Gijs Joost Brouwer

My Mycelium project is an extensive online course about machine learning, data science, and artificial intelligence, primarily focusing on machine learning but written for a broad audience. Specifically, it is for those of us who find ourselves on highly interdisciplinary technology teams but who are not data scientists or machine learning engineers but would like to learn.

Please note that this is currently a work in progress.

Samson is a concept project I introduced to my innovation team while working at Memorial Sloan Kettering, a large oncology center in New York City.

Samson The Cat is a virtual companion that can accompany a child during their journey at the hospital and even after they leave. But unlike an app or a toy, my team and I thought of Samson as an embedded AI, a ghost in the shell brought to life by any available technology. This makes Samson largely device and technology-independent. A phone, a screen, a toy, a creature living in a metaverse, while also capable of taking on a physical form through any IoT technology available. This creates a sense of presence and continuity.

This project aimed at detecting questionable, offensive, or otherwise undesired content on web pages through the images it hosted. To do so, I augmented a pre-trained convolutional deep belief neural network with categories automatically discovered through a clustering approach that used my company's existing NLP metrics.

At the United Nations, I worked on The D4D-Senegal challenge, an open innovation data challenge on anonymous call patterns of Orange's mobile phone users in Senegal. The challenge's goal is to help address society development questions in novel ways by contributing to the socio-economic development and well-being of the Senegalese population.

I worked with a large mobile phone dataset based on call detail records (CDR) of phone calls and text exchanges between more than 9 million customers of a large telecommunication company. I created an algorithm that captures both the regularities and anomalies in the patterns of hum mobility. I extract the most common cell phone tower each user's phone is connected to, from which the algorithm learned to output a 'surprise feature' as well as a probability of a transition between two separate towers from the input data.

The algorithm could detect regular and anomalous patterns by tracking these metrics over time to alert humans. I designed the algorithm to guide infrastructural planning from the regularities it detects and load higher capacity roads on public transport. Second, the algorithm can serve as an early warning system of disruptive events involving many human beings.

Intensive care units and emergency rooms are extremely noisy places in hospitals. They are filled with medical equipment that all produce various sounds. Some of these sounds are essential indicators, for example, heart rate monitors. Other sounds are the inevitable result of care given and the communication between staff and patients.

Within all of this, clinical staff will need to be able to hear specific alarms from a distance, as they cannot have eyes on every patient. However, a lot of the sounds are not relevant to anyone clinician. Similarly, not all conversations at any time are relevant for anyone clinician.

Finally, most sounds are irrelevant from the patient's point of view. They only really need to comprehend and understand things that are said to them, not necessarily to anyone else. Perhaps they would prefer focusing on the sound from a TV for distraction.

The real solution seems obvious: both patient and clinician should have some ability to tune in to specific sounds or voices while others are muted. Therefore, I experimented with a known technique called source separation that theoretically will allow for this, but as applied to a hospital setting, through simulation.