{"id":22509,"date":"2020-12-17T10:17:12","date_gmt":"2020-12-17T10:17:12","guid":{"rendered":"https:\/\/www.experfy.com\/blog\/limitations-of-machine-learning\/"},"modified":"2023-09-21T16:31:02","modified_gmt":"2023-09-21T16:31:02","slug":"limitations-of-machine-learning","status":"publish","type":"post","link":"https:\/\/www.experfy.com\/blog\/ai-ml\/limitations-of-machine-learning\/","title":{"rendered":"The Limitations of Machine Learning"},"content":{"rendered":"\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Machine learning is now seen as a silver bullet for solving all problems, but sometimes it is not the answer.<\/em><\/p>\n\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

\u201cIf a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future.\u201d<\/p>

\u2014 Andrew Ng<\/em><\/strong><\/p><\/blockquote>\n\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t

\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Most people reading this are likely familiar with machine learning and the relevant algorithms used to classify or predict outcomes based on data. However, it is important to understand that machine learning is not the answer to all problems. Given the usefulness of machine learning, it can be hard to accept that sometimes it is not the best solution to a problem.<\/p>\n\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t

In this article, I aim to convince the reader that there are times when machine learning is the right solution, and times when it is the wrong solution.<\/h2>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Machine learning, a subset of artificial intelligence, has revolutionalized the world as we know it in the past decade. The information explosion has resulted in the collection of massive amounts of data, especially by large companies such as Facebook and Google. This amount of data, coupled with the rapid development of processor power and computer parallelization, has now made it possible to obtain and study huge amounts of data with relative ease.<\/p>\n

Nowadays, hyperbole about machine learning and artificial intelligence is ubiquitous. This is perhaps rightly so, given the potential for this field is massive. The number of AI consulting agencies has soared in the past few years, and, according to a report from\u00a0Indeed<\/a>, the number of jobs related to AI ballooned by 100% between 2015 and 2018.<\/p>\n

As of December 2018,\u00a0Forbes<\/a>\u00a0found that 47% of business had at least one AI capability in their business process, and a report by\u00a0Deloitte<\/a>\u00a0projects that a penetration rate of enterprise software with AI built-in, and cloud-based AI development services, will reach an estimated 87 and 83 percent respectively. These numbers are impressive \u2014 if you are planning to change careers anytime soon, AI seems like a pretty good bet.<\/p>\n\n

So it all seems great right? Companies are happy and, presumably, consumers are also happy \u2014 otherwise, the companies would not be using AI.<\/p>\n\n

It is great, and I am a huge fan of machine learning and AI. However, there are times when using machine learning is just unnecessary, does not make sense, and other times when its implementation can get you into difficulties.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t

Limitation 1 \u2014 Ethics<\/h3>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Machine learning, a subset of artificial intelligence, has revolutionalized the world as we know it in the past decade. The information explosion has resulted in the collection of massive amounts of data, especially by large companies such as Facebook and Google. This amount of data, coupled with the rapid development of processor power and computer parallelization, has now made it possible to obtain and study huge amounts of data with relative ease.<\/p>\n\n

It is easy to understand why machine learning has had such a profound impact on the world, what is less clear is exactly what its capabilities are, and perhaps more importantly, what its limitations are. Yuval Noah Harari famously coined the term \u2018dataism\u2019, which refers to a putative new stage of civilization we are entering in which we trust algorithms and data more than our own judgment and logic.<\/p>\n\n

Whilst you may find this idea laughable, remember the last time you went on vacation and followed the instructions of a GPS rather than your own judgment on a map \u2014 do you question the judgment of the GPS? People have literally driven into lakes because they blindly followed the instructions from their GPS.<\/p>\n\n

The idea of trusting data and algorithms more than our own judgment has its pros and cons. Obviously, we benefit from these algorithms, otherwise, we wouldn\u2019t be using them in the first place. These algorithms allow us to automate processes by making informed judgments using available data. Sometimes, however, this means replacing someone\u2019s job with an algorithm, which comes with ethical ramifications. Additionally, who do we blame if something goes wrong?<\/p>\n\n

The most commonly discussed case currently is self-driving cars \u2014 how do we choose how the vehicle should react in the event of a fatal collision? In the future will we have to select which ethical framework we want our self-driving car to follow when we are purchasing the vehicle?<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t

If my self-driving car kills someone on the road, whose fault is it?<\/h4>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Whilst these are all fascinating questions, they are not the main purpose of this article. Clearly, however, machine learning cannot tell us anything about what normative values we should accept, i.e. how we should act in the world in a given situation. As David Hume famously said, one cannot \u2018derive an ought from an is\u2019.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t

Limitation 2 \u2014 Deterministic Problems<\/h3>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

This is a limitation I personally have had to deal with. My field of expertise is environmental science, which relies heavily on computational modeling and using sensors\/IoT devices.<\/p>\n\n

