ica-lens

name

ica-lens

description

Use this skill when applying Independent Component Analysis as a training-free interpretability lens — decomposing a target activation site (residual stream, MLP output, attention-head output, or any cached hook point) into maximally non-Gaussian directions and treating each direction as a candidate monosemantic component for sparse-probing, targeted perturbation, annotation, and cross-comparison against trained SAE / transcoder / crosscoder features.

ICA Lens: A Training-Free Interpretability Lens on Model Components

When to Use

Activate this skill when the goal is to characterise the interpretable structure of a component's activation distribution without paying the up-front cost of training a sparse dictionary.

Typical triggers:

• A new model, layer, or hook point needs an initial inventory of candidate interpretable directions before committing to SAE / transcoder training.
• An existing SAE / transcoder feature set needs an independent baseline of interpretable directions to compare against (overlap, coverage, perturbation strength, sparse-probe accuracy under a matched budget).
• A specific concept or behaviour is suspected to live on a low-dimensional, non-Gaussian subspace of a residual / MLP / head-output state, and a fast, gradient-free extractor is needed to expose it.
• A study of non-Gaussianity itself — where interpretable structure concentrates across depth, modality, and model family — is the object of interest.
• Targeted perturbation / steering needs a direction handle whose discovery is decoupled from the model's training signal.
• An interpretability artefact needs to be browsed, annotated, or shared via an explorer UI rather than recomputed each session.

Keywords: ICA, Independent Component Analysis, FastICA, non-Gaussianity, dictionary-free interpretability, interpretable directions, component interpretability, residual-stream directions, MLP-output directions, attention-head directions, sparse probing, targeted probe perturbation, SAE alternative, SAE baseline, ICA explorer.

Method in One Paragraph

The ICA Lens treats each cached activation $\mathbf{a} \in \mathbb{R}^{d_{\text{model}}}$ at a chosen hook point as a sample from an unknown source distribution and fits an unmixing matrix $\mathbf{W}{\mathrm{ICA}}$ such that the components $\mathbf{f}(\mathbf{a}) = \mathbf{W}{\mathrm{ICA}},\mathbf{a}$ are maximally non-Gaussian and mutually independent. The motivating observation is that interpretable, token-selective directions in language models appear systematically less Gaussian than random directions in the same space, so non-Gaussianity is a sufficient surrogate objective to recover candidate monosemantic axes without any reconstruction or sparsity loss. Compared with sparse-autoencoder-family methods, no over-complete dictionary is trained, no per-feature dead-direction / split-feature pathology is incurred, and the recovered directions can be evaluated by exactly the same downstream protocols — sparse probing, targeted probe perturbation, top-activating example mining, and feature dashboards.

When to Prefer ICA Lens Over a Trained Dictionary

Situation	ICA Lens	SAE / Transcoder / Crosscoder
First-pass inventory on a new site / model	preferred	overkill
Budget-constrained perturbation study	preferred (better TPP under budget)	weaker under matched budget
Need an independent baseline to score a learned dictionary against	preferred	the object under test
Need more directions than the activation has dimensions (over-complete)	not applicable (ICA returns $\le d_{\text{model}}$ components)	preferred
Need explicit input → output feature mapping for circuit replacement	not applicable	transcoder preferred
Need cross-site / cross-checkpoint shared basis	not applicable	crosscoder preferred

The general rule: start with ICA Lens, escalate to a learned dictionary only when ICA's resolution, over-completeness, or circuit-replacement guarantees are demonstrably insufficient.

Quick Reference

• Repository: https://github.com/liusida/ica-lens-paper
• Paper: https://arxiv.org/abs/2606.11722
• Project page: https://liusida.github.io/ica-lens-paper/
• Hugging Face Space (hosted explorer): https://huggingface.co/spaces/EEEAILab/ICAExplorer
• Hugging Face artifacts dataset: https://huggingface.co/datasets/sida/ica-lens-paper
• Docs:
  • docs/quickstart.md
  • docs/fit_one_layer_qwen36_27b.md — worked example for fitting ICA on a new model / layer (use as template)
  • docs/reproduction.md
  • docs/troubleshooting.md
  • docs/artifact_contract.md, docs/artifact_datasets.md, docs/model_and_layer_conventions.md

Installation / Setup

Prerequisites

• Python 3.10+
• PyTorch (CUDA recommended for activation capture and ICA fitting)
• uv package manager
• FastAPI-based explorer

