Machine Learning models are everywhere – in the voice recognition feature on our phones, image understanding in the text-to-speech option, in our movie and music recommendations, virtual chatbots and so many more areas.
They make decisions almost every minute that impact our lives greatly.
But one of the downsides of allowing a machine to make complex decisions depends upon the historical data on which it is trained. They arrived at that conclusion.
The gap between what the machine learning model predicts, and a human being’s understanding is getting minimum. Sometimes machines are promoted to do a better job.
This is what interpretable machine learning models aim to do – bridge that gap and make machine learning results understandable to humans.
An introduction to Machine Learning:
Machine Learning is a machine’s ability to make use of past data to make predictions.
An example is the recommended products list on any online shopping platform. Using past search data, the machine learning algorithms bring up several products similar to what you had searched or purchased, assuming that you would be interested in them.
Employing machine learning methods to automate any function is a smart decision because ML takes much less time to train.
ML algorithms also have higher replicability, reliability and speed compared to the human mind. But they are also buried in several layers of complex hidden models that we humans just cannot decipher.
This is a major disadvantage especially if one wants to understand the reasoning behind a certain prediction or output.
What is interpretability?
Simply put, interpretability means presenting or explaining something in a manner that is understandable to a human being. [Finale Doshi-Valez].
Data scientists also call the Interpretability of a machine as human-interpretable interpretations (HII).
In his book Interpretable Machine Learning, Christopher Molnar defines Interpretable Machine Learning as “methods and models that make behavior and predictions of machine learning systems understandable to humans.”
Methods to interpret ML models
Complex functions demand higher accuracy and require complex algorithm structures, like neural networks. These deep neural networks aren’t exactly interpretable because of several hidden layers of millions of neurons.
It is for this reason that data scientists sometimes call them ‘black boxes'.
To interpret any ML model, we need to know:
- what features the system considers most important,
- what effect each feature in the dataset has on a single prediction by the model, and
- the effect of each feature on all the possible predictions.
There are many methods that we can use to gain insight into these questions and help us interpret ML models.
- Model-agnostic methods can be applied to a range of ML models. They function by analyzing feature inputs and output pairs and help us understand a system regardless of the algorithm it uses.
- Model-specific methods differ from model to model and depend on their features and coefficients.
To achieve transparency and make the model interpretable for even those not well-versed in data science, agnostic global and local methods help.
For example, the Local Interpretable Model-Agnostic Explanation (LIME) explains why the model arrived at a specific decision. The Global Surrogate method, on the other hand, explains the logic behind how the model functions as a whole to make predictions.
Both these methods are useful to explain the scope of the model interpretation.
Other model-agnostic methods include Partial Dependence Plots (PDPs) and SHAP values.
Partial Dependence Plots focus on understanding the marginal impact one or two features will have on a model’s prediction.
SHAP values or Shapley Additive explanation is derived from game theory and aims to understand how much of an impact each feature has on a predicted outcome.
It assumes that each feature is a player in the game.
Then, it assesses the contribution or impact of each of the players by adding or removing a player from each of the subsets.
Why do we need interpretable ML models?
Machine learning models are interpretable if, as a programmer or data scientist, you can understand them without any extra aid. People need explanations for results; we need to understand how something happened, even if it was a positive result.
Imagine that a public surveillance system is connected to a system running on a machine learning model. The ML model, for whatever reason, identifies members of only a certain community as probable perpetrators of crime based on their physical and facial features.
The real-world implications of this are massive.
This kind of decision, if followed through blindly by people, could lead to wrongful and harmful assumptions made about that community, ultimately leading to discrimination and unfairness.
Similarly, in 2018, Amazon came under fire for employing an AI tool run on machine learning that automatically rejected female applicants for its developer positions, as reported by Reuters.
When they dug deep into why this happened, they found something very interesting.
It was discovered that the system responsible for filtering candidates was trained primarily on data from male candidates’ resumes. It also flagged the word “women” and viewed them as undesirable. This was discriminatory and unfair.
Now, if there was no way to find out how this decision was arrived at and if nobody questioned the unfair decision being churned out by the machine learning model, then we would remain with grossly inaccurate data without knowing how we can fix or change it.
Whatever decisions or predictions machines make will be just a set of algorithms for them. Because, no matter what, machines do not possess the real-world insight and emotions that we do.
How do we know we can trust the machine’s decisions?
This is why we need interpretability- for fairness, transparency, and accountability of the model.
Can we avoid these inaccurate decisions?
We can prevent such gross mishaps from occurring because we can carefully train these trustable and reliable models ourselves.
Areas where we need interpretable machine learning models
Interpretation is especially important in business, law and finance, where we need to clearly outline the why’s and how’s of each step of a process.
Even in industries like medicine, healthcare, and retail, it is crucial to understand why the model has presented a solution.
Even though it might be tricky to exactly learn why, learning how is important in such high-stakes industries because, when applied to real-life situations, there can be no room for errors or bias.
Read also: The Difference Between Deep Learning And Machine Learning | Insights - Tooliqa
Tooliqa specializes in AI, Computer Vision and Deep Technology to help businesses simplify and automate their processes with our strong team of experts across various domains.
Want to know more on how AI can result in business process improvement? Let our experts guide you.
Reach out to us at business@tooli.qa.
.jpg)
