Machine learning has been proven to be very useful, but it is also easy to assume that it can solve all problems and is applicable in all situations. Like other tools, machine learning is useful in specific domains, especially those problems you always know exist but can never hire enough people to solve; or those with clear goals but no clear methods to achieve them.
Nonetheless, every business organization can leverage machine learning in some way. In a recent survey by Accenture, 42% of executives stated that they expect AI to support almost all innovation projects by 2021. However, if you can achieve better results by understanding what machine learning can and cannot do, you can avoid common misconceptions.
01
Myth 1: Machine Learning Is Artificial Intelligence
Machine learning and artificial intelligence are often used synonymously. Machine learning is one of the most successful technologies to emerge from research labs into the real world, while artificial intelligence is a broad field that covers areas such as computer vision, robotics, and natural language processing, as well as methods like constraint satisfaction that do not include machine learning. Artificial intelligence can be seen as any method that makes machines smarter. None of this is the kind of ‘artificial intelligence’ that some people worry will compete with or even attack humans.
Be cautious with various buzzwords and strive for precision. Machine learning is about learning patterns and predicting outcomes from large datasets; its results may appear ‘smart’, but at its core, it applies statistics at an unprecedented speed and scale.
02
Myth 2: All Data Is Useful
To do machine learning, you need data, but not all data is suitable for machine learning. To train a system, you need representative data that covers the patterns and outcomes the machine learning system needs to handle. The data you need should not include irrelevant patterns (e.g., photos showing all standing men and all seated women, or all vehicles in a garage and all bicycles in a muddy area), because the machine learning model you create will reflect those overly specific patterns found in your data center. All data used for training needs to be labeled, with features that match the questions you are asking the machine learning system, which requires a lot of work.
Do not assume that you already have clean, clear, representative, or easy-to-label data.
03
Myth 3: You Always Need a Lot of Data
Recent significant advancements in image recognition, machine reading comprehension, language translation, and other areas are mainly due to the availability of better tools, computing hardware like GPUs that can process large amounts of data in parallel, and large labeled datasets like ImageNet and Stanford Question Answering Dataset. However, because of a technique called transfer learning, you do not always need a large amount of data to achieve good results in a specific domain; instead, you can train a machine learning system to learn from a large dataset and then transfer that knowledge to your own small training dataset. This is how Salesforce and Microsoft Azure’s Custom Vision API works: you only need 30-50 images to show the content you want to classify to achieve good results.
Transfer learning allows you to customize pre-trained systems for your problems using relatively little data.
04
Myth 4: Anyone Can Build a Machine Learning System
There are many open-source tools and frameworks for machine learning, as well as countless courses teaching you how to use machine learning. However, machine learning remains a proprietary technology; you need to know how to prepare data and partition, train, and test it. You need to know how to select the best algorithms and which heuristics to use, and how to translate that into a reliable production system. You also need to monitor the system to ensure that the results remain relevant over time; regardless of market changes or whether your machine learning system is good enough, you will eventually encounter different customer groups, and you need to continuously check whether the model aligns with your questions.
Effectively leveraging machine learning requires experience; if you are just starting, you can leverage APIs that allow you to call pre-trained models from code while hiring data science experts and machine learning specialists to build custom systems.
05
Myth 5: All Patterns in Data Are Useful
Patients with asthma, chest pain, or heart disease, and any elderly person over 100 years old have a survival rate from pneumonia that is much higher than expected. Indeed, a simple machine learning system designed to automatically send hospitalization notifications could notify them to go home (a rule-based system trained on the same data as a neural network). The reason for such a high survival rate is that pneumonia is very dangerous, and patients are immediately sent to the hospital.
This system saw an effective pattern from the data; however, this is not a useful pattern for selecting who needs to be hospitalized (but it can help insurance companies predict treatment costs). More dangerously, you do not know that your dataset contains such useless counter-patterns unless you already know of their existence.
