Till Hartmann

Hi! I am a senior engineer at Zoox. Previously, I was director of research at Perceptive Automata, building AI that understands pedestrian for self driving cars and a visiting scientist at Harvard Medical School. But over a decade ago, I thought of myself as a physicist due to my university training. However, through my work in graduate school and post-doctoral fellowship, I have developed into a systems neuroscientist. Before working at Perceptive Automata, I was employed as a research associate at Harvard Medical School and teach at Harvard College. I'm interested in machine learning and neural networks, but also systems neuroscience and computational neuroscience. On this site, you can find a list of my publications, a brief biography, and a description of my academic research.

Recurrent CNNs

In my latest conference submission, I proposed to use recurrent connections in convolutional layers to improve object classification when it is getting dark, i.e. the signal-to-noise ratio is getting low. The plot below shows that for bad image quality, the classical convolutional neural network (cCNN) performs just above chance over time, whereas the CNN with GRU architecture (a recurrent network) increases classification performance over time.

gruCNN
Recurrent convolutional neural network (gruCNN) outperforms classical convolutional neural network (cCNN).

Pre-print: arXiv

Gamma dependency on cortical feedback

Gamma Oscillations in primary visual cortex are a well-established phenomenon—their role in cortical information processing is unclear. We found strong evidence that these oscillations are heavily dependent on cortico-cortical feedback from higher cortical areas.

Oscillation
LFP with and without cortico-cortical feedback