Inductive Biases for Deep Learning of Higher-Level Cognition

In plain words

Today's deep learning systems do very well on test data that looks like their training data, but they often break when conditions shift, or when a new combination of familiar parts shows up. People learn from far fewer examples and handle such novelty much better. The authors ask what built-in assumptions, the kind a learner uses to prefer one explanation over another, would close that gap.

Their answer draws on how human conscious thought seems to work. Only a few things are held in mind at once, and those things can usually be put into words. The authors suggest that a model whose basic units are such verbal, cause-and-effect concepts could reuse knowledge, combine it in new ways, and adapt quickly when only one part of the world changes.

The paper is a position piece, not a method with results. It lists the assumptions it thinks are worth adding to deep learning and points to earlier work that tries some of them. It does not show how to combine them in one system, and the authors say that is still open.

TL;DR

• A position paper, not a method: it argues that deep learning needs extra inductive biases, drawn from high-level conscious processing, for out-of-distribution and systematic generalization §1 §2.
• Proposed biases: high-level variables are verbalisable and causal §3.4 §3.5; changes are sparse and local §3.6; a sparse factor graph links them §3.8; knowledge is generic and reusable §3.9; relevant causal chains are short §3.10; top-down and bottom-up signals combine at every level §3.11. Summary in tab. 2.
• It gives no training procedure, no experiments and no benchmark numbers. The authors call combining these biases in a single system an open question §1.

Problem

Current systems are often not robust to distribution changes, adversarial examples or spurious correlations §1.1. Without assumptions about how training and test distributions relate, nothing can be said about generalization to a changed distribution §1.1. The no-free-lunch argument says that some preference over functions is needed to generalize at all §2. A test set drawn from the training distribution cannot reveal overfitting to that distribution, so evaluation has to change as well [§2, "Systematic Generalization and Out-of-Distribution Generalization"].

The paper argues that data, compute and architecture alone are insufficient without further inductive biases §1. Its stated assumptions are that the training distribution is not fixed in sequential settings and that agents act on their environment [§2, "Agency, Sequential Decision Making and Non-Stationary Data Streams"].

Method

The paper has no algorithm. Its contribution is a list of inductive biases, each argued from cognitive science and causal modelling, and a summary table tab. 2. Table 1 lists biases already used in deep learning, such as convolution, recurrence and soft attention tab. 1.

1. Verbalisable high-level variables. High-level variables are those manipulated consciously, and they are generally verbalizable §3.4. The authors say they hypothesise that these variables are often causal too §3.5, remark "Semantic variables are often also causal variables".
2. Sparse, local changes. Most changes in distribution come from one or a few variables or mechanisms, which can be named in a short sentence §3.6. Interventions start at one or a few nodes of the causal graph §3.6.
3. Stable and fast knowledge. Stable parts of the world should be learned slowly and shared, while volatile parts are learned fast and are the only ones that need adapting §3.7. The paper says this calls for at least two learning speeds, as in meta-learning's outer and inner loops §3.7 §2.
4. Sparse factor graph. The joint distribution over high-level variables is a sparse factor graph, not a set of independent marginals §3.8. The sparsity mirrors the limited capacity of the global-workspace bottleneck §3.8 §3.3.
5. Generic, reusable knowledge. Knowledge is written as schemas, or factor templates, whose argument variables are bound to instances §3.9. The same rule can apply to different objects, and the knowledge is not stored per instance §3.9.
6. Short causal chains. Learning and inference chains are short in depth, even when the events they link are far apart in time §3.10. The authors say earlier work they cite found this prior the strongest remedy for vanishing gradients over long dependencies §3.10.
7. Top-down and bottom-up combination. Top-down context should combine with bottom-up sensory signals dynamically, at every level of the hierarchy §3.11.
8. Independent causal mechanisms. An intervention on one mechanism should not change the others, and knowing one mechanism should tell you nothing about another §4.1. The formal statement of structural causal models is in §4, eq. 1 and §4, eq. 2.
9. Use non-stationarity. Data should not be shuffled into an i.i.d. sample, because the changes between sources carry causal information §4.2.

Two supporting ideas appear in the text. First, soft key-value attention is presented as a soft form of binding: a query selects which object fills an argument, with a separate query per argument [§3.2, paragraph "Attention implements variable binding"]. The paper reads current attention as already doing soft binding, and says discrete types and hard choices are missing §3.2. Second, the global workspace is a bottleneck through which only a few variables are broadcast at a time, and the authors use it to motivate the sparse graph §3.3 §3.8.

The one formula in the discussion of attention is the scaled dot-product form. The paper does not number it.

