Artificial intelligence and machine learning

"Even a machine will learn if you tighten its string!" That's how I answered someone who was wondering about the subject. It also describes the fact that it isn't a case of
magic. It's just the twisting and turning of data.

Matti Leskinen
Research Software Engineer, ATR Soft

Artificial intelligence guide: machine learning

Since the previous decade, research into artificial intelligence has been experiencing a new boom, and many practical artificial intelligence applications have become part of our everyday lives. At their best, computer systems are capable of bewilderingly human-like functions: people and objects are automatically recognised from images, virtual assistants recognise speech and understand instructions, self-driving cars are becoming a reality.

How has this discipline, which has been studied for decades, and which is also a subject of science fiction, once again become a hot topic and AI a buzzword?

Are you interested in the basics of AI?

Download the guide here

AI expertise
in an internal project

“….The hourly records of a specialist are checked weekly by a supervisor and approved or rejected. We didn't want to build such a bureaucratic and inefficient process”

Teemu Tasanto
CEO of ATR Soft

A few years ago, we mastered modern AI tools through an internal development project. The review process of our timesheets was selected as a problem to be solved. Traditionally, a process has been built into the review of specialists' timesheets, where, for example, the supervisor reviews the timesheets on a weekly basis and approves or rejects them. We didn't want to build such a bureaucratic and inefficient process, so we wondered whether artificial intelligence could come to our aid in this matter.

We got going on the existing data. We quickly went through the hourly records for the last ten years and divided the records into those for approval and those to be commented on.

We selected the features from the data to be examined, i.e. the features from which a category can be inferred (e.g. explanation of an hourly entry, project, task, etc.). For data security reasons the data was converted into such a form that no outsider could deduce personal data from it or any other sensitive information. Conversion took place using, for example, ID codes instead of text and by converting the text into Soundex form. From the existing data, we extracted part of the data as a machine learning training set and the rest as test data. We selected the Azure Machine Learning tool as a platform. From there, we selected suitable algorithms for data processing, taught the system with teaching data and tested it with test data. After a few iterations, we found the test results to be good enough and began to put the solution to good use..

The final project was just basic work for us, implementing the integration between the timesheet system and Azure, developing the user interface for the timesheet system and implementing the business logic. The result was our own artificial intelligence, AIMO, which tirelessly checks all of our records and sends friendly messages if there is something wrong with them.

AI meets many kinds of needs.

Following internal takeover projects, we have used AI and machine learning tools to address a range of customer needs.

Python

Pandas

Numpy

Scikit-learn. TensorFlow

PyTorch

Fast.ai

HF. transformers

FastText

Jupyter

Azure ML

Case 1

Highlighting important information in the mass of documents

Important information can often be lost amidst the avalanche of documents. It is impossible for for a person to go through such a large number of documents. We taught the FinBERT language comprehension model to identify suggested actions in maintenance and inspection reports, and implemented an interface to find, browse and highlight them in the documents.

Case 2

Digitising material and mining for information

An enormous volume of material is still on paper or in scanned form. Packaged solutions for digitisation do not always provide good enough results for reading different document bases and handwritten text. We have developed precision methods for reading handwritten text from scanned tabular documents.

Case 3

Noticing errors in the manufacturing process

Measurement data for industrial processes is valuable for the development of operations. Using time series analysis and machine learning, it is possible to make more accurate interpretations and predictions than by looking at visualisations of the data. We have developed models and methods to identify and classify faults in the semiconductor component manufacturing process.

Case 4

Understanding the purpose of search in a search system

A search that works badly is frustrating. An intelligent search on the Internet has been an everyday thing for so long that people want the same thing from all systems. We have redesigned the text search functionality of the old public administration system, adding semantic search features, i.e. a better understanding of the meaning of the search and the meaning behind words and sentences.

Case 5

Workload and project completion assessment

Machine learning can also be utilised in software development. Our students have explored how project management tools and time tracking data could be used to assess the workload of development tasks and the degree of completion of a development project.

AI and ML tools continue to evolve at an accelerating pace, so continuous learning about them is really important. We minimise the risk of developing technologies for our clients, for example by proving the achievable benefits right at the start of the project.
We deepen knowledge about our articles

Working together and
discovering things

Why is now precisely the right time to start making use of artificial intelligence in business, even though it has already been talked about for 20 or so years?

Developments in the sector have led to a situation where it is possible to achieve really good results as technology becomes ever more accessible and efficient. Costs and risks are much smaller when you can take advantage of ready-made libraries of programs and develop tools. It is no longer necessary to develop and implement new algorithms for each application from scratch with the scientific community.

The threshold is already low. With the right data, it is possible to gain sufficient confidence in the performance and benefits of a machine learning solution with very little effort, before having to invest more.

We help our customers to find out what their data can be used for.

Shall we discuss more about the topic?

You are sending a message to our CEO Teemu who knows most about our customer accounts.