A little over a year ago, when I wrote about Predictive Analytics for Man and Machine, I discussed how predictive analysis of sensor data from humans and machines can help create wellness-driven systems for people’s health and predictive maintenance for machines, respectively. We also wrote a book on the internet of things (IoT), in which we discussed at length how the return on investment and business models of such systems depend on the promise of going towards preventive care systems from the current reactive ones.
In this period, we have seen tremendous development in the automation of analytics through artificial intelligence (AI). DeepMind, whose AlphaGo system beat a human professional in the game of Go using AI techniques, recently introduced an advanced version that learns the game by playing it on its own (without referring to any previous human games), making it more ‘human-like’. We have also seen advances in deep learning and deep reinforcement learning techniques that can automatically learn, recognize, and classify images, text, and speech with great accuracy. All this has led to impactful applications in computer vision, medical imaging analytics, and knowledge mining from text and speech recognition.
Now, the next natural question is, how much of this advancement in AI can be applied to sensor data analytics for predictive maintenance and preventive wellness systems? For example, can it reliably predict when a machine is going to fail by looking at, say, data from a vibration sensor mounted on the machine? Or can it reliably screen for early onset of heart problems from ECG sensor data, so that measures can be prescribed to prevent further deterioration, possibly even bringing about improvement? The answer, as it turns out, is not that simple. The current AI developments are focused on two key issues that impede its wider application, especially in critical areas like healthcare and predictive maintenance.
AI Systems Need Data with an Underlying Structure
There is a fundamental difference between AI systems that learn the Go game or recognize faces from images or diagnose disease from x-ray images, and systems that need to diagnose machine failure from vibration data or need to predict heart diseases from ECG data. The former all work on human-generated data, which has an inherent structure; this is missing in the latter scenarios. In other words, the former systems are driven by a language structure understandable by humans, while the latter systems don’t.
As pointed out by Josh Tenenbaum, professor of cognitive science and computation at MIT, “There’s no way you can have an AI system that’s human-like that doesn’t have language at the heart of it. It’s one of the most obvious things that set human intelligence apart.” The Go game, images, speech – all have this inherent structure that is referred to as ‘language’ – essentially because they were created by human beings. But when we deal with raw sensor data generated naturally by a machine or human body, such structure may not be obvious, thereby making it difficult for deep learning systems to derive value from it.
It has already been seen that deep learning systems are good at predicting anomalies from sensor data once they have learnt the ‘good’ models from previous data (any deviation from it is then triggered as an anomaly). However, they are not yet good at creating prognostic/diagnostic systems because of the absence of the underlying ‘language’ or ‘structure’ that characterizes a malfunction or a disease.
Interpretability of Artificial Intelligence
The other issue with such predictive diagnostic systems is that they are quite critical in nature, and have large costs associated with wrong diagnoses (the criticality of predictive health is obvious, as is the criticality of a factory machine going down due to a wrong prognosis). This makes it imperative that the models used/learnt by the AI systems are physically interpretable from the current scientific knowledge base.
For example, models/features used in predicting machine failure from vibration sensor data should be interpretable and explainable using laws of physics, whereas models/features used in predicting heart problems from ECG sensor data should use knowledge of biology and medical sciences. Otherwise, even if they produce good results, it will be difficult to certify and deploy them in different scenarios.
As has been highlighted in the recent ICML workshop on Human Interpretability in Machine Learning, “Supervised machine learning models boast of remarkable predictive capabilities. But can you trust your model? Will it work in deployment? What else can it tell you about the world? We want models to be not only good, but interpretable. And yet, the task of interpretation appears underspecified.”
We need to understand that without interpretability, we cannot bring in ownership and liability of wrong diagnosis/prognosis, which is extremely important in the given use cases of human and machine health. Nor can we build systems that are certified by domain experts through multiple trials and experiments. A very interesting blog lucidly explains this problem of interpretability in deep learning systems.
So, what is the way forward? How do we create deep learning systems that can automatically discover structure in seemingly unstructured natural sensor data, or offer physically interpretable models? Perhaps, as Carlos E. Perez writes in his blog, the only way to make deep learning interpretable is to have it explain itself. That is likely to warrant another post.