§15.6Providing Clues to Discover Underlying Causes

Part III DL pp. 554–557 · ~6 min read

§15.6 The catalog of clues

To close the chapter, back to the original question one last time: what makes one representation better than another? The answer of §15.3: an ideal representation disentangles the underlying causal factors — especially those relevant to our applications. Most representation-learning strategies work by introducing clues that help the learner find and separate those factors. Supervised learning is the strongest clue — a label yy delivered with each x\boldsymbol{x} usually specifies at least one factor of variation directly. To exploit abundant unlabeled data, learners rely on less direct hints: implicit prior beliefs that we, the algorithm’s designers, impose. The no-free-lunch theorem says regularization of some kind is necessary, and no universally superior choice exists — deep learning’s wager is that a set of fairly generic priors covers the kinds of tasks people and animals solve.

The clue map — eleven generic priors (Bengio et al. 2013, expanded), each a different way of pointing the learner at the underlying causal factors h. Grouped by what they constrain: the function’s geometry, the factor structure, the architecture, and the dynamics.
underlying causalfactors hsmoothnesslinearitymanifoldsmultiple factorscausal factorsnatural clusteringtemporal coherencedepth / hierarchysparsityshared factors (tasks)simple dependenciesgeometry (blue) · factor structure (green) · architecture (red) · dynamics (amber)

Each entry, with where you have already met it in this book:

The eleven generic priors. Click a cell for the caveat or the mechanism.
What it assumesWhere it appeared
Smoothnessf(x + εd) ≈ f(x) — generalize to nearby pointsnearly every algorithm; §5.11.2
Linearitysome variable relationships are linear — predict far from the datalinear models; §7.13
Multiple explanatory factorsdata generated by several underlying factors; tasks easy GIVEN the factorsthe §15.3 semi-supervised argument; §15.4 directions
Causal factorsthe learned h are CAUSES of x, not effects§15.3 (Schölkopf robustness)
Depth / hierarchyabstract concepts defined via simpler ones; a multi-step program§6.4.1, §15.5
Shared factors across taskseach task's yᵢ ties to a SUBSET of a common factor pool hmulti-task §7.7; transfer §15.2
Manifoldsprobability mass concentrates on low-dimensional, locally connected regions§5.11.3, §13.5, §14.6 (autoencoders learn them explicitly)
Natural clusteringeach CONNECTED manifold carries a single classtangent prop & manifold tangent classifier §7.14; adversarial training §7.13
Temporal & spatial coherenceimportant factors change slowly over time (or are at least easy to predict)slow feature analysis §13.3
Sparsitymost features are irrelevant to most inputs — "present/absent" features should mostly be absent§7.10, §13.4, §14.2.1
Simplicity of factor dependencieshigh-level factors relate through SIMPLE (marginal-independent, linear, or shallow) dependenciesfactorial priors §13.1/§14.2; linear classifiers on h §15.4
Dotted-underlined cells have explanations — click one.

One catalog, one book

Read the rows again and this catalog is a map of everything so far: smoothness from Part I, sparsity and the manifold hypothesis from the regularization and autoencoder chapters, the slowness principle from linear factor models, depth from the feedforward chapters, shared factors from multi-task and transfer learning, causal disentangling from this chapter. Representation learning ties together all the many forms of deep learning — feedforward and recurrent networks, autoencoders and deep probabilistic models all learn and exploit representations. Learning the best possible one remains an open and exciting research problem.

That closes Chapter 15 — and sets up the rest of Part III: chapters 16–19 build the machinery of structured probabilistic models, sampling, partition functions and inference, so that chapter 20’s deep generative models can actually learn the causal factors this chapter argued for.

Check yourself — clues to underlying causes

1.Why must representation learning rely on "clues" (implicit priors) at all?

2.Smoothness and linearity are often conflated. How do they actually differ?

3.Which catalog priors do slow feature analysis and adversarial training respectively instantiate?

4.What does the "natural clustering" prior assert?

5.The "simplicity of factor dependencies" prior underlies which everyday modeling choice?

5 questions