Part I
Develop probabilistic models
of behavior capable of handling complex interactive environments.
Inspiration in modern probabilistic machine learning.
High dimensional interpretable latent variables capturing perception, planning and action.
Qian, Xuelin, et al. "fMRI-PTE: A Large-scale fMRI Pretrained Transformer Encoder for Multi-Subject Brain Activity Decoding." arXiv preprint arXiv:2311.00342 (2023).
Going beyond fMRI to Image.
Develop probabilistic models
of behavior capable of handling complex interactive environments.
Inspiration in modern probabilistic machine learning.
High dimensional interpretable latent variables capturing perception, planning and action.
Develop probabilistic models
of behavior capable of handling complex interactive environments.
Inspiration in modern probabilistic machine learning.
High dimensional interpretable latent variables capturing perception, planning and action.
dSprites is a dataset of 2D shapes procedurally generated from 6 latent factors:
Higgins et al. "beta-VAE: Learning basic visual concepts with a constrained variational framework." In Proceedings of the International Conference on Learning Representations (ICLR). 2017.
Transfrom dSprites dataset into an
interactive environment with
movements along 4 latent factors:
https://github.com/dimarkov/active-dsprites
Multi-color multi-object environments
Continuous latent spaces?
Adding speed and acceleration to objects?
ML way of doing Bayesian predictive coding.
Marino, Joseph. "Predictive coding, variational autoencoders, and biological connections." Neural Computation 34.1 (2022): 1-44.
\(\hat{x}^n\)
\( x^n\)
\( z^n \)
Encoder
Decoder
Variational free energy or negative ELBO
Amortized inference
Two problems with amortized inference:
Friston, Karl. "A theory of cortical responses." Philosophical transactions of the Royal Society B: Biological sciences 360.1456 (2005): 815-836.
Marino, Joe, Yisong Yue, and Stephan Mandt. "Iterative amortized inference." International Conference on Machine Learning. PMLR, 2018.
\( \phi_n^{(k+1)} = f\left(\phi_n^{(k)}, \nabla_{\phi_n^{(k)}} \hat{F}_n, \pmb{W} \right) \)
\( \phi_n^{(k + 1)} = \phi_n^{(k)} + \beta_t \nabla_{\phi_n^{(k)}} \hat{F}_n \)
Stochastic Gradient descent
Learnable optimization algorithm
Iterative amortized inference
\( \phi_n = (\mu_n, \sigma_n) \)
Requirements:
Spatial Transformer Networks
Jaderberg, Max, Karen Simonyan, and Andrew Zisserman. "Spatial transformer networks." Advances in neural information processing systems 28 (2015).
Spatial Transformer Networks
Jaderberg, Max, Karen Simonyan, and Andrew Zisserman. "Spatial transformer networks." Advances in neural information processing systems 28 (2015).
+ bilinear interpolation
Spatial Transformer Networks
Amortized inference
Iterative inference
Amortized inference
Iterative inference
Khan, Mohammad Emtiyaz, and Håvard Rue. "The Bayesian learning rule." arXiv preprint arXiv:2107.04562 (2021).
Khan, Mohammad, et al. "Fast and scalable bayesian deep learning by weight-perturbation in adam." International conference on machine learning. PMLR, 2018.
Natural momentum for natural gradient SVI
Iterative inference
Iterative inference
Iterative inference
Iterative inference
Iterative inference
Representing latent state dynamics
Learning of A, B, and L.
Action selection via expected free energy minimization.
Integration with Bayesian sparse-coding
Hierarchical extensions for complex object representation and learning.
Position/Orientation/Scale