Clone the repository (required first — all commands below assume the repo root as cwd)

git clone https://github.com/liusida/ica-lens-paper.git
cd ica-lens-paper

Minimal install (any task)

uv sync

Extra steps required for any SAEBench-touching evaluation (TPP, sparse probe, SAE-overlap)

git submodule update --init --recursive
bash scripts/setup_saebench_envs.sh

The Four Things You Actually Do With This Skill

1. Browse pre-fitted ICA components in the explorer

The released artifacts ship with non-Gaussianity scores, top-activating contexts, ERF entries, and SAE-overlap annotations for GPT-2 Small, Gemma 2 2B, and Qwen 3.5 2B Base. This is the fastest way to answer interpretability questions on those models without re-fitting anything.

uv sync
uv run python scripts/fetch_artifacts.py --models --databases
uv run python scripts/verify_artifacts.py
uv run python -m server.app --port 8001

Switch to the full database when you need token-coloured context examples:

uv run python scripts/fetch_artifacts.py --models --databases --database-variant full
uv run python scripts/verify_artifacts.py --database-variant full
ICA_EXPLORER_DB_PATH=artifacts/fetched/databases/ica_probe_full.sqlite \
uv run python -m server.app --port 8001

2. Fit ICA on a new component (your own model, layer, or hook point)

This is the central reusable workflow. The recipe is the same regardless of model: capture activations at the target hook point, then fit FastICA on them. Use docs/fit_one_layer_qwen36_27b.md as a template — it documents the only place model-specific config matters.

# 1. Capture activations at the chosen hook point(s)
uv run python workflows/01_capture_activations.py \
    --config configs/activations/<your_model>.toml \
    --output-root results/<your_run>/activations \
    --include-embedding

# 2. Fit FastICA on the captured activations
uv run python workflows/02_fit_ica.py \
    --config configs/fit_ica/<your_model>.toml \
    --activation-root results/<your_run>/activations \
    --output-root results/<your_run>/ica

The TOML configs control:

• which hook points / layers are captured (configs/activations/*.toml)
• the token budget and FastICA hyperparameters (configs/fit_ica/*.toml)
• model-tokenizer wiring (configs/models/*.toml)

Author a new TOML pair by copying the closest released model's pair and editing the hook-point list.

3. Score the fitted components

Once ICA components exist, the same downstream protocols apply regardless of which model produced them.

# Non-Gaussianity diagnostics (per-direction interpretability prior)
uv run python workflows/03_compute_nongaussianity.py \
    --models <your_model> --token-budget 1000000 \
    --activation-root results/<your_run>/activations \
    --ica-root results/<your_run>/ica \
    --output-root results/<your_run>/nongaussianity

# Top-activating examples + explorer DB (for browsing / annotation)
uv run python workflows/04_build_explorer_db.py \
    --models <your_model> --token-budget 1000000 \
    --activation-root results/<your_run>/activations \
    --ica-root results/<your_run>/ica \
    --output-db results/<your_run>/databases/ica_probe.sqlite
uv run python workflows/05_populate_erf.py \
    --models <your_model> --token-budget 1000000 \
    --db-path results/<your_run>/databases/ica_probe.sqlite \
    --ica-root results/<your_run>/ica

# Component-level evaluation against trained SAE features
uv run python workflows/06_compare_ica_sae_overlap.py \
    --models <your_model> --token-budget 1000000 \
    --ica-root results/<your_run>/ica \
    --output-root results/<your_run>/ica_sae_overlap

# Targeted Probe Perturbation (TPP) — the headline SAEBench score
uv run python workflows/07_run_saebench_tpp.py \
    --model <your_model> --token-budget 1000000 \
    --ica-root results/<your_run>/ica

# Sparse probing — recoverability of labelled properties from ICA components
uv run python workflows/08_run_saebench_sparse_probe.py \
    --model <your_model> --token-budget 1000000 \
    --ica-root results/<your_run>/ica --methods all

Use these as interpretability instruments, not just as paper-reproduction commands: the per-direction non-Gaussianity score is a cheap prior on which components to inspect first; SAE overlap localises which learned dictionary atoms a given ICA direction stands in for; TPP and sparse probing quantify the direction's causal handle on labelled behaviour.

