Knowledge and expertise are the core values ​​of CIMSOLUTIONS. Focused on the quality of its secondment services, CIMSOLUTIONS continuously invests in the development of its own employees in the form of thematic growth paths, Special Interest Groups and Competence Centers. The latter focus on two aspects: building internal knowledge and expertise while developing the viable and demonstrable products for the external clients – both are necessary ingredients for the asset-based consulting [1]. In the field of data and artificial intelligence (AI), the 'Competence Center AI, Machine Learning and Robotics' focused on the innovative development of data-driven applications and the use of artificial intelligence; this includes a number of in-house projects, one of which can be seen at the bottom right of this page at the green button.

Traditional programming vs. self learning system
Developing AI systems is slightly different from traditional software. In traditional programming, explicit rules are laid down in the code. These fixed rules along with the input of new data ensure that a simple output is always produced by the system. In contrast, a different approach is needed to let AI learn on its own. Instead of the established rules (which we often don't even know ourselves) we give a number of known answers so that the machine itself can formulate the rules based on these answers and related input. Once AI has established the rules, the system can determine its own response for new input.

Figure 1 – Functional difference between traditional programming and AI system

AI is a broad concept and can be divided into two parts (Figure 2):

machine learning focuses on self-learning based on some statistical properties of the input data and connected responses; 
deep learning (part of both AI and machine learning) focuses more on self-learning from large (often unstructured) data by training complex multi-layered neural networks. 
The latter is becoming increasingly popular nowadays with growing amounts of data and computer computing power, because it can solve much more complex problems such as image, speech or text recognition. This trend is also reflected in the market – Gartner has 7,777 references for the term “deep learning”, far more than for “machine learning” (5,639) and “AI” (4,457) [3].

Figure 2 – the scope of AI / Machine Learning / Deep Learning [2]

How does modern software learn to see, listen and talk?
Just like traditional software, the neural network works with the digital numerical input (digits). For this reason, this system cannot learn directly from text, audio or video input and must first convert this information to numbers. The ways in which this input is converted varies per type of problem:

For image/video recognition, the network learns based on the pixel encoding of an image: length x width x RGB (color) scale; the same goes for videos, with the difference that it looks at all the single frames as if they were single images;
For text recognition, the network first converts all words into numbers, which are determined based on the similarities in meaning between the words. The closer the meanings of the words match, the closer their numbers will be.
For speech recognition, the analog audio sound must first be converted to a digital signal, for example by an analog-to-digital converter. The problem is further reduced to text recognition by breaking the digital signal into phonemes (smallest units of sound) and matching the most probable words to them;
Standing on the shoulders of giants
Training deep learning models against huge datasets can be computationally intensive and require multiple learning iterations that can take many days, even on powerful graphics processing units (GPUs). Therefore, it is often a good idea to first look for pre-existing models that have already been trained on similar data and then adjust the model parameters of these so that only a few specific iterations can be used to train the model on the specific data .

Figure 3 – concept of transfer learning [4]

From prototype to production
The main goal of training AI systems is to eventually put these systems into production so that they can make predictions based on new data. Several additional components such as data management, model serving and monitoring are required for the robust operation of the self-predicting system. As part of this, a discipline has recently emerged called Machine Learning Operations (MLOps , or DevOps for Machine Learning [5]) that specifically focuses on making the end-to-end lifecycle of predictive model development robust, reproducible, and manageable in any part of it.

Figure 4 – Iterative-Incremental Process in MLOps [5]

Following these principles ensures stable predictions and timely signaling when the model needs to be retrained (for example, if input data or business processes change).

Klim Mikhailov
Machine Learning Engineer

[1] https://www.gartner.com/en/documents/3990228

[2] https://master-iesc-angers.com/artificial-intelligence-machine-learning-and-deep-learning-same-context-different-concepts/

[3] Overview of Gartner search results (until publication date of this blog post)

[4] https://www.v7labs.com/blog/transfer-learning-guide

[5] https://ml-ops.org/