A subset of the universe of AI (artificial intelligence), machine learning, or "machine learning", arouses a lot of interest, including among SMEs. What professions can it are used for? What resources does it require and with what skills?
Perhaps the most accessible machine learning to understand is image recognition. We submit to software a sufficient set of pictures or drawings that it is able to analyze by identifying shapes or objects recognized as being close to a model, because they meet similarity criteria in a given category, compared to those in memory.
It is an algorithm that does the job automatically, repetitively and with a "zero faults" approach.
In addition to image recognition (video surveillance, biometric identity, medical imaging, board games, autonomous vehicles, etc.), it can be combinations of complex or non-complex data (tables, histograms), incidents or behaviors, empirical situations - linked to databases or information on the web, including social networks data that is of interest to the business, marketing, sales or general management. It can also be words, phonemes (written or vocal translation), musical notes, organic components, materials, gases or liquids, specified by all kinds of sensors.

Generate decisions or actions

The results of the analysis can generate decisions, predictions, alarms, fraud detection, etc. - or even launch actions automatically or semi-automatically (human validation or removal of doubt may prove to be necessary, practically or legally). This means that all trades can find their account there.
On the methodology side, we must consider two phases, which can be in a virtuous loop: learning, or "training", of the model from the observations or data collected; and the release phase with generation of valid, unbiased results.
Most often, we ensure that the learning phase is "supervised": the data is labeled, pre-classified in order to direct the machine towards such or such "pattern" (model) sought. Otherwise, if we want to avoid any bias, the machine must first ingest huge volumes of data (see Big Data insights).
There is also the option of learning "by reinforcement": the algorithm learns by its errors (rewarded or penalized) - option typically applicable to games. It is always preferable that the data, once collected, is organized, cleaned to avoid biased modeling. Then, we select the algorithm (decision tree, linear regression, clustering, "k" nearest neighbors, neural networks, etc.) and we proceed to its training by iteration.
Then we run the variables again until the algorithm produces a correct result. Then, we can submit a new data set to it in order to further improve the model and avoid its "over-learning".

What skills?

The business user or manager wishing to use a "machine learning" application already configured and "trained" does not need specific skills other, precisely, than those of his business expertise which will allow the modeling and relevance to be validated results. The user interacts with the system without needing to know the coding or the technical details of the model used.
On the other hand, for designers and developers of "business" models and applications, a certain mastery of programming languages, such as Python, "R" or JavaScript is required. Likewise, a good knowledge of statistics or maths will be useful and necessary. There are many tutorials but also specific certifying training (from 3 to 5 k €). You can start by exercising on a Colab notebook (Google, TensorFlow) or on Amazon AWS or Microsoft Azure (ML Free then ML Studio), or consult the R Consortium (Microsoft, Dell, Oracle, etc.) for the various distributions. of the market.

Which resources: in the cloud or not?

Until recently, "machine learning" was confined to supercomputers or HPCs, in research centers or universities. But the market is opening up. As an indication, Dell offers, for example, a "formula 1" configuration for IA / machine learning application from 75 k € (ref. PowerEdge R940xa), with the possibility of associating 4 processors (Intel Xeon Gold 6240L) to 4 graphics processors (GPU / FPGA Tesla V100 accelerator card, 32 GB from Nvidia and Mellanox card (acquired by Nvidia) ConnectX-6 DX dual port 100 Gigabit-Ethernet ... with 6 TB of memory, one can dream of.
At Synapse Développement, French expert in AI (like DC Brain, Onogone, TinyClues ... a flowering of very dynamic start-ups), we believe that it is possible to start with a "standard" server equipped with at less 8 GB of RAM and 10 to 100 GB of "hard memory" (SSD). "The type of processor is not important but, in the absence of a graphics card, the more powerful it will be, the faster the models will train and respond," recommends Clément Tourné, development engineer at Synapse. “Good network access will facilitate deployments, but are not essential."
Which manufacturers to recommend? “Those who offer configurations with Nvidia graphics card may be preferred, because these cards significantly improve performance. But above all, it should be remembered that the material requirements are specific to each project and must therefore be defined during development,” observes Thiziri Belkacem, research engineer at Synapse.
Failing to be able to invest immediately, there is still the possibility of getting started on virtual machines using cloud platforms: all the tools are provided there, including use cases and tutorials. With pay-per-use, it's risk-free!