Machine learning is incredibly powerful for sensors and can be used to help calibrate and correct sensors when connected to other sensors measuring environmental variables such as temperature, pressure, and humidity. The correlations between the signals from these sensors can be used to develop self-calibration procedures and this is a hot research topic in my research field of atmospheric chemistry.<\/p>\n\n

However, things get a bit more interesting when it comes to computational modeling.<\/p>\n\n

Running computer models that simulate global weather, emissions from the planet, and transport of these emissions is very computationally expensive. In fact, it is so computationally expensive, that a research-level simulation can take weeks even when running on a supercomputer.<\/p>\n\n

Good examples of this are MM5 and WRF, which are numerical weather prediction models that are used for climate research and for giving you weather forecasts on the morning news. Wonder what weather forecasters do all day? Run and study these models.<\/p>\n\n

Running weather models is fine, but now that we have machine learning, can we just use this instead to obtain our weather forecasts? Can we leverage data from satellites, weather stations, and use an elementary predictive algorithm to discern whether it is going to rain tomorrow?<\/p>\n\n

The answer is, surprisingly, yes. If we have knowledge of the air pressures around a certain region, the levels of moisture in the air, wind speeds, and information about neighboring points and their own variables, it becomes possible to train, for example, a neural network. But at what cost?<\/p>\n\n

Using a neural network with a thousand inputs to determine whether it will rain tomorrow in Boston is possible. However, utilizing a neural network misses the entire physics of the weather system.<\/p>\n\n

Machine learning is stochastic, not deterministic.<\/strong><\/mark><\/p>\n\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t

A neural network does not understand Newton\u2019s second law, or that density cannot be negative \u2014 there are no physical constraints.<\/h4>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

However, this may not be a limitation for long. There are multiple researchers looking at adding physical constraints to neural networks and other algorithms so that they can be used for purposes such as this.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t

Limitation 3 \u2014 Data<\/h3>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

This is the most obvious limitation. If you feed a model poorly, then it will only give you poor results. This can manifest itself in two ways: lack of data, and lack of good<\/strong> data.<\/p>\n\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t

Lack of Data<\/h4>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Many machine learning algorithms require large amounts of data before they begin to give useful results. A good example of this is a neural network. Neural networks are data-eating machines that require copious amounts of training data. The larger the architecture, the more data is needed to produce viable results. Reusing data is a bad idea, and data augmentation is useful to some extent, but having more data is always the preferred solution.<\/p>\n\n

If you can get the data, then use it.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t

Lack of Good Data<\/h4>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Despite the appearance, this is not the same as the above comment. Let\u2019s imagine you think you can cheat by generating ten thousand fake data points to put in your neural network. What happens when you put it in?<\/p>\n\n

It will train itself, and then when you come to test it on an unseen data set, it will not perform well. You had the data but the quality of the data was not up to scratch.<\/p>\n\n

In the same way that having a lack of good features can cause your algorithm to perform poorly, having a lack of good ground truth data can also limit the capabilities of your model. No company is going to implement a machine learning model that performs worse than human-level error.<\/p>\n\n

Similarly, applying a model that was trained on a set of data in one situation may not necessarily apply as well to a second situation. The best example of this I have found so far is in breast cancer prediction.<\/p>\n\n

Mammography databases have a lot of images in them, but they suffer from one problem that has caused significant issues in recent years \u2014 almost all of the x-rays are from white women. This may not sound like a big deal, but actually, black women have been shown to be 42 percent more likely to die from breast cancer<\/a> due to a wide range of factors that may include differences in detection and access to health care. Thus, training an algorithm primarily on white women adversely impacts black women in this case.<\/p>\n\n

What is needed in this specific case is a larger number of x-rays of black patients in the training database, more features relevant to the cause of this 42 percent increased likelihood, and for the algorithm to be more equitable by stratifying the dataset along the relevant axes.<\/p>\n\n

If you are skeptical of this or would like to know more, I recommend you look at this article<\/a>.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t

Limitation 4 \u2014 Misapplication<\/h3>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Related to the second limitation discussed previously, there is purported to be a \u201ccrisis of machine learning in academic research<\/em><\/a>\u201d whereby people blindly use machine learning to try and analyze systems that are either deterministic or stochastic in nature.<\/p>\n\n

For reasons discussed in limitation two, applying machine learning on deterministic systems will succeed, but the algorithm which not be learning the relationship between the two variables, and will not know when it is violating physical laws. We simply gave some inputs and outputs to the system and told it to learn the relationship \u2014 like someone translating word for word out of a dictionary, the algorithm will only appear to have a facile grasp of the underlying physics.<\/p>\n\n

For stochastic (random) systems, things are a little less obvious. The crisis of machine learning for random systems manifests itself in two ways:<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t