How can you avoid common Machine Learning myths?

It's important for the stakeholders of any commercial project to have realistic expectations at the outset, given the time and resources available. There can often be a large gap between engineers' and stakeholders' understanding of ML capabilities.

That last sentence is a big one! An ongoing service needs a model that is carefully cultivated and maintained. Only using a static model risks failing to adapt to new patterns or characteristics, but failing to carefully groom new input brings problems as well. For example, using model-generated content in training can lead to a collapse of any worthwhile features of the original model, or malicious users may engineer input to their own ends. It is vitally important to commit to regular upkeep and evaluation of your models. It is not a "one and done" project.

There is no silver bullet. There never was and Machine Learning is no magical solution to all problems either. It's true that ML algorithms have a tradition of being proven to work before being understood mathematically, and Deep Learning models are the worst in this aspect, so it's easy to naively think of it as "magic". But, actually, because of this very reason, building a model that is good, useful and creates value for the company is very hard and involves uncertainty in all steps, from the input data, the model architecture to the evaluation metric. Only an online A/B test can truly tell if it works, so ML is far from magic, it's science through and through.

Well, I partially disagree. There was a time where you needed a PhD and knowledge of calculus, linear algebra, distributed systems, data structures and algorithms, and the list goes on. Nowadays, the tools are very user-friendly, they require some mastery and some coding, granted, but no deep knowledge of what happens under the hood is required anymore. Like Photoshop experts have no clue how most of the tools work, but they know the effect of each of them and know what is the desired outcome. But, because now a one-page tutorial can guide you through building your first toy model, the ML landscape is much more dynamic and broad than it was a decade ago. So, it's easy to start, but it's a whole career to master.

Additionally, the success of machine learning depends on various factors, including data quality, quantity, and relevance. Preprocessing and feature engineering are crucial steps to tailor the data to the specific requirements of the chosen machine learning model. Furthermore, hyperparameter tuning and model selection are necessary to optimize the performance of the model and achieve the desired outcomes. Customization and optimization are essential to ensure that the machine learning solution aligns with the unique characteristics and intricacies of your problem domain.

One common mistake is to go straight to neural networks, when a simple naive bayes classifier (for example) would do the trick. Knowing the more traditional models solves most of the problems at hand. A second regretful mistake is to think most of the work lies in training the model, where in my experience, 70% of the work lies in exploratory data analysis and feature engineering. Then you'll know what kind of model you will need.

Doing data science or building machine learning models is different from traditional software development, it's more akin to empirical science. You make assumptions, create experiments, run them, measure the results, rinse, repeat. In this whole process, a lot of human decisions and interpretations goes on, and (like science) you need a really rigorous and systematic methodology to be the most objective and unbiased as possible. And companies investing in AI and ML need to be aware that the experiments may fail to produce the expected results or, worse, have unexpected outcomes. It's an inherent risk, even with the best of methodologies.

In dispelling the myth that "Machine Learning is Enough," it's essential to tackle another associated misconception: that machine learning entirely eradicates human bias. While machine learning has the potential to reduce certain human biases by automating some aspects of decision-making, it does not completely eliminate them. In fact, machine learning models can inadvertently perpetuate and even intensify existing human biases if the training data they learn from is biased. This underscores the need for vigilance, critical analysis, and continual refinement in machine learning applications, emphasizing that machine learning is a part of the solution, but not the entire solution.

One of the first questions I have for any data-driven problem is the need for an ML solution. Points that you should consider - Complexity of Problems: Machine learning is powerful but not a magic bullet that can solve all problems. - Data Limitations: Machine learning heavily relies on data availability and quality. The effectiveness of machine learning algorithms can be severely limited due to data quality. - Interpretability and Explainability: Machine learning models often operate as black boxes, making it difficult to interpret their decision-making process. - Human Expertise and Intuition: In many domains, human domain experts possess invaluable knowledge and insights that cannot be easily captured by data alone.

One way to illustrate this myth is think of an organization with a strong machine learning and data science team, building and deploying incredible models, for example a recommender system, but the UX is terrible and buggy and the API is sluggish. No amount of data or training or neural layers will make people more likely to click that widget. Every IT discipline, from governance and compliance, through site reliability and quality assurance, and machine learning and front-end development, and everyone in the middle of that multidimensional spectrum have to work together to create a successful product.

Moreover, the deployment and production phase of machine learning models require ongoing monitoring and maintenance to ensure their continued accuracy and effectiveness. This involves regularly retraining models with new data, evaluating their performance, and making necessary adjustments or refinements to keep up with evolving patterns and dynamics in the data and the problem domain. Staying proactive and agile in your approach to machine learning is crucial for maximizing its potential and achieving sustainable results.

Machine Learning is a wonderfully inclusive realm, welcoming individuals from all backgrounds. Regardless of your personal history, it is crucial to infuse your distinct perspective when constructing machine learning models. While crafting these models might seem straightforward, the true challenge lies in interpreting their outcomes, requiring diligent effort. Let's embrace diversity and recognize the significance of our individual contributions in this ever-evolving field of Machine Learning.