The power of active learning: AI models that self-diagnose and self-improve
There’s a common saying in machine learning and statistics that “All models are wrong, but some are useful”. The phrase is an acknowledgment that while no model can ever be perfect they can still hold tremendous value. The goal then of good machine learning is to constantly strive for better. After all, the real world is not static nor 100% consistent. Tweaking and adjustments are required to ensure models continue to perform.
Take construction sites. They change over time, operate in different seasons and are crawling with a variety of machinery. Anomalies are everywhere. In the case of AI vision, models need to be adaptable or else they will quickly find their predictions more wrong than right. A non-negotiable in high-risk environments like construction.
What if an AI model could identify when it was wrong in order to make itself more useful? In short, it’s game-changing.
Active learning in action
At Presien, we use an active learning system to continuously improve our AI vision models. First things first, AI vision allows a machine to see something, understand it and take action. AI can be trained to detect any object you like - from vehicles to people.
So where does active learning come in? Our technology reviews real-world images provided by in-field devices (Blindsight Processing Units) and filters them for ones that cause uncertain and inaccurate detections. This data is particularly useful for training. In fact, it’s the fuel that optimises Presien’s AI vision models. Thanks to a constant stream of data from our fleet of processing units, the improvements are continuous. The active learning system fixes immediate issues the model encounters plus generalises the learnings to a wide set of conditions the model might encounter in the future. Win, win.
But how does a model know when it is wrong? One way is to test the impact that slight modifications to the input images have on the model’s predictions. Here at Presien, we employ tools that hunt for uncertainties in our model. For example, purposely changing the rotation or colour of an image to see how the system responds. If it detects the same thing every time we know the model is doing its job. If not we can use the result of this and other related techniques to identify types of data that challenge our model.
Additionally, by comparing the outputs of our efficient edge model with larger, more resource-hungry models we can hone in on sources of uncertainty. Combining all these approaches allows us to gain insights into how well the model is performing on a particular image. In doing so, we can identify images that are potentially causing an error or may be novel. This regular pressure testing and iteration is essential to produce a robust model that can keep up with industry demands.
Automating the process
The real potential of active learning is unleashed when the process is automated. This allows the process to run continuously in the background, iterating through loops of error diagnosis toward solutions. This feedback flow, from production data to training data, allows the model to constantly improve and adapt to a changing world with minimal user intervention.
Presien’s active learning system is currently set up to capture data from across our fleet of devices. This allows Blindsight Processing Units to be automatically updated in the field with enhanced models, care of our Active Learning system. These ongoing enhancements will continue to aid heavy industry in their efforts to operate efficiently and safely.