m8ta
use https for features.
text: sort by
tags: modified
type: chronology
{1455}
hide / / print
ref: -0 tags: credit assignment distributed feedback alignment penn state MNIST fashion backprop date: 03-16-2019 02:21 gmt revision:1 [0] [head]

Conducting credit assignment by aligning local distributed representations

  • Alexander G. Ororbia, Ankur Mali, Daniel Kifer, C. Lee Giles
  • Propose two related algorithms: Local Representation Alignment (LRA)-diff and LRA-fdbk.
    • LRA-diff is basically a modified form of backprop.
    • LRA-fdbk is a modified version of feedback alignment. {1432} {1423}
  • Test on MNIST (easy -- many digits can be discriminated with one pixel!) and fashion-MNIST (harder -- humans only get about 85% right!)
  • Use a Cauchy or log-penalty loss at each layer, which is somewhat unique and interesting: L(z,y)= i=1 nlog(1+(y iz i) 2)L(z,y) = \sum_{i=1}^n{ log(1 + (y_i - z_i)^2)} .
    • This is hence a saturating loss.
  1. Normal multi-layer-perceptron feedforward network. pre activation h h^\ell and post activation z z^\ell are stored.
  2. Update the weights to minimize loss. This gradient calculation is identical to backprop, only they constrain the update to have a norm no bigger than c 1c_1 . Z and Y are actual and desired output of the layer, as commented. Gradient includes the derivative of the nonlinear activation function.
  3. Generaete update for the pre-nonlinearity h 1h^{\ell-1} to minimize the loss in the layer above. This again is very similar to backprop; its' the chain rule -- but the derivatives are vectors, of course, so those should be element-wise multiplication, not outer produts (i think).
    1. Note hh is updated -- derivatives of two nonlinearities.
  4. Feedback-alignment version, with random matrix E E_{\ell} (elements drawn from a gaussian distribution, σ=1\sigma = 1 ish.
    1. Only one nonlinearity derivative here -- bug?
  5. Move the rep and post activations in the specified gradient direction.
    1. Those h¯ 1\bar{h}^{\ell-1} variables are temporary holding -- but note that both lower and higher layers are updated.
  6. Do this K of times, K=1-50.
  • In practice K=1, with the LRA-fdbk algorithm, for the majority of the paper -- it works much better than LRA-diff (interesting .. bug?). Hence, this basically reduces to feedback alignment.
  • Demonstrate that LRA works much better with small initial weights, but basically because they tweak the algorithm to do this.
    • Need to see a positive control for this to be conclusive.
    • Again, why is FA so different from LRA-fdbk? Suspicious. Positive controls.
  • Attempted a network with Local Winner Take All (LWTA), which is a hard nonlinearity that LFA was able to account for & train through.
  • Also used Bernoulli neurons, and were able to successfully train. Unlike drop-out, these were stochastic at test time, and things still worked OK.

Lit review.
  • Logistic sigmoid can slow down learning, due to it's non-zero mean (Glorot & Bengio 2010).
  • Recirculation algorithm (or generalized recirculation) is a precursor for target propagation.
  • Target propagation is all about the inverse of the forward propagation: if we had access to the inverse of the network of forward propagations, we could compute which input values at the lower levels of the network would result in better values at the top that would please the global cost.
    • This is a very different way of looking at it -- almost backwards!
    • And indeed, it's not really all that different from contrastive divergence. (even though CD doesn't work well with non-Bernoulli units)
  • Contractive Hebbian learning also has two phases, one to fantasize, and done to try to make the fantasies look more like the input data.
  • Decoupled neural interfaces (Jaderberg et al 2016): learn a predictive model of error gradients (and inputs) nistead of trying to use local information to estimate updated weights.

  • Yeah, call me a critic, but I'm not clear on the contribution of this paper; it smells precocious and over-sold.
    • Even the title. I was hoping for something more 'local' than per-layer computation. BP does that already!
  • They primarily report supportive tests, not discriminative or stressing tests; how does the algorithm fail?
    • Certainly a lot of work went into it..
  • I still don't see how the computation of a target through a ransom matrix, then using delta/loss/error between that target and the feedforward activation to update weights, is much different than propagating the errors directly through a random feedback matrix. Eg. subtract then multiply, or multiply then subtract?