finding

active

finding:das-achieves-100-iia-on-hierarchical-equality-task-with-n-16-intervention-size-8-layer-1

DAS achieves 100% IIA on hierarchical equality task with |N|=16, intervention size 8, Layer 1

DAS discovers a perfect alignment between the feed-forward network and the Both Equality Relations high-level model.

Source paper

extracted_from

Finding Alignments Between Interpretable Causal Variables and Distributed Neural Representations

(2023) · Atticus Geiger · Zhengxuan Wu · Christopher Potts · Thomas Icard +1

Neighborhood — ranked by edge-count

Claims (3)

claim

DAS overcomes the localist limitation of prior causal abstraction by allowing individual neurons to play multiple roles via non-standard bases
supports
Central claim motivating DAS over prior methods.
The feed-forward network truly implements a symbolic, tree-structured algorithm for hierarchical equality, with abstract equality relations not decomposable into input identities
supports
DAS reveals that the neural network encodes abstract relational structure rather than raw input identities.
The discovery of perfect abstract equality representations that cannot be decomposed into entity representations is a foundational result informing our understanding of how symbolic and connectionist architectures coexist
supports
Concluding claim about theoretical significance of the hierarchical equality finding.

Questions (1)

question

Does the hierarchical equality network implement a program that computes w=x and y=z as intermediate values?
answered_by
Specific research question for the first experiment.

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

DAS runs in 502 seconds for hierarchical equality vs. estimated 6e8 seconds for exhaustive brute-force searchfinding0.804
DAS runtime is invariant with number of testing hypotheses, unlike brute-force search.
Brute-force search achieves best IIA of 0.60 on hierarchical equality Both Equality Relations in Layer 1finding0.803
DAS substantially outperforms brute-force search (1.00 vs 0.60 IIA) on the hierarchical equality task.
DAS achieves 100% IIA for combined Negation and Lexical Entailment model on MoNLI at Layer 9, intervention size 256finding0.795
Perfect abstraction relation between BERT and symbolic algorithm with negation and lexical entailment variables.
Over 80% IIA achieved using complex non-linear alignment maps on randomly initialised MLPs in hierarchical equality taskfinding0.793
Demonstrates that high IIA can be obtained even when model cannot solve the task
Best localist alignment achieves IIA of 0.73 on hierarchical equality Both Equality Relations in Layer 1finding0.770
Shows localist alignment fails to capture the distributed structure found by DAS.
Trainable intervention (DAS) finds sparser gender representations than linear probing, suggesting probing overestimates causal coverageclaim0.769
Interpretive claim from Case Study II about the distinction between correlational probes and causal interventions
DAS behavioral loss achieves IIA of 0.997±0.001 on synthetic 10-class dataset training/test setsfinding0.767
IIA baseline for DAS behavioral loss on synthetic dataset
DAS trainable intervention finds sparser gender representations across layers compared to linear probe in Pythia-6.9Bfinding0.767
Case Study II result showing DAS identifies fewer causally relevant positions than a probe