It is possible to quickly create an AI business

It is possible to quickly create an AI business

There are several techniques to accelerate the development of a business-oriented machine learning model: from “all-in-one” algorithms to the transfer of learning via auto ML.

Rapidly developing business learning models is now fully possible. There are many tools and methods to accelerate this type of development. In the database we will of course find machine learning or deep learning libraries such as Keras, Apache MXNet, Pytorch, Scikit-Learn or Tensorflow. Websites (or “zoos”) also refer to a large number of models from these libraries, but tailored to specific objectives: automatic translation, image recognition, chatbot-oriented conversation management, product recommendation via reinforcing learning, etc. These sites include, for example, Model Zoo in deep learning or Rlzoo in reinforcement learning.

“The main public clouds, at the forefront of which AWS, Google Cloud or Microsoft Azure, offer managed services of all-in-one models,” comments Marc Fanget, Technical Director at Umanis. These offerings cover the basic uses of AI. Microsoft’s (Azure Cognitive Service) integrates e.g. detection of anomalies, identification of unwanted content, speech recognition and synthesis, natural language comprehension, sentiment analysis, etc. Translation and recognition of images are again part of the list. “These models will be trained with data specific to the company to respond to a specific business context. However, their structure can not be changed and the format of the training datasets will be limited,” emphasizes Marc Fanget.

Automated machine learning

Another technique for building an AI fast: automated machine learning (or auto ML). There are now many tools in this area. DataRobot and H2O.ai launched the very first auto ML platforms in 2012. AWS, Google Cloud and Microsoft Azure followed suit in 2018 and 2019. At the same time, computer science studies came on the scene. This is the case with Dataiku, Knime, Rapidminer and SAS. “Automatic machine learning enables a data analyst to quickly create a simple model, such as an image classifier, by training it from a dataset of tagged photos,” explains Didier Gaultier, director of computer science and artificial intelligence at Business & Decision, a subsidiary by Orange expert in data.

But to achieve a satisfactory result on complex predictions, for example in econometrics, in pharmaceutical research or in fraud detection, it will be necessary to personalize the models that are the result of auto ML, or even to combine them. “Let’s take the example of computer attack detection systems. Some attacks change their signature in real time. Therefore, it is no longer possible to detect them using conventional algorithms. To spot them, we have to go through semi- supervised or unattended, which requires the intervention of a computer scientist “, illustrates specialist Aymen Chakhari (read the article Will automated machine learning replace the computer scientist?). Auto ML is still a great tool for speeding up projects.

“To verticalize a neural network, we can simply complete our learning with a base of specific examples”

In this quest for speed, the transfer of learning is also not to be neglected. It consists of taking branched neural networks and then adapting them to the target problem. Explanation: A neural network that has learned to recognize a dog in a photo can be used to detect a cat. The latter also has a head, two ears, four legs, etc. The corresponding network layers can thus be reused. Only the top layers need to be added.

Obviously, we’re immediately thinking of Google’s BERT language model, and especially of OpenAI’s GTP-3. “To date, this is the best neurolinguistic model ever designed,” acknowledges Antoine Simoulin, R&D Manager within Quantmetry’s NLP expertise. The use cases of GTP-3 are abundant: of course translation, but also synthesis of text or adaptation of it to everyday language, automatic writing of articles, generation of an application from instructions in natural language, human-machine dialogue … Here, the neural networks can be supplemented with a layer centered on a specific task, such as filtering emails (with the parameters “spam”, “professional message”, “message about this or such themes” …). “But to adapt it, we do not necessarily have to change its architecture and its weights, we can also complete its learning with a specific base of qualified and correctly labeled examples that allow it to specialize in a particular vocabulary.” , adds Antoine Simoulin.

Transfer of learning

Resumption of such neural networks nevertheless generates a great limitation. “BERT has between 100 and 300 million parameters depending on the version, and GPT3 about 175 billion. Which makes them cumbersome to handle and train, including on small specialization datasets. The computer infrastructure to manage this learning will be a cost that few companies will have. afford “, warns Antoine Simoulin, before adding:” To meet this challenge, OpenAI has developed an API that facilitates access to GPT3. ” Another limitation to consider is that the results of a neural network are generally not mathematically explanatory. It can also be completely wrong in rare cases. Which makes it difficult to apply in certain areas (read Autonomous car: what artificial intelligence under the bonnet?).

This does not mean that transfer learning cannot be exploited. The AI ​​factory in a large group will be able to reuse the neural networks it has developed, for example to recognize categories of products to be referred within a sorting center (a non-critical process that therefore supports a small margin of error). They can be reused for other projects, typically to identify and automate routing of other categories of goods from other factories in the group. Focused on a specific task, the model in question will not be as large as BERT or GPT3, therefore the transfer learning will not involve as much computing power as for the latter.

Leave a Comment