This is not an AI: contextual classification with Brainvivo's Digitized Brain
Look at the image above.
Your brain tells you that this is a banana. You can imagine its taste, its smell. You can do all this because your brain has been wired through evolution and life experience to identify this object, cluster it as a food, and deduce characteristics of this banana like its smell and taste.
Conventional artificial intelligence (AI) models can learn to recognize bananas, but only if they are trained on thousands of images of bananas from a huge variety of sizes, colors, camera angles, and backgrounds. Even then, these models cannot infer any other information about the banana like its texture or its ripeness - they only recognize its visual properties.
The human brain is the most sophisticated trained network in existence, and can accomplish the same and so much more after seeing only a few bananas. Imagine a technology that could capture your brain's response to that image of a banana, or any image at all, and generate real human insights!
Brainvivo is on a journey of digitizing the brain and building a groundbreaking humanized AI platform that mimics how the human brain processes data. Based on over 20 years of neuroscience research, Brainvivo’s cutting-edge technology unlocks the sophisticated biological network of the brain to automate tasks that require human decision-making, emotional insight, judgment, and intuition. Using its proprietary MRI technologies, Brainvivo has digitized the brain into a humanized computational model.
Click here to see how our technology works!
Brainvivo ran an experiment that succeeded in digitizing the human understanding of visual data by asking people to look at images while they were being scanned in the MRI. Participants were exposed to a very small dataset of 36 images from 10 distinct categories. Using their proprietary technology, the Brainvivo team captured the real brain responses to these images. From these responses, Brainvivo built a model called the 'Digitized Brain' that mimics the way the brain processes and interprets visual data. The Digitized Brain was tested on 180 new and unseen images and for each image it generated a simulated brain response, before finally clustering those responses into the 10 different categories.
The diagram above shows how the Digitized Brain predicted the label for each image. What you see are the probabilities of each label for each image, represented by bars coloured for each category - for example, the 'dog' label is red and the 'butterfly' label is yellow.
The model was 100% accurate in predicting the category of each image, using a fraction of the training data required even by state-of-the-art AI models!
Most importantly, the Digitized Brain revealed similarities between different categories of the same type. For example, images of Zebras showed high probability for the 'zebra' label but also for the 'dog' and 'butterfly' labels, suggesting the model recognized all of these labels as animals. Similar probabilities can be found between labels that belong in the same higher class!
The Digitized Brain doesn't just label images... it understands them.
To verify the model’s ability to understand the context of images, Brainvivo ran another experiment in which they took seven categories, as seen below. The Digitized Brain was trained on 60 images per category, and tested with 20 images per category. For good measure and validation, the same data was used to train and test three classical machine learning models: CNN, KNN, and Random Forest.
As you can see above, Brainvivo’s Digitized Brain achieved 80% accuracy, almost 2x more accurate than the next best machine learning model given such a small dataset. Beside surpassing the classical machine learning models in accuracy, the Digitized Brain also understood the context of the images.
Though the Digitized Brain was trained and tested on 7 categories, Brainvivo found it clustered these categories together into three higher classes - Faces, Foods, and Landscapes - without them being explicitly defined! Even if the Digitized Brain made mistakes predicting the category of the image, it was still correct in predicting its class; the model may have mistaken a steak for a waffle, however it still recognized it as food. These results suggest the Digitized Brain is not only a powerful classifier but is actually capable of understanding the context of images!
By digitizing the human brain, Brainvivo is digitizing the most sophisticated trained network out there and making it accessible. One of the uses of this incredible technology is training complex classification models which don’t just classify the objects - they capture the hidden context of objects and scenes, just like humans. The ability to understand context as humans is crucial to make decisions and judgment in our day to day lives.