\mathrm{Attention}(Q,K,V)=\mathrm{softmax}\left(\frac{QK^T}{\sqrt{d}}\right)V §3.2, unnumbered display

The structural causal model is defined by two statements. The joint density factorises over a DAG §4, eq. 1. Each variable is a deterministic function of its parents and an independent noise §4, eq. 2.

p(X_1,\dots,X_n) = \prod_{i=1}^{m} p(X_i \mid PA_i) eq. 1

X_i := f_i\big(X_{pa(i,C)}, N_i\big), \quad \forall i \in \{1,\dots,M\} eq. 2

Given the structure, sampling one joint draw follows eq. 2 directly. The paper does not give this procedure, so the order below is a standard way to honour the DAG and not the paper's own algorithm.

# Sample one joint draw from an SCM  [§4 eq. 1] [§4 eq. 2]
# Input:  parents[i] for each of M variables (from the DAG C)
#         functions[i], noise_samplers[i]
# Output: values of X_1..X_M for one draw

def sample_scm(parents, functions, noise_samplers):
    # One independent noise per variable  [§4 eq. 2]
    noises = [draw(noise_samplers[i]) for i in range(M)]
    values = [None] * M
    # Visit variables so each parent is set before its children
    for i in topological_order(parents):
        # Structural assignment X_i := f_i(parents of i, N_i)  [§4 eq. 2]
        values[i] = functions[i](values[parents[i]], noises[i])
    return values

Hyperparameters

The paper trains nothing, so no training hyperparameters are stated.

Results

The paper reports no experiments and no numerical results §1. Its statements about better out-of-distribution generalization are motivation from earlier work it cites §1 §4.3. The one simulation discussed, in the meta-learning and causality section, is that of cited work and not of this paper §4.3. The paper says the cited preliminary work suggests that parameterising the correct variables lets a model adapt faster to interventions §4.3.

No results table is given, because the paper has none.

Paper vs code

No official code.

The paper is a position piece with no implementation. Its LaTeX source has no project or code link beyond package and archive URLs. Checked: the paper text, its LaTeX source, and a GitHub repository search for "inductive biases higher-level cognition", which returned no results.

Pitfalls

• Position, not evidence. The authors call their biases hypotheses and state that combining them in one system is open §1. Treat every claim in Method as a design argument.
• Table 2 references are not endorsements. The "Relevant References" column cites the work the authors treat as progress on each bias tab. 2. This card does not use their content, and the paper does not evaluate them in a common setting.
• Causal claims are argued, not measured. The authors say they hypothesise that semantic variables are causal §3.5. The sparsity of verbalizable knowledge is argued from the length of sentences §3.8. No measurement is given.
• Disagreement with disentangling. The paper rejects the assumption, common in disentangling work it cites, that high-level variables are marginally independent, and asks instead for a sparse joint structure §3.8. An implementer who adds an independence penalty goes against this paper.
• Two senses of "independent". The paper defines mechanism independence with algorithmic mutual information §3.5 and, for the causal principle, with information theory §4.1. Neither gives a computable test.
• Counts in the SCM equations. Equation 1 writes n variables but multiplies over i up to m. Equation 2 has M variables and noises N_1 to N_m. The paper does not say these counts are equal §4 eq. 1 §4 eq. 2. Use one count.
• Sampling is not specified. The SCM is defined as assignments over a DAG, but no sampling order is stated §4 eq. 2. The topological order in the pseudocode is standard, not the paper's.
• Exponential advantages are asserted. The paper says some compositional biases may be worth "even possibly exponentially more data" [§2, "Inductive Biases as Data"]. It gives no derivation and relies on cited work.
• Non-stationary data is harder to evaluate. The authors say that non-stationary, interactive settings make comparing algorithms harder, and that they consider this necessary and feasible [§2, "Agency, Sequential Decision Making and Non-Stationary Data Streams"]. They give no benchmark.
• Compiled section numbers. Section numbers here follow the LaTeX source with \iffalse blocks skipped, which removes a block titled "What do we want to achieve?" and a subsection on programming-language inspiration. The compiled PDF was not read, so its numbering may differ. Check before citing the section numbers in other pages.
• Attention is only partly covered. The paper says current attention supports soft binding but lacks discrete typed arguments and hard choices §3.2. It does not describe a method for hard typed binding.

Lineage

• Builds on: structural causal models, as in Peters et al. §4. The paper presents the meta-learning causal example as coming from Bengio et al. (2019), a meta-transfer objective; the paper's bibliography gives arXiv id 1901.10912 §4.3. That paper has no card in this wiki, so it is not linked.
• Compared against: none. The paper contains no comparison of methods.
• Superseded by: not stated.