the available data and the problem to be solved, foresee a training phase to take in the data and create the overall mathematical model of the neural network (calculating the results from the single neurons and the modalities of inter-neuronal interaction). In the case of supervised learning, there is more than one testing phase to assess how similar the network’s results are to the actual past results. The third phase, that of prediction, allows to “guess” new results and use the network for the reason for which it was created: to predict outputs based on new data.
These are mathematical models, complex, but models nonetheless. A human being designs a network and executes it. The network, using the formulas governing it, produces an output which, however, at least ideally, could also be obtained by solving a very complex formula. But it is not a matter of reasoning, of something that could be called ‘intelligent’. Not even close. It is the complexity that conveys the idea of reasoning, but it is code and iterations.
Neural Networks, the speech at WAICF
This concept must be clear because it allows us to get away from so many misunderstandings and myths about artificial intelligence. Unfortunately, or fortunately, these calculations, which in the case of Premoneo serve to implement analysis and output in software developed to support the commercial activities of companies, still fall within the sphere of mathematics. Intelligence is just an illusion.
Massimo Dell’Erba, CTO and co-founder of Premoneo, will talk about the use of Neural Networks for business on Friday 15 April from 3pm at the Palais des Festival de Cannes, during the WAICF.
