Adaptive networks for cognitive architectures: from advanced learning to neurorobotics and neuromorphic applications
 

What we do

The central ambition of the Work Package consists in achieving a measurable step forward in our understanding of human cognition; specifically, how biological learning networks enable human visuo-motor and cognitive functions.

The approach implemented is to emulate the architecture and operation of the brain. In practice, this corresponds to the design of functional cognitive architectures, addressing challenging visuo-motor and cognitive problems. This work is conducted through numerical simulations, with an emphasis on embodiment, and, for a selection of specific cognitive tasks, on real-world systems. The embodiment provides an agent through which the architecture can express itself, and a physically realistic environment to afford a simulation of task performance that is faithful to reality. Additional emphasis is placed on the use of neuromorphic computation, which allows rapid simulation of large-scale models. The work performed targets the development of a novel modular functional cognitive framework. This framework will support the integration of a range of functional modules, organized within an algorithmic architecture. These modules will emulate relevant brain functions and will express a range of cognitive functions in an embodied setting. Training of involved modules will be pursued, relying on a range of learning approaches. Particular emphasis will be placed on biologically plausible learning rules, further grounding the developed functional architectures within the biological reality they are designed for. This will be accomplished with the perspective of not only exploiting network plasticity to reproduce behavioral data, but doing so in a manner that is closer to the natural learning process than standard approaches. The developed framework will be exploited to support modular integration and embodiment of models of brain function implemented at distinct spatio-temporal scales and degrees of biological realism.

How we're organized

Work in the WP is structured across a number of complementary Tasks. Activities therein address the development of functional models, descriptive of particular brain regions. These models are developed to exhibit different levels of detail and reflect varying degrees of biological plausibility. Functions considered span a wide range of abstraction levels, from sensorimotor aspects to complex cognitive behaviors.

The active collaboration between modellers and learning experts allows the application of novel training methodologies, with a special emphasis on biologically plausible learning schemes. This training is necessary to achieve functionality and pursue different levels of performance. These activities are structured around a set of integrative demonstrators. Functional models are integrated within modular architectures, developed to address complex problems in an embodied setting. This integrative work is made to progress iteratively, building upon lower-abstraction levels towards more complex problems. This simulation and integration work is pursued in direct collaboration with EBRAINS developers, who provide supporting services and tools. Simulation of large-scale models, in particular, is facilitated by the use of neuromorphic platforms (SpiNNaker and BrainScaleS).

Developed showcases, publically hosted on EBRAINS, illustrate the merit of the approach and demonstrate the use of EBRAINS’ unique set of services and workflows.

Publication highlights

  1. Saxe, A., Nelli, S., & Summerfield, C. (2020). If deep learning is the answer, what is the question?. Nature Reviews Neuroscience, 1-13.
    https://www.nature.com/articles/s41583-020-00395-8
  2. Summerfield, C., Luyckx, F., & Sheahan, H. (2020). Structure learning and the posterior parietal cortex. Progress in neurobiology184, 101717.
    https://doi.org/10.1016/j.pneurobio.2019.101717
  3. S. Coppolino, G. Giacopelli and M. Migliore, "Sequence Learning in a Single Trial: A Spiking Neurons Model Based on Hippocampal Circuitry," in IEEE Transactions on Neural Networks and Learning Systems, doi: 10.1109/TNNLS.2021.3049281.
    https://ieeexplore.ieee.org/document/9333591
  4. McCauley, J.P., Petroccione, M.A., D’Brant, L.Y., Todd, G.C., Affinnih, N., Wisnoski, J.J., Zahid, S., Shree, S., Sousa, A.A., De Guzman, R.M. and Migliore, R., 2020. Circadian modulation of neurons and astrocytes controls synaptic plasticity in hippocampal area CA1. Cell reports33(2), p.108255.
    https://doi.org/10.1016/j.celrep.2020.108255
  5. Pozzi, I., Bohte, S., & Roelfsema, P. (2020). Attention-Gated Brain Propagation: How the brain can implement reward-based error backpropagation. Advances in Neural Information Processing Systems, 33.
    https://proceedings.neurips.cc/paper/2020/file/1abb1e1ea5f481b589da52303b091cbb-Paper.pdf