Cloud Computing and the Democratization of AI
By Raul Reguillo (graduated ESI and Cloud Architect at BBVA Next Technologies).
1. A walk in the clouds.
For years, technologies cloud They play a crucial role in any company. Whether it's a Startups, an SME or a large company, the cloud has positioned itself as an essential ingredient when defining the technological strategy [1]. Characteristics such as the elasticity of resources, adjusted and predictable costs, security or the incipient number of services offered, position cloud technologies as a very attractive value for companies. In data, [2] shows us an example of adoption and projection.
Within these services, they can be found from the classic ones, aimed at storage, databases or computation, as well as the most avant-garde, exploring the horizons of Quantum Computing [3] or pushing the limits of Artificial Intelligence further.
This quality of bringing resources, knowledge and services closer to people is known as democratization. Throughout history, many examples of democratization have been seen, one of the most relevant and initial being that of the invention of the printing press; When books were luxury objects available to only a few who knew how to read, interpret and take advantage of them, the printing press made it possible to make them more accessible to the masses and, therefore, bring knowledge to them. We can thus think of many other examples that can range from transport and energy to news and telecommunications. The democratizing role of the cloud in technological trends is remarkable, extensive and it is possible that it will be one of those paradigms that change the way computing works [4]. Likewise, the number of cloud service providers favors global coverage as well as very affordable prices [5]. It is no longer necessary to make a large investment in servers and disks to host information, but virtual space can be rented for the precise time required at very affordable prices and delegate aspects of security, maintenance and compliance to the provider itself cloud. Likewise, a small company or even an individual can thus rent an entire cluster of parallel processing (potentially hundreds of nodes) to make use of it the time strictly necessary, and then release it. This powerful concept can be extended beyond computing or storage, reaching other classes of more specific services (the list is very extensive [6]) such as IoT, Blockchain, data migration, robotics or Artificial Intelligence.
2. Democracy
It is in this last field, Artificial Intelligence, where the effects of this democratization seem most evident and in very diverse fields. The cloud came to democratize many things in the form of services (PaaS, SaaS, IaaS), the field of Artificial Intelligence being a particular case where several of these branches converge. In this article we will talk about three aspects where the democratization of AI has the most impact: resources, knowledge and operations.
2.1 Resources
If we think about training a model of Deep Learning the amount of resources that some of the algorithms require as well as the need to hardware As are high-end GPUs or TPUs, they can be out of the budget of many modest companies, students or researchers. This material would only be available to companies with sufficient solvency that could buy it and take full advantage of it. Let's see it with an example:
ResNet-50 [7] is a model of Deep Learning oriented to the classification of images. If we train ResNet-50 for 90 epochs with the ImageNet dataset, it can take up to a week with a single GPU. As we increase the number of GPUs, the training time can be reduced, reaching more reasonable times with at least 4 GPUs. On the other hand, the dataset itself is quite large: around 150 GB. Since this is a simple (and relatively old) problem, we begin to see that this kind of data is perhaps better to have, as well, in a distributed environment.
It is quite evident that for certain problems of deep learning the solution goes through distributed environments, both for computation, data storage and let's not forget the inference or exploitation of the model in productive environments. At this point, we can make calculations of how many nodes we want to have, what use we are going to give them, how many GPUs we can afford and how we are going to keep everything updated (library versions, patches, etc). Nothing that is not affordable, but it does require a offset considerable metawork.
In an environment cloud however we enjoy a layer of abstraction of all this. Distributed storage is extremely cheap (cents of euro per GB per month) and the different training services already offer standard configurations to carry out this kind of training: number of nodes, capacity of the machines, types of GPUs, libraries and versions to use. Thus, with a few clicks, you can count on an environment ready to train our ResNet-50 in a matter of minutes, with the cost of the service being around a few euros.
2.2 Knowledge
Perhaps one of the most powerful aspects where the cloud can bring AI closer to the user is in relation to the exploitation of knowledge. There is already a catalog of trained models available for use. A company doesn't have to train an image classifier or recommendation system from scratch; There are already services that make use of pre-trained models that can be exploited by anyone. Among the most successful are sentiment analysis services in text or image, models of speech to text or text to speech, fraud detection, entity recognition, forecasting And a long etcetera. Thus, a company can make use of this knowledge in the form of services, connecting them to its solutions as a component of more and paying a monthly fee to exploit them. It is no longer necessary to have trained a model within the company with the problems that this entails; lack of data, prediction errors, retraining and mainly acquiring the necessary knowledge and profiles to carry it out.
2.3 Operations
Finally, one of the last chapters added to the service catalog is operations, that is, facilitating the necessary work that must be carried out (that meta-work mentioned before) to train and put an Artificial Intelligence model into production.
Cloud providers, aware of the complex loop that governs this kind of projects, greatly facilitate the creation of pipelines necessary to extract the data, train the model, take it to production and monitor its performance. This set of processes, generally known as MLOps [10] is also democratized by cloud providers through an abstraction layer, so that with a few clicks all the necessary operations are covered.
3. CONCLUSIONS
The cloud has come to revolutionize the way we approach processes and services. In the field of AI, the impact has been notorious and the number of services that are launched each year related to this field seems to know no end. Of course, the technologies themselves cloud they require a certain learning curve to be used, so we are not exempt from this meta-work to make efficient use of them, although they tend to be smooth learning curves and the benefit obtained in terms of time and cost optimization is more than evident. Now more than ever, Artificial Intelligence at any scale is within everyone's reach.
