README.md

Overview

Attempted PyTorch implementation of Active Neural Generative Coding (ANGC) from Backprop-Free Reinforcement Learning with Active Neural Generative Coding. This builds on Neural Generative Coding (NGC) from The neural coding framework for learning generative models and applies it to RL problems.

Variables

My notation differs in spots from the paper. One major difference is that I use row-major convention (1-d vectors are row vectors instead of column vectors), so that I don't have to worry about transposing things when going from math to implementation.

Variable	Description
$L + 1$	number of layers
$T$	number of interence time steps (paper calls this $K$)
$\gamma_v$	inference update leak coefficient
$\beta_e$	prediction error coefficient
$\beta$	inference state update coefficient
$\gamma_e$	error weight update coefficient. "controls the time-scale at which the error synapses are adjusted (usually values in the range of [0.9, 1.0] are used)
$J_\ell$	dimensionality of layer $\ell$
$z^\ell$, $\forall \ell = 1, \ldots, K-1$	hidden layer state vectors. $z^\ell \in \mathbb{R}^{1 \times J_\ell}$
$z^0$	bottom sensory vector, clamped to sensory input $z^0 = x^o$
$z^K$	top sensory vector, clamped to sensory input $z^K = x^i$
$W^{\ell} \in \mathbb{R}^{J_{\ell} \times J_{\ell - 1}}$	(top-down) prediction weights for layer $\ell$, also called the generative weights
$E^{\ell} \in \mathbb{R}^{J_{\ell-1} \times J_{\ell}}$	(bottom-up) error weights for layer $\ell$
$\phi^\ell$	activation function for layer $\ell$
$g^\ell$	another activation function for layer $\ell$. the paper says they use the identity function $g^\ell = \text{id}[J_\ell]$
$\bar{z}^\ell$	top-down prediction vector of $z^\ell$
$e^\ell$	prediction error vector for layer $\ell$
$d^\ell$	bottom-up + top-down inference pressure
$M_W^\ell \in \mathbb{R}^{J_{\ell+1} \times J_{\ell}}$	generative weights modulation matrix for layer $\ell$
$M_E^\ell \in \mathbb{R}^{J_{\ell} \times J_{\ell+1}}$	error weights modulation matrix for layer $\ell$

Inference

There's an inference phase that iterates, for timesteps $t = 1, \dots, T$:

$$\bar{z}^\ell(t) = g^{\ell}(\phi^{\ell + 1}[z^{\ell + 1}(t)] \cdot W^{\ell + 1})$$

$$e^\ell(t) = \frac{1}{2 \beta_e} (\phi^\ell[z^\ell(t)] - \bar{z}^\ell(t))$$

$$d^\ell(t) = -e^{\ell}(t) + e^{\ell-1}(t) E^{\ell}$$

$$z^\ell(t) = z^\ell(t-1) + \beta (- \gamma_v z^{\ell}(t) + d^{\ell}(t))$$

(where the last equation is slightly simplified to exclude the NGC lateral term, which is not used in ANGC):

TODO pseudocode

Update

$$\Delta W^\ell = \phi^\ell[z^{\ell + 1}]^\top e^{\ell} \odot M_W^\ell$$

$$\Delta E^\ell = \gamma_e (\Delta W^{\ell})^\top \odot M_E^\ell$$