Valentino Maiorca

My research focuses on understanding the mechanisms behind the emergence of the "semantic convergence" of distinct neural networks - both artificial and biological - to similar latent representations, exploring the potential for leveraging this synergy for broad interdisciplinary applications.
Central to our understanding is the distinction between intrinsic and extrinsic factors influencing the network representations. Intrinsic ones solely emerge from the fundamental characteristics encoded in the input data, revealing how networks shape information when processing similar concepts. Extrinsic factors, including random seed initialization or data modality, mainly influence the ambient space where these representations are embedded, serving as confounding variables that obfuscate the underlying alignment.
My work tackles this topic from three modality-agnostic perspectives: i) with the formalization of "latent communication" with respect to the semantics encoded in the input, the neural manifolds involved, and the transformations relating them; ii) by developing methods to make latent spaces "communicate" across different and independently trained neural networks, thereby harmonizing them; iii) by showing the practical, interdisciplinary implications of this alignment such as model stitching in deep learning, single-cell alignment in biology, and brain decoding in neuroscience.
Given its fundamental nature, anchored in the geometric properties of neural manifolds and the universal dynamics of data processing, latent communication has the potential to influence a wide range of domains, extending from scientific to societal impacts.