4. Inspect your own components in the explorer

The explorer reads any SQLite that follows the artifact contract, so you can point it at a self-fitted run while keeping released artifacts available for cross-comparison:

ICA_EXPLORER_DB_PATH=results/<your_run>/databases/ica_probe.sqlite \
ICA_EXPLORER_ICA_ROOT=results/<your_run>/ica \
uv run python -m server.app --port 8001

For a low-cost end-to-end dry run of the entire pipeline (capture → fit → score → explorer DB), use the demo orchestrator before committing GPU time:

uv run python scripts/reproduce_all.py --mode demo --clean --force --erf-limit 1

Programmatic Surface

Construct the FastAPI app in-process

from server.app import create_app

app = create_app()                       # honours ICA_EXPLORER_* env vars
# mount, test with fastapi.testclient.TestClient, or inspect routes

See bundled scripts/explorer_app_usage.py for a runnable demonstration including OpenAPI inspection.

Programmatic artifact-variant selection

from scripts.fetch_artifacts import _with_database_variant

selected = _with_database_variant(artifact_set, "full")

See bundled scripts/artifact_variant_example.py.

Key APIs / Surfaces

Workflow entry points (the operational interface)

Entry point	Component-interpretability role
workflows/01_capture_activations.py	Stream a chosen hook point's activations to disk
workflows/02_fit_ica.py	Recover non-Gaussian directions at that hook point
workflows/03_compute_nongaussianity.py	Rank directions by interpretability prior
workflows/04_build_explorer_db.py	Index top-activating contexts per direction
workflows/05_populate_erf.py	Attach example-receptive-field entries to each direction
workflows/06_compare_ica_sae_overlap.py	Cross-reference ICA directions with trained-SAE features
workflows/07_run_saebench_tpp.py	Causal-handle quality under a perturbation budget
workflows/08_run_saebench_sparse_probe.py	Labelled-property recoverability from ICA components

Core server / scripts surface

Symbol	Module	Description
create_app(settings)	server/app.py	FastAPI factory for the explorer
ProbeRequest, SaeProbeRequest, AnnotationUpdate	server/app.py	Explorer request / annotation models
_model_settings(app, model_name)	server/app.py	Per-model settings resolver
_raw_gpt2_block_index_for_probe(app, model_name, layer)	server/app.py	Layer-id translator
main(argv), _with_database_variant(artifact_set, variant)	scripts/fetch_artifacts.py	Artifact fetch + variant selection
main(argv), _filter_database_checksums(reports, variant)	scripts/verify_artifacts.py	Artifact verification
main(argv), _run_step(report, name, cmd), _clean_generated()	scripts/reproduce_all.py	Pipeline orchestration
main(argv), _selected_models(args, demo_cfg), _run_model_demo()	scripts/run_demo.py	Demo runner

Environment variables

Variable	Effect
ICA_EXPLORER_DB_PATH	Selects which SQLite the explorer reads
ICA_EXPLORER_ICA_ROOT	Selects which fitted ICA directory the explorer reads components from

Released artifacts (use as ready-made components for the three covered models)

Model	Activations config	Fit-ICA config	Released artifact path
GPT-2 Small	configs/activations/gpt2.toml	configs/fit_ica/gpt2.toml	artifacts/fetched/models/gpt2/
Gemma 2 2B	configs/activations/gemma2_2b.toml	configs/fit_ica/gemma2_2b.toml	artifacts/fetched/models/gemma2_2b/
Qwen 3.5 2B Base	configs/activations/qwen3_5_2b_base.toml	configs/fit_ica/qwen3_5_2b_base.toml	artifacts/fetched/models/qwen3_5_2b_base/
Qwen3.6-27B	worked-example template — see docs/fit_one_layer_qwen36_27b.md	same	not pre-shipped

For any other model, follow the Qwen3.6-27B doc as the template.

Related Tools

ICA Lens sits inside the feature-dictionary-learning family as the dictionary-free baseline. Pair it with the other family members along a cost–fidelity gradient:

• SAELens (../SAE/) — when over-complete coverage and per-feature monosemanticity are required at the same site.
• Transcoder circuits (../transcoder/) — when an interpretable input → output mapping is needed to slot into circuit-replacement / attribution graphs.
• Crosscoders (../crosscoder/) — when a shared basis across layers or fine-tuning checkpoints is the object of interest.
• SAEBench (submodule, shared infrastructure) — supplies the TPP and sparse-probe evaluators used by workflows/07 and workflows/08; ICA-Lens results report on the same axes that SAE / transcoder evaluations use, so cross-comparison is one-to-one.

