In medical imaging assays, artifacts can introduce major biases if not detected before the analysis. For one of the leading manufacturers of medical and Life Sciences instruments, we developed a machine learning system able to detect local outliers on an image based on imperfect training sets, the only assumption being that more regions of the images are valid than not. We thus eliminated the the need for a perfect - or even labeled - truth set, which would not have been obtainable given the sheer number of images.
When dealing with Machine Learning predictors that require certification, it is necessary to provide explanations how decisions are made so that a human certifier may trace and approve the decisions. Such explanations are necessary in certifying most health-related solutions, but also in, for instance, mortgage and other loan application, as well as in understanding a potential scientific discovery. Our approach for explaining classifiers is also applicable to typical black-box systems, such as deep learning networks.
Electroencephalography (EEG) plays an important roles in monitoring the brain activity of patients with epilepsy, Alzheimer’s, and other brain related diseases. While currently an expert is needed to analyze all EEG recordings to detect anomalous activity, we are working with a large national clinical institution in North America to automate the process to pre-screen patients on a large scale, thus reducing response time for each patient, and reducing cost for the health care system.
Every instrument is slightly different, every optical module, and every robot, which in turn means that the analysis software needs to be able to automatically adjust to these individual differences without ever having seen them. For our customer, a leading provider of health care instruments, we developed a machine learning normalization system able to apply one single model to recognize objects in images from very different platforms. We devised this novel strategy working closely with our customer and its manufacturing division.
Auto-immune diseases are among the ailments that are difficult to diagnose. While several markers have been established, biological variation makes it difficult to define single-variate cut-off values, requiring expert medical expertise for taking decisions on treatment. For our customer, one of the largest providers of diagnostics, we implemented a system to monitor changes in patient's health by applying a deep learning algorithm to simultaneously process a multitude of marker values and to issue early warnings if the patient's condition changes.
Hundreds of articles are published every day in a single newspaper, often not labeled appropriately for someone following a specific area of interest. Simple key word searches both produce articles that are not relevant, as well as missing articles that are. By training a model on a moderate data set, we can achieve up to 98% accuracy in classifying newspaper articles as 20 different categories, including politics, international, sports, science, web, opinion, etc. We will publish the results and software shortly.
Names for companies, products, or web domains should be short and memorable. However, they cannot clash with existing names, a vast multitude of which exists on the web. We thus developed a system to learn from moderately to large sized data corpora what looks and sounds good to humans, and to generate novel names that are not yet taken. With this algorithm, it is still possible to create web domains with 6 letters of fewer that are still available for registration. Fun fact: our company's name, Methority, was created by our creative AI system.
Developing automated intelligence system has great merits. At the same time, we also recognize that a lot of work still needs to be done to educate people about artificial intelligence, machine learning, statistics, etc. As a leader in automated intelligence, we take on this challenge by offering courses for corporations, both standard and customized to their needs. Our courses emphasize the relationship between teachers and students, which is reflected in the low student to teacher ratio that we offer.