In other cases, a system can learn an effective pattern (like a controversial facial recognition system that can accurately predict sexual orientation from selfies) because it lacks a clear and obvious explanation, making it useless (in this case, the photos show some social cues, such as posing for the photo, rather than displaying innate characteristics).
“Black box” models are effective, but they do not clarify what patterns they have learned. More transparent algorithms, like generalized additive models, can help us understand what the model has learned, allowing us to decide whether it is suitable for deployment.
06
Myth 6: Reinforcement Learning Is Always Available
Almost all machine learning systems in use today employ supervised learning; in most cases, the systems are trained on clearly labeled datasets with human involvement in preparing these datasets. Organizing these datasets requires time and effort, which is why there is great interest in non-supervised forms of learning, especially reinforcement learning (RL)—which refers to learners learning through trial and error by interacting with the environment and being rewarded for predicting correct actions. DeepMind’s AlphaGo system combined reinforcement learning with supervised learning to defeat top Go players, while the Libratus system built by a Carnegie Mellon University team combined reinforcement learning with two other AI techniques to beat the world’s top poker players in no-limit Texas Hold’em. Researchers are experimenting with applying reinforcement learning to various aspects, from robotics to testing secure software.
However, outside of the lab, reinforcement learning is not common. Google DeepMind reduced power consumption in data centers by learning to cool more effectively; Microsoft employed a specific and limited version of reinforcement learning called Contextual Bandit to show personalized news headlines to visitors of MSN.com. The problem is that very few real-world environments have rewards and immediate feedback that can be easily discovered, especially when agents take multiple actions before anything happens.
07
Myth 7: Machine Learning Is Free of Bias
Because machine learning learns from data, it will replicate any biases present in the dataset. Searching for images of CEOs may yield photos of white male CEOs, as more CEOs are white males. However, it turns out that machine learning can also amplify biases.
The COCO dataset, often used to train image recognition systems, contains photos of both men and women; however, more female images appear next to kitchen appliances, while more male images are associated with computer keyboards, mice, or tennis rackets and skis. Training a system on COCO will bind men more closely to computer hardware.
A machine learning system can also impose biases on another machine learning system. Training a machine learning system using popular frameworks with vectors that express relationships between words can learn stereotypes like “men are to women as computer programmers are to housewives” or “doctors are to nurses as bosses are to receptionists.” If you use this system, it may translate he and she (in English) into gender-neutral pronouns in languages like Finnish or Turkish, turning “they are doctors” into “he is a doctor” and “they are nurses” into “she is a nurse.”
While getting similar recommendations on shopping sites can be useful, it can lead to issues when it involves sensitive matters and can create feedback loops; for example, if you join an anti-vaccine group on Facebook, Facebook’s recommendation engine may suggest other groups focused on conspiracy theories or that believe the Earth is flat.
Understanding the bias issue in machine learning is important. If you cannot eliminate biases in the training dataset, you can use some techniques, such as normalizing gender associations between word pairs, to reduce bias or adding irrelevant entries to recommendations to avoid ‘filter bubbles.’
08
Myth 8: There Is Only a Positive Side to Using Machine Learning
Machine learning provides powerful capabilities for antivirus tools, further looking into new attack behaviors to discover them as quickly as possible. However, hackers are also leveraging machine learning to study defenses against antivirus tools by analyzing vast amounts of public data or previous successful phishing incidents to launch large-scale targeted phishing attacks.
09
Myth 9: Machine Learning Will Replace Humans
People often worry that artificial intelligence will take away human jobs and will certainly change the way we work; machine learning systems can improve efficiency and compliance while reducing costs. In the long run, AI will create new roles in business and eliminate some current positions. However, many tasks automated by machine learning were previously unimaginable in terms of complexity or scale; for instance, you cannot hire enough people to review every image posted on social media to see if it contains your company’s brand features.
Now, machine learning has begun to create new business opportunities, such as improving customer experience through predictive maintenance and providing recommendations and support to business decision-makers. Like previous generations of automation, machine learning allows employees to leverage their expertise and creativity.