Common Patterns & Best Practices

1. Pick the site before the method. ICA's quality depends entirely on the activation distribution at the chosen hook point — residual stream, MLP output, attention-head output, post-LN, etc. Decide where the interpretability claim lives, then capture exactly that hook point.
2. Always start from released artifacts for GPT-2 / Gemma 2 2B / Qwen 3.5 2B Base; only re-fit when extending to a new model, layer, or hook point.
3. Use non-Gaussianity as the first filter when triaging hundreds of components — rank by workflows/03 output before opening the explorer.
4. Score ICA and any candidate SAE on the same protocol (TPP, sparse probe, SAE overlap) before claiming one is "more interpretable" than the other.
5. Mini DB by default for exploration; full DB only when token-coloured contexts are needed.
6. Demo before full reproduction — scripts/reproduce_all.py --mode demo validates the orchestration end-to-end before committing GPU time.
7. Config-driven everything — for a new model / layer, copy and edit a TOML pair (configs/activations/*.toml + configs/fit_ica/*.toml) rather than passing flags.
8. Respect output layout — fetched data under artifacts/fetched/, generated data under results/; the explorer assumes this split.
9. Use the env vars (ICA_EXPLORER_DB_PATH, ICA_EXPLORER_ICA_ROOT) to mix released and self-fitted artifacts in one explorer instance.
10. Don't claim circuit-level mechanism from ICA alone — ICA produces correlated, non-causal directions. Promote a candidate to a causal claim with targeted perturbation (workflows/07) or by switching to a transcoder for replacement-context circuit work.

Demo Scripts

scripts/explorer_app_usage.py

#!/usr/bin/env python3
"""
Programmatic usage example for the ICA Lens FastAPI explorer.

Demonstrates importing `create_app`, building the FastAPI app, inspecting
registered routes, and fetching the generated OpenAPI schema via TestClient.

    uv sync
    uv run python scripts/explorer_app_usage.py
"""

from __future__ import annotations

import json
import sys
from pathlib import Path
from typing import Any

try:
    from fastapi.testclient import TestClient
except ImportError as exc:
    raise SystemExit(
        "fastapi is required to run this example. "
        "Install project dependencies with `uv sync` first."
    ) from exc


def ensure_repo_root_on_path() -> Path:
    repo_root = Path(__file__).resolve().parents[1]
    repo_root_str = str(repo_root)
    if repo_root_str not in sys.path:
        sys.path.insert(0, repo_root_str)
    return repo_root


def build_app():
    from server.app import create_app
    return create_app()


def summarize_routes(app: Any) -> list[dict[str, Any]]:
    route_summaries: list[dict[, ]] = []
     route  (app, , []):
        methods = ((route, , [])  [])
        path = (route, , )
        name = (route, , )
        route_summaries.append({: name, : path, : methods})
     route_summaries


 () -> [, ]:
     TestClient(app)  client:
        response = client.get()
        response.raise_for_status()
         response.json()


 () -> :
    repo_root = ensure_repo_root_on_path()
    ()
    app = build_app()
    ()
    routes = summarize_routes(app)
    ()
    schema = fetch_openapi_schema(app)
    ()
     


 __name__ == :
     SystemExit(main())

scripts/artifact_variant_example.py

#!/usr/bin/env python3
"""
Programmatic usage example for ICA Lens artifact variant selection.

Demonstrates `_with_database_variant(artifact_set, variant)` from
`scripts.fetch_artifacts`, used to switch a manifest between the mini and
full SQLite explorer databases without running the full download flow.

    uv sync
    uv run python scripts/artifact_variant_example.py
"""

from __future__ import annotations

import json
import sys
from pathlib import Path
from typing import Any


def ensure_repo_root_on_path() -> Path:
    repo_root = Path(__file__).resolve().parents[1]
    repo_root_str = str(repo_root)
    if repo_root_str not in sys.path:
        sys.path.insert(0, repo_root_str)
    return repo_root


def build_sample_artifact_set() -> dict[str, Any]:
    return {
        "models": [
            {"name": "gpt2", "path": "artifacts/fetched/models/gpt2/"},
            {"name": "gemma2_2b", "path": "artifacts/fetched/models/gemma2_2b/"},
            {"name": "qwen3_5_2b_base", "path": "artifacts/fetched/models/qwen3_5_2b_base/"},
        ],
        "databases": {
            "mini": {"path": "artifacts/fetched/databases/ica_probe_mini.sqlite"},
            : {: },
        },
    }


 () -> :
    ensure_repo_root_on_path()
     scripts.fetch_artifacts  _with_database_variant

    artifact_set = build_sample_artifact_set()
     variant  (, ):
        selected = _with_database_variant(artifact_set, variant)
        ()
        (json.dumps(selected, indent=, sort_keys=))
     


 __name__ == :
     SystemExit(main())

Attribution

• Sida Liu — Independent Researcher — [e-mail]
• Feijiang Han — University of Maryland — [e-mail]

License: MIT.