Now that you have implemented the AI-powered solution into your network, you can just relax and let the artificial intelligence work for you, right? Wrong!
Why?
Because the relationship between a human and a solution powered by machine learning capabilities still must be a two-way road, with continuous “traffic” in both directions.
To be more precise, the formula for success goes like this: raw network data + human insight + artificial intelligence = high-performance network + accurate business decisions.
So, to be on the right track from the start, artificial intelligence is not omnipotent in itself, even if it looks like that to most users. In many cases, it blends seamlessly into the product to the point that users don’t even notice its existence. As one research showed, 77 percent of surveyed consumers utilized a machine’s AI capabilities in one form or another, but only 33 percent were aware they were doing so (source: Artboard 12 (pega.com). Maybe that’s why we take the potentials of solutions and products powered by artificial intelligence for granted.
The potentials of AI technology are immense, but…
By now, it is well established that the adoption of AI technology helps organizations generate long-term value and reduce risks in managing complexity.
But, as said before – AI is not a magical tool that will solve your issues and make your wishes come true as soon as applied! And there is no one-size-fits-all, even if you have an advanced solution powered by machine learning capabilities in your hands.
Despite massive investments and rapid advancements in AI technology, its field of application is still narrow and very use-case sensitive.
Businesses and institutions must still learn how to capitalize on AI technology's full potential. Although, as surveys show, the trendline is ascending.
For example, the Stanford University report “2023 State of AI in 14 Charts” reveals that 2023 sees an increase in job postings seeking AI skills across all sectors, and the number of AI job postings overall was notably higher in 2022 over the prior year. Also, even though corporate investment in AI dipped in 2022 from 2021 highs, the number has still increased 13-fold in the last decade.
It remains to figure out – how to harness AI potential to the fullest.
The experts claim that the key is the scaled AI introduction and continuous customization, which involves human intervention.
How AI-powered system works – a backstage view
To understand this thesis better, let´s start by looking at the different logic between traditional and AI model software deployment.
Traditional software has a code as a foundation and is programmed to perform a certain task.
On the other hand, with AI model software, code serves a different purpose. It still has to be written, but the foundation here is data, from which AI learns to perform the task.
AI model software consists of several subcomponents, data pipeline items, ML pipeline items, and workflows.
Thus, processing techniques and machine learning reach the levels that automation requires. And unlike traditional software, this is a continuous, adaptive process with periodic human intervention.
What happens with an “unguided” machine learning/AI-powered system?
When exposed to continuous new input data flow without human control and without models to rely on, “perfect” algorithms and AI models in live production environments can become disparate and devious, prone to different biases. It makes them potentially fragile and exposed to post-deployment risks associated with ethical, safety, and logical issues.
Through the AI-powered software history, we have examples of unsupervised learnings that brought bizarre outcomes, such as when a facial recognition system wrongly identified more lawmakers as criminals or when an AI bot learned racism with biased and unchecked data received from bad-minded users.
Shortly, here are the underlying reasons why “unguided” machine learning systems can find themselves in problems, together with their creators.
Raw data without human annotation and/or interaction pose a significant risk to the AI-powered system's relevancy and accuracy.
The quality of decisions is the crucial benefit of machine learning systems. In the AI world, the quality decision is made by collecting the correct ingredients and the correct reading of these ingredients. One of those wrong readings is compute, and wrong ingredients can be data. When reaching the scale, as one of the main drivers of automation, a critical issue that consequently appears is the quality of decisions. An insufficient or incorrect dataset drives a huge risk compared to incorrect human insight.
Therefore, the process of producing annotated ingredients by combining raw data and human insights is fundamental while designing and developing AI-based services and platforms.
That´s why synergy between AI technology and a human guidance is indispensable
The necessity for synergy between AI technology and human guidance comes especially in focus in AI-powered use-case solutions.
Here, a blend of data source competence, data science competence, and data engineering, together with deployment scenarios and capabilities to handle post-deployment streams, represent the receipt to get the maximum from an AI-powered solution.
The power of AI-powered solutions comes from the mix of knowledge gained from:
- engineering of machine learning applications and solutions per use-case deployment scenario tracking and acquiring factors that matter.
- experiences and understandings gained from the network optimization during and post-deployment of use-cases using AI.
- working hand in hand with customers to navigate the AI adoption journey and ability to observe, discern and adjust AI models periodically.
Use cases are among the most distinctive examples of the necessity for synergy between AI technology and human guidance.
As gaps and deviations in operating AI model driven use cases are understood, changes are continuously being made in the AI components embodied in the solution architecture runway, with the aim to meet the specific situations or, to be more precise – specific use cases.
AI model driven use cases – what and how
When considering the adopters of AI model driven use cases, financial, retail, healthcare, and manufacturing industries top the list. Each of the listed industries has one area where the level of AI adoption is the highest, with the perspective to spread to more areas, as shown in the 451 Research - New Survey Sheds Light on Top Use Cases for AI and Machine Learning in Key Industries.
In the financial industry, machine learning is used primarily for fraud detection, followed by another security-centered use case – digital and data security. In retail, customer engagement is the most popular use case. In healthcare, the highest on the list of surveyed use cases using machine learning is patient monitoring systems, and the next is clinician workflow optimization.Maintenance forecasting is the most popular use case for machine learning in manufacturing.
In the telecom industry, improving customer satisfaction and reducing time to market are stated as the main reasons for pursuing zero touch operations among CSPs interviewed by Omdia (report How CSP network executives are overcoming barriers to ZTO).
As the transformation of industries, new requirements from IoT, and new use cases require intelligent networks, Ericsson is on track, developing AI-driven solutions targeting these specific use cases. Our efforts in AI address new network complexities, increase network performance, and enable its automation.
Service Continuity AI app suite, recently added to our Network Support Services portfolio, applies AI and machine learning to provide predictive and preemptive support for each specific use of the network. Intelligent algorithms adjust and adapt on the go to suit target use cases.
How does it work?
The ML model implemented in each use case is a combination of offline and real-time AI model using unsupervised learning methods. When the model is pushed for the first time, it’s a pre-trained offline model driven from the development to the production environment, which is then trained online.
A model developer uses the development environment to write and register the code. Two tasks are carried out in the staging environment - online inference and retraining. In the production environment, the user can request for prediction using the endpoint exposed by the application. Users may monitor the staging model and production model for their performance. When this happens, the user further increases the quality of the ingredients in the staging environment to facilitate even greater accuracy in attaining a more valuable result. There are also insistences of discovery by eliminating variables that have construed the results into different results. In all cases, AI and humans are both learning in the process, working together for the desired outcomes.
If the user concludes that the staging model is better than the production model, an option is provided in the user interface to replace the production model with the staging model. This user interface also allows users to monitor inference data using an ML management dashboard. In production, the outcome generated by the new input dataset is stored for retraining purposes. The online ML retraining is then triggered periodically or on demand.
No matter what blockbusters tell you - AI intelligence will work along with human intelligence!
All AI-powered systems of the future are going to be heavily reliant on human insights. Even if we have incredible advancements in technology, we will have to oversee these systems to make sure that they are built to fit our expectations and produce results that we expect and that make sense.
We developed a suite with more than 200 AI apps to make preemptive and predictive network support a reality. Learn more about how our network support services work, powered by artificial intelligence and machine learning, and guided by human intelligence.
Read more:
Network Support Services powered by AI and ML - Ericsson
Explore telecom AI
Explore AI in networks
