Evidence of Intent

What parts of Deep Learning are modern?

December 14, 2015 Deep Learning

Conclusion: outside of a very brief period in which pre-training with Unsupervised Learning was shown to be helpful, Deep Learning has largely been about hardware brute force, and learning how to use brute force to solve problems.

Terms I need to learn more about

  • Pattern Deformations

  • Hessian-free learning

  • Batch Normalisation (Thanks A Breitman)

  • Competing Units


Neural networks have been around for a long time. What exactly changed between the 80s and now? Why the resurgence of late?

Reading this review paper (page 4),

  1. Shallow NN-like models with few such stages have been around for many decades if not centuries, and models with several successive nonlinear layers of neurons date back at least to the 1960s and 1970s
  2. Training of deeper architectures (ANNs with more layers, including RNNs) only became feasible through the use of unsupervised learning techniques in the 2000s
  3. Training of deeper architectures without unsupervised learning became possible later

I'm a bit puzzled by (3), because it's not clear where it differs from (1). There are 3 candidates:

  1. More data
  2. More computation
  3. Better techniques

More data

I think this is actually conflated with More computation below. Without fast computation, you can't get through enough data.

More computation

Page 23 of the reference - in 2010, a new MNIST record was set using backpropagation and pattern deformations, both of which are decades old.

these results seemed to suggest that advances in exploiting modern computing hardware were more important than advances in algorithms

Better techniques

Unsupervised learning followed by supervised learning

It seems that there was a year or two where this was popular, and then it just went away.

Vanishing / Exploding Gradients

LSTMs solve this, and are a recent invention. ReLU activations do not explode but have other problems.

Proper Initialization

(Thanks to A Breitman for mentioning this)

When they are not suffering from vanishing / exploding gradients, activation functions can be too linear. Reference