In an animal brain, there are cells called neurons. Schematically, they consist of a nucleus and an axon. Neurons interact with each other through interactions between the axon of a first neuron and the nucleus of a second neuron. A neuron can be connected to many other neurons. We can see that when an animal learns, new connections are made between neurons, and that when an animal thinks, there is a stimulation of neurons, with activation/inhibition.

An artificial neural network (machine learning) is composed of artificial neurons. These neurons are grouped in layers. Each neuron is connected to the neurons in the neighbouring layers. Depending on the stimuli received by the axons of the neurons of the previous layer, the neuron will itself stimulate or not the neurons of the next layer.

The training of the neural network consists in presenting several data (Xi) as input, the result of which is known for each (Yi). Using a regression algorithm, the weights of the neurons are modified to minimise the difference between the expected result and the result obtained.

For example, we provide photos with numbers ‘1’, ‘2’ and ‘3’ as input: 1000 photos with ‘1’, 1000 photos with ‘2’ and 1000 photos with ‘3’. We start with neurons with random weights. We input the 3,000 photos successively and observe the number of errors in the output predictions, for classes ‘1’, ‘2’ and ‘3’. We then modify the weights to try to achieve a lower total error. The model is finished being trained when the minimum error is reached for the training dataset. A test dataset (photos not used in training) can then be used to check that the neural network is capable of predicting something other than the photos it has already learned.

There are other algorithms in the same family, such as random forests or boosting gradient.

What all these methods have in common is that they require a large training dataset, and are black-box, i.e. the weights of the neurons do not make it explicitly clear why it chooses this or that output class.

(There are, however, different variants that allow to limit the input data, and to explain part of the black box).

Rational human reasoning is formulated in the form of symbols: ‘if’, ‘and’, ‘or’, ‘implies’… For at least two and a half thousand years, scholars have sought to refine these symbols so that they can represent the most complex reasoning. The most famous ancient scholar in this respect is Plato. But since then, every generation has sought to improve this language, even in the Middle Ages. The current form used is mainly the work of Bertrand Russell in the early 20th century.

Symbolic AI aims to use this formalism of logic rules. For example, to recognise numbers, a number of rules will be defined: a ‘1’ has a large vertical bar and a small slash at the top left, etc.

By its nature, symbolic AI does not need a large amount of training data, and there is no black box, so there is full explicability of the results. On the other hand, input rules must be defined, which requires a human analysis of the problem.

Connectionism and symbolism are the two most popular forms of AI today.
They aim to replicate the functioning of a human brain, but at different levels of abstraction.

Connectionism reconstructs the local functioning of neurons, while symbolism reproduces human reasoning methods.

Connectionism, by nature, consumes a lot of resources, but is very powerful.

Symbolism, by nature, is a low user of resources but is currently not very efficient for industrial tasks.

Hybrid AI (or Neurosymbolic AI) is the fusion of these two types of AI.

Synonymous with moderation and sobriety by its etymology, the concept of frugal AI is a more responsible way of designing AI. Frugal AIs are capable of learning from very little training data to reduce their consumption. Frugal AI also aims to optimise the resource consumption of an infrastructure or system as much as possible, to reduce its impact on the environment or even make it positive.

Muvraline is certified ISO/IEC 27001:2017 for information systems security. This certification represents our commitment that we do our utmost to secure our customers’ information, whether digital, paper or cloud-hosted.  In this way, we limit our vulnerability to the growing threat of cyber-attacks.

Know more