Cospectral is an MIT spinoff developing new conceptual and computational architectures for physical continual learning1 systems. Our goal is to address the limitations of offline statistical models, which require ever-increasing amounts of hand-coded data and are inextricably bound by computational resources. They are inherently inefficient since the past is rarely a good model of the future.

We instead revisit the learning problem using its original framing: to act directly on continuous, unstructured datastreams using a universal joint representation, guided by sparse2 abstractions of causal relationships3 and identified components4 which are simplified over time5,6. Existing systems look backwards to make predictions. We focus on describing and evaluating the present.

Our researchers have backgrounds in robotics and reinforcement learning, with experience developing autonomous vehicles, personal robots, spacecraft, multimodal models, and safety‑critical systems. Drawing on algorithmic information theory, sparse reconstruction, approximation theory, harmonic analysis, and compressive sensing, our proprietary algorithms yield systems that are online7, adaptive8, interpretable9,10, and minimal5,6.

1. Ring, Mark B. “Child: A first step towards continual learning.” Machine Learning 28, no. 1 (1997)
2. Foucart, Simon, and Holger Rauhut. “Sparse solutions of underdetermined systems.”
   In A Mathematical Introduction to Compressive Sensing, pp. 41-59. 2013.
3. Good, Irving J. “A causal calculus (I).” The British journal for the philosophy of science 11, no. 44 (1961)
4. Soderstrom, Torsten, and Petre Stoica. “System identification.“ Prentice-Hall, 1989.
5. Rissanen, Jorma. “Modeling by shortest data description.” Automatica 14, no. 5 (1978)
6. Kolmogorov, Andrei N. “Three approaches to the quantitative definition of information.”
   Problems of Information Transmission 1, no. 1 (1965)
7. Sutton, Richard S., and Steven D. Whitehead. “Online learning with random representations.” In ICML, 1993.
8. Li, Puheng, Tijana Zrnic, and Emmanuel Candes. “Robust sampling for active statistical inference.”
   Advances in Neural Information Processing Systems 38 (2026)
9. Behboodi, Arash, Holger Rauhut, and Ekkehard Schnoor. “Compressive sensing and neural networks from a statistical learning perspective.” In Compressed Sensing in Information Processing, 2022.
10. Zeger, Emi, Yifei Wang, Aaron Mishkin, Tolga Ergen, Emmanuel Candès, and Mert Pilanci. “A Library of Mirrors: Deep Neural Nets in Low Dimensions are Convex Lasso Models with Reflection Features.” 2024.