We are currently demonstrating quantum artificial intelligence using our 2nd generation quantum processing unit. Our first two use cases are cell membrane wall detection and satellite imaging. Below, we detail how we used quantum coherent noise from our machine to improve neural network training significantly.
1. Cell Membrane Wall Detection
The first example is a small set of 30 consecutive 256×256-pixel monochrome images from an electron microscopy (ssTEM) dataset (ref) of the Drosophila first instar larva ventral nerve cord. Below are two examples of the raw images and the corresponding cell walls, which are the target for our semantic segmentation task.
The images represent real-world images: some noise, image registration errors, and even a small stitching error in one section. None of these would lead to any difficulties in the manual labelling of each element in the image stack by an expert human neuroanatomist. A software application that aims at removing or reducing human operation must be able to cope with all these issues.
We inject quantum coherent noise generated from our 424-quantum dot array structure into the neural network used to detect membrane walls. We employ this enhanced network to make prediction results in improved ice coefficient scores. The images below show a comparison of the UNET model trained with default settings and a model trained with the unitary noise. The most evident comparison can be observed in the last image in Fig. 2 (purple segmentation), where we apply harder thresholds to create sharper imagery.
2. Satellite Imaging
The second use case focuses on how humanitarian organizations could use satellite imagery (such as www.crowdai.org). Following a natural disaster, it would be advantageous to map impassable sections of roads and identify the most damaged residential areas and the most vulnerable schools, hospitals, and public buildings. The objective is to adapt to the situation as quickly as possible to enable intervention procedures as the crisis evolves.
In the first days following such an event, it is essential to have detailed maps of communication networks, housing areas, infrastructure, and areas dedicated to agriculture, etc. Images are available from various sources, including nano-satellites, drones, and conventional high-altitude satellites.
When new maps are required, they are often drawn by hand, often by volunteers who participate in so-called Mapathons. A machine learning approach can automate maps’ production with relevant features in a short timeframe and from disparate data sources.
Our training data contains individual tiles of satellite imagery in RGB format, and labels (color segmentation super-imposed on the images) are used to annotate recognized information. The goal is to train a model, which is given a new tile (satellite image), to annotate all buildings. The typical training dataset has 280,741 tiles (as 300×300-pixel RGB images), so we restrict our test case to 0.5% in the interest of speed.
We form predictions with the trained model to visually see how these improvements come through. The results are illustrated below, where we compare a UNET where we added a Gaussian noise layer (Figure 1) to the performance achieved with the Equal1 unitary noise layer (Figure 2).
These use cases are available on our GitHub for you to explore further. In the meantime, we are busy exploring use cases in medical imaging, materials science, and others. Please contact us to discuss your specific use case.