TECH | Jan 11, 2018

Name of the medicine: Deep Learning

If Deep Learning were a medicine: read the warnings and instructions for use very carefully


Sometimes a part becomes confused for its whole. For example, it is common to confuse the active ingredient of Deep Learning with a medicine like Artificial Intelligence. In fact Deep Learning is to Artificial Intelligence more or less what acetylsalicylic acid is to an antipyretic: aspirin is made of the former, but not all antipyretics employ acetylsalicylic acid to reduce a fever. Deep Learning, according to the Wikipedia definition,  is therefore “that field  of machine learning and artificial intelligence research that is based on  different levels of representation, corresponding to hierarchies of factors or concepts, where high-level concepts are defined on the basis of low-level ones“. A definition for sophisticated understanding that explains how Deep Learning is based on the development of Neural Networks in which each neuron is “trained” (and therefore “learns”) to perform a certain particularly simple function, added to the likewise simple one performed by other neurons, enabling the completion – thanks to a continuous learning process – of even very complex tasks with an ever lower coefficient of error.

The active principle therefore consists of a set of methods and algorithms that, exploiting the Neural Networks, grant the system an automatic learning process (Machine Learning) based on the structuring of successive levels of complexity, increasing in depth, which simulate the way in which knowledge is acquired by our brain.

Deep Learning is therefore part of the family of active principles called Neural Networks, which are in turn part of the family of Machine Learning’s active principles, the whole all incorporated into the generic medicine, Artificial Intelligence.


Like the Artificial Intelligence medicine, Deep Learning facilitates the automation of certain processes. One of the best known applications of Deep Learning relates to the recognition of images, shapes, sounds, voices, and their classification. When an image is provided to a Deep Learning algorithm, it is divided into many parts that will be passed to the “neurons” of the first state, which in turn will pass them to those of the next state and so on. The various steps allow the analysis and identification of significant traits and patterns to be passed to the next level to obtain a probability vector in which the algorithm assigns a value to each of the different possible categories.

Deep Learning can be used for the automatic detection of objects (e.g. traffic signs or traffic lights) in autonomous driving, for the identification via satellites of objects that allow the location of areas of interest and identify safe and unsafe areas, for detection of cancer cells in medicine, to improve the safety of workers in the use of heavy machinery by detecting the distance between people and any heavy objects in industrial automation, for electronic auditory and vocal translation, for domestic assistance and much more.


Deep Learning must be used in all contexts in which it is useful to “feed” an image or data algorithm in order to allow the automation of certain operations.  It is particularly effective where there is a need to manage large amounts of data that can be used as a “basis” for learning: in fact the system becomes more reliable the more it is able to optimize the probability vector; namely to learn from experience.


The effect of the medicine can be modified by the use of processors that facilitate improved and faster data-processing by the active ingredient. It has been scientifically proven that, in the case of Deep Learning applied to graphics, the use of graphic processors (GPUs) used in parallel to the CPU enhances the effects of the active ingredient. On the contrary, poor quality hardware resources completely negate its effects. Beneficial effects on the efficacy of the medication have also obviously been observed from the availability of enormous quantities of data to be processed.


How much

Deep Learning can be taken in large quantities and for long periods as long as the effects of the main ingredient are kept in mind. It is not harmful to health, except in the cases described under “Undesirable Effects”.

When and for how long

Deep Learning can be taken regularly. It must be stored in a cool place, possibly air-conditioned in order to allow the operation of the processors used to prepare the medicine.  It is shown that the effects of the medication are cumulative, so it makes no sense to adopt it for too short periods, in which the side effects are likely to be greater than the benefits induced by its consumption. It may be addictive.


Given the particular characteristic of the active ingredient there is no risk of overdosing, there is instead the risk of not achieving results or, worse, obtaining negative effects if the doses taken are too low.


Panic attacks in those who are not used to the efficiency of the machines, in those who believe that Artificial Intelligence does not produce beneficial effects and in all those people who are skeptical and frightened with regard to the possibility that automation will reduce jobs.

Disorientation in subjects who believe that Deep Learning is the only medicine in the Artificial Intelligence family.  Disappointment, frustration and depression in those who try to apply this active principle in contexts in which something else should be done regarding the application of a simple and continuous geometric chain of transformations that map a vector space in another.

Sonia Montegiove – Stefano Epifani