Sometimes what you want is not what you need – deep learning explained
I guess most of you have come across the concepts of deep learning and neural networks. Artificial intelligence, especially in terms of deep neural networks, is a very popular subject of conversations about modern technology. Recent years have seen the success of artificial neural networks in a number of disciplines; however; this solution is not bound to be successful in every possible case.
The popularity of deep learning
The notion of neural networks has been commonly known for many years now. Its beginnings date back to the 1960s, when a Soviet mathematician Alexey Ivakhnenko developed the Group Method of Data Handling. Despite the recent drop, the interest in deep learning is still immense. Google search results for “deep learning” have shown a growing interest in the subject, especially since 2014.
The main factors that influence the development of deep learning include the continuously growing computing power and accessibility of a lot of large sets of data that can be analysed. It has also been fuelled by the rise in popularity of AIaaS (Artificial Intelligence as a Service) in business.
Over-hyped deep neural networks
In fact, deep neural networks have been definitely over-hyped recently. This is confirmed by Gartner Hype Cycle for Emerging Technologies. The 2020 Hype Cycle for Artificial Intelligence indicates that deep learning is at the stage of “Peak of Inflated Expectations”.
This tendency is also evidenced in the number of searches of “deep learning” in Google Trends mentioned above. This, in turn, may be caused by the current media trend, where the number of views and the emotional value of the news is more important than reliable presentation of information.
An article published in The Guardian in 2020 created a great sensation around AI. The article titled “A robot wrote this entire article. Are you scared yet, human?” was written entirely by GPT-3 language generator, according to the editors. The quality of the text and the way it was narrated spurred a series of comments that expressed fear of artificial intelligence.
Readers felt alarmed because the language used by the robot was as good as the language used by humans and the content of the piece was a deceitful means of winning the readers over. The article was not written entirely by AI, though. The GPT-3 model generated eight different texts; next, The Guardian selected the most interesting excerpts, combined them, and edited them to form the final version. The article misled the readers and exploited human fears concerning AI in order to make money. Articles of this type do nothing but generate over-hype and lead to misinformation.
The debate on AI is also whipped up by advertising catchphrases created by AI-selling companies that lure their clients with “space flights”. Both the over-hyped media coverage and the sellers that overpraise their products heighten the clients’ expectations to such an extent that they start planning doing their projects on Mars. In result, data scientists have either to meet the users’ demands or to get them back down to Earth.
A trendy watch
One of the outcomes of the growing popularity of deep learning and artificial intelligence discussed here is that all these issues have been turned into catchy buzzwords. AI has become a must when it comes to functionalities offered by a product. Most of the large technology companies work on smart solutions and they manage to prove the effectiveness of AI-based products.
As a result, AI, neural networks, and machine learning are now like a smartwatch that everyone wants to have. A misconception arose, according to which AI-based products are better, and promoting them with the phrase “AI-powered” is supposed to increase their popularity and sales. And yet, there is nothing to worry about in this respect: over-hyped descriptions of technological products are a common occurrence.
Killing a fly with a cannon
Numerous ML models have demonstrated their usability. Hence, it comes as no surprise that people who are not really tech-savvy choose the methods that have already been successful. The problem is that inexperienced users often end up striving to have specific ML models.
Every method of machine learning has its use and a specific task it is meant to solve. Currently, there is no universal model that could deal with every possible problem. This is why data scientists need to select solutions carefully to match particular tasks and their requirements: available data and resources, the complexity of the problem and the algorithms, and the model interpretability.
It is rather understandable that advanced ML methods are very popular: especially if a company can boast having used them as the first one in a given sector. Then again, chasing technology trends may lead to the disproportionate use of resources in relation to the problem at hand.
The use of popular neural networks to solve every single problem may be like killing a fly with a cannon. Solutions based on neural networks require much more computing power than traditional methods, which causes larger expenses.
You don’t need an AI model – you need a solution to your problem
Users don’t need an algorithm as a product; what they actually need is to get rid of a bothersome problem. The goal of ML-based solutions is to identify the user’s pain points, to present the problem as a measurable function, and then to eliminate the pain point to make sure the user is fully satisfied.
AI-powered products look attractive to the clients, but the clients’ priority is to meet their goals. Some users need to solve an everyday problem to save time for more pressing matters; other users wish to show off with a fashionable product. If you call a plumber because your tap is dripping, you don’t care what tools and materials the plumber is going to use to fix it: what you want is for the tap to stop dripping and never do it again.
100% effective model
Due to the misleading ways of presenting information in the media, the public is often wrong about AI. People believe that AI models must be infallible to be usable. The clients who want to implement smart solutions search for models that are nearly 100% effective.
Treating effectiveness as the main factor (or even the only factor) that determines the success of a given solution is not really the best one. The effectiveness of a model is a manner of approximating the correctness of its functioning. It is like a test that students take at school. Seeking unerring models is like striving blindly to get 100% points in every school test. When a learned model is tested in a real-life environment, the results may be devastating.
The key factor that should be taken into account is how the client’s objectives correspond to the model’s objectives. Developing the right model is all about balancing effectiveness with business interpretability and the required outlays. To bring all of these issues together, it is necessary to talk to the client to evaluate the situation and reach a compromise.
Even if the model’s effectiveness is much lower than 100%, it can be still useful. A solution is sufficient if it releases humans from some part of their duties or recognises and detects danger quicker than humans. Anti-spam software for email is a great example of this kind of solution. Sometimes spam does get through to your inbox, but the solution can still be deemed effective because it checks every message for you.
This might not seem obvious, but sometimes less functional models turn out to be more successful than great models with the highest effectiveness rates. The solution that won the Netflix Prize may serve as a good example here. The winning solution was never used because it was too complex. The complexity meant high implementation costs.
A great deal of myths and expectations surround the subject of artificial intelligence. The media exploit the appeal of this subject by boosting the hype around it: both positive and negative. Thus, data scientists need to educate the clients in the right way. People who do not know much about AI should be able to trust experts. The work of data scientists is based on constant cooperation with the clients aimed at developing the right solution. The goal of designers is to put themselves in the clients’ shoes and to identify the major problems.
Once they know exactly what the problem is, they can define the goal clearly and in a measurable form, select suitable tools, and, consequently, eliminate the pain point in an effective way. For every minute that is taken to design a ML model, an hour should be spent analysing the client’s problem. After all, the main goal of implementing smart solutions is to solve a problem successfully. If the problem can be solved by means of simple ML models or just by data analysis, these methods should be used to spare the client’s time and money.
Data scientists do deserve your trust – and if you have any doubts, let them know. Creation of smart solutions involves dialogue and data analysis. Specialists are here not only to deliver solutions but also to educate the clients about AI by explaining the way algorithms work and presenting their possible applications.
Thanks to close cooperation with data scientists, the clients can achieve their goals and prevent unnecessary losses.