Research
It is a standard notion in the AI community that deep learning has the potential to perform well at human-level tasks partly because of its neuroscience roots. In the neuroscience community, it is commonly believed that deep networks can be helpful as models of the brain. But the computations underlying artificial networks and the diversity of neuronal function go well beyond these first analogies. How exactly to introduce biological features into artificial networks to enhance robustness and performance, and how to properly analyze and explain biological as well as artificial networks are key challenges facing machine learning and computational neuroscience today. The goal of my research is to rigorously address such questions from both perspectives.
Selected papers
Learning dynamic modular structure of functional connectomes
The functional connectome in neurobiology, like community structure in social networks, is dynamic: it changes in time with behavioral state, sensory inputs, and internal constraints. I have developed an unsupervised algorithm to learn the dynamic affinities between neurons in live, behaving animals—the worm C. elegans—and to reveal which communities form among neurons at different times.
Dyballa, L., Lang, S., Haslung-Gourley, A., Yemini, E., Zucker, S., 2024. "Learning dynamic representations of the functional connectome in neurobiological networks", The Twelfth International Conference on Learning Representations (ICLR '24). Link
(See also my work in the modular structure of anatomical connectivity in the primate below.)
Neural encoding manifolds reveal the topology of neural networks
I have combined my contributions to manifold learning and tensor factorization to demonstrate, in a data-driven way, that visual feature representation at the population level is qualitatively different between two areas in the mouse visual system: retina encodes features discretely (neurons form clusters), while primary visual cortex covers feature space relatively uniformly. I have also shown that the same method can be directly applied to artificial networks.
Dyballa, L., Rudzite, A., Hoseini, M., Thapa, M., Stryker, M., Field, G., Zucker, S. (2024), “Population coding of stimulus features along the visual hierarchy”, PNAS, 121(4). https://doi.org/10.1073/pnas.2317773121
Unsupervised similarity kernels for manifold learning and dimensionality estimation
Manifold learning, the state of the art in non-linear low-dimensional embeddings, is based on a data graph expressing which data points are similar. Standard approaches assume the data are sampled uniformly from a pure manifold of known dimension, thereby imposing that an arbitrary neighborhood size (e.g., k=8) should hold for each point. Real world datasets often do not follow these assumptions; manifolds can be nonpure, disconnected, or highly curved.
My Iterative Adaptive Neighborhoods (IAN) algorithm tackles these issues directly. First, it is able infer an appropriate data graph from only pairwise distance information in an unsupervised way. This involves learning geometry-constrained local neighborhoods. It can be used with most manifold-learning algorithms and provides reliable intrinsic dimensionality estimates efficiently for real-world datasets.
Dyballa, L., Zucker, S. (2023), “IAN kernel: Iterated Adaptive Neighborhoods for manifold learning and dimensionality estimation”, Neural Computation, 35(3): 453-524. https://doi.org/10.1162/neco_a_01566 Preprint: https://arxiv.org/abs/2208.09123
Dyballa, L., Hoseini, M., Dadarlat, M., Zucker, S., Stryker, M. (2018), “Flow stimuli reveal ecologically appropriate responses in mouse visual cortex”, PNAS, 115(44). https://doi.org/10.1073/pnas.1811265115
The visual system of the mouse is now widely studied as a model for development and disease in humans. Studies of its primary visual cortex (V1) using conventional grating stimuli to construct linear–nonlinear receptive fields suggest that the mouse must have very poor vision. Using stimuli resembling the flow of images across the retina as the mouse moves through the grass, we find that most V1 neurons respond reliably to very much finer details of the visual scene than previously believed. Our findings suggest that the conventional notion of a unique receptive field does not capture the operation of the neural network in mouse V1.
Gerritz, E., Dyballa, L., Zucker, S. (2024), "Zero-shot generalization across architectures for visual classification", The Twelfth International Conference on Learning Representations (ICLR '24) – TinyPapers Track (to appear).
Generalization to unseen data is a key desideratum for deep networks, but its relation to classification accuracy is unclear. Using a minimalist vision dataset and a measure of generalizability, we show that popular networks, from deep convolutional networks (CNNs) to transformers, vary in their power to extrapolate to unseen classes both across layers and across architectures. Accuracy is not a good predictor of generalizability, and generalization varies non-monotonically with layer depth.
Dyballa, L., Barbosa, V. (2015), “Further insights into the interareal connectivity of a cortical network”, Network Science, 3(4). https://doi.org/10.1017/nws.2015.19
Over the past years, network science has proven invaluable as a means to better understand many of the processes taking place in the brain. We explore new aspects of a densely connected interareal atlas, such as a correlation between connection weights and cortical hierarchy. We extend a link-community detection algorithm to allow for directed connections, and analyze structure that emerges from the data to uncover the major communication pathways in the network.