AI systems once required a plethora of inputs. New techniques are filling the gaps. 

“It can’t be done.” Previously, this might have been the response from engineers when asked to build an effective model from a small data set. Sounds sensible, right? Traditional AI methods require a certain amount of quality data to produce a workable model.

But what if a small data set could be amplified? What if we had smarter AI models that could learn from less data?

That’s exactly what the engineers at Presien have achieved. Using cutting-edge techniques they are turning up the volume on small data sets to allow customers to create custom AI models from only a few images and videos.

‍The hunger for big data

Data is the fuel that drives AI vision. Collecting high-quality, varied and accurately annotated data is par for the course. In practice, this means sourcing thousands of example images and manually annotating them with bounding boxes and class labels.

This data requirement grows even more extreme in less structured environments. For example,  construction sites require an even higher threshold of example imagery in order to be ready to respond accurately to differences in lighting, occlusion, shadows and background changes. Not to mention, entry to sites often comes with a lot of red tape. 

The effect is that many scenarios where AI could make a significant impact are missed. 

‍Overcoming the data burden

If your data is sparse, don’t despair. This hurdle can be lowered through a variety of approaches. 

One way to fill in data gaps is by siphoning learning from another AI model that is pre-trained on a similar domain or application. For those in heavy industry, the likes of Blindsight would be ideal. The product offers seven years of training data plus the ongoing network effect of deployed units. Whatever system you choose,  leveraging an existing dataset of similar images means an AI model can expand its scope, fast.

Other techniques involve leveraging ‘Data Lakes’, augmenting existing example images to inject variance and applying “Few Shot Learning’ techniques. 

Let’s break down a couple of these key ways forward. 

‍AI models that study smarter

AI models can be designed to learn from limited data using an approach called 'Few Shot Learning'.

One solution involves using a more permissive model trained on a larger dataset to suggest potential objects in an image and then training a smaller model to classify those objects using limited labeled data. By training the AI model's hidden representation of each object in a way that better separates useful visual features, the classifier model can better generalize to new examples, improving its performance even with less labeled data.

This method also allows learning in an unsupervised manner, where larger unlabelled datasets can be used to further improve the separation of objects within the latent space. One such source of this data is our Data Lake.

‍Soaking up the benefits of the data lake

By using machine learning techniques such as weak supervision on a large dataset without labels, a user can leverage their limited existing data more effectively during training.

Weak supervision uses various sources of supervision, such as heuristics, rules, or weak labels, to generate training data. This approach can be useful when manual labeling is time-consuming, expensive, or infeasible.

At Presien, our partners are putting this idea to practice and leveraging our Heavy Industry Data Lake. In doing so, they are tapping into an extensive database of heavy industry objects and environments that hold relevancy to the challenge they want to tackle.

‍Spicing up sparse data: augmentation and generation

Ready to make your data do more? Enter augmentation and generation. By visually augmenting existing example images we can inject variance into the data and amplify what already exists.

To achieve this, strategies such as object blending, image generation and computer vision augmentation are employed. Introducing different scenes, varied backgrounds and lighting into an existing image are several common tactics. The incorporated tweaks are specific to variations commonly encountered in heavy industries to drive real-world performance.

At Presien, our Active Learning Pipeline ensures we’re regularly pressure-testing our model by introducing imagery with heightened complexity. For example, one of the reviews involves purposely changing an image e.g. rotation or colour to see how the system responds. If it detects the same object every single time we know the model is doing its job. If not, it gets to work improving itself. 

Our approach has been proven over several years and already exists within Presien’s automated MLOps pipelines.

‍Getting started

By harnessing these new AI techniques, the challenge of sparse data is reduced. By lowering this ‘data barrier’ to entry, customers can access AI faster and easier. It opens the door for more customers to apply our AI technology to what it does best: make an impact.

What custom ML model would you like to create?