My research targets fundamentally distinct, but interdependent, levels of Bayesian optimisation and investigates the necessary computations required for: 

• building flexible generative models for life-long learning that can aptly encode (temporal and spatial) information even when the embodied environment changes e.g., Fountas, Sajid, et al, NeurIPs, 2020 (DAI-MC); Yuan, Sajid, et al, 2021. 

• that can support adaptive planning strategies. This can be introduced using information-based exploration e.g., Sajid, Ball, et al, Neural Computation, 2021; Sajid, Da Costa, et al., The Drive for Knowledge 2022. Conversely, we have formulated latent state preference model learning to induce adaptive behaviour under static deep generative models e.g., Sajid, Tigas, et al, ICML URL, 2021 (PEPPER); Sajid, Tigas, et al, COLLAs Workshop, 2022 (NORE).

• afford appropriate optimisation of perceptual learning and inference i.e., what is the right level of approximation for posterior estimation and ensuing message passing schemes for different contexts e.g., Sajid, Faccio, et al Neural Computations, 2022; Sajid, Convertino, Friston, Entropy, 2021; Parr, Sajid, Friston, Entropy, 2020. 

Additionally, I am focused on investigating how these generative models may adapt, adjust and (re-)learn after perturbations via computational lesions. For this, we have introduced:

• theoretical understanding of functional degeneracy e.g., Sajid, Parr, et al, Cerebral Cortex, 2020; 

• quantitative understanding of how distributed structural lesions disrupt functional outcomes e.g., Sajid, Holmes, et al, Scientific Reports, 2021, Sajid, Parr, et al, Brain Communications, 2021;

• empirical evidence of degenerate functional architecture e.g., Sajid, Gajardo-Vidal, et al, 2022.