FEATURE_MAP: variable counts per group, level table, and one-page .docx

- Add (N) variable counts to every feature-type entry that lacked one
  (image stats 12, acoustics 34, faces 7, saliency/depth 8, etc.).
- New "Feature summary table" organizing groups by class x level
  (low / mid / high-level).
- docs/feature_summary_table.docx: one-page landscape Word table for
  Google Docs import, generated by tools/build_feature_summary.py
  (counts pulled from schema/channel_template.json).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: torwager

docs/FEATURE_MAP.md

@@ -20,6 +20,27 @@ The map is a plain, flat **SVG** built for editing:
   python3 tools/build_feature_map.py     # reads schema/channel_template.json
   ```
 
+## Feature summary table
+
+Feature groups organized by class (row) and by level of abstraction (column). Numbers in
+parentheses are the count of **variables** in that group — each scalar/flag counts as 1
+and each vector counts as its dimensionality (so an embedding contributes many). "Level"
+is the standard perception hierarchy: **low-level** = raw physical/perceptual signal,
+**mid-level** = perceptual organization/structure, **high-level** = semantic/conceptual.
+
+| Feature class | Low-level | Mid-level | High-level |
+|---|---|---|---|
+| **Visual** | image statistics (12); optical-flow motion (3); shot cuts (2) | saliency & depth (8); faces & bodies (7) | SigLIP semantics — embedding + scene probes (784); DINOv2 embedding (384); action recognition (400) |
+| **Audio** | acoustics — loudness / pitch / timbre / MFCC / chroma (34) | speech presence & rate (2) | AudioSet sound tags (527); CLAP semantics — embedding + probes (524); vocal affect V/A/D (3) |
+| **Language** | word frequency & length (2) | lexical norms — valence / arousal / dominance, concreteness, AoA (5); syntactic complexity (5) | word surprisal & next-word entropy (2) |
+| **Social** | — | agent count & closest-pair distance (2) | interaction type & social dominance (2) |
+| **Situation** | — | — | scene description / setting / indoor–outdoor (3); event structure — boundaries & event id (2) |
+| **Affect** | — | facial affect — 8 expressions + valence + arousal (10) | EmoNet image emotion (20); text emotion & sentiment (32); VLM depicted emotion (3) |
+
+> A one-page Word version of this table (for import into Google Docs) is at
+> [`docs/feature_summary_table.docx`](feature_summary_table.docx), regenerated by
+> `python3 tools/build_feature_summary.py`.
+
 ## The major feature types
 
 What each model in the pipeline is, what it outputs, why it is useful (with an eye to
@@ -29,7 +50,7 @@ copies live in `PDFs/` (Dropbox-only).
 
 ### Visual — what the frame looks like
 
-- **Low-level image statistics** — Classic per-frame pixel/color measures computed with
+- **Low-level image statistics (12)** — Classic per-frame pixel/color measures computed with
   [scikit-image](https://doi.org/10.7717/peerj.453): luminance, RMS contrast, mean red/
   green/blue and hue/saturation/brightness, edge density, image entropy, the slope of the
   spatial-frequency (Fourier) power spectrum, and perceptual
@@ -62,25 +83,25 @@ copies live in `PDFs/` (Dropbox-only).
   probability vector plus the single top action label per second — a semantic
   "what are people doing" regressor for action-observation and social/motor brain systems.
 
-- **Motion (optical flow)** — [RAFT](https://arxiv.org/abs/2003.12039) estimates dense
+- **Motion — optical flow (3)** — [RAFT](https://arxiv.org/abs/2003.12039) estimates dense
   per-pixel motion between consecutive frames. From it the pipeline derives overall
   optical-flow magnitude, a camera-motion estimate, and residual (object) motion. Low-
   level motion energy is a classic driver of MT/MST and dorsal-stream responses and a
   standard motion regressor in naturalistic fMRI.
 
-- **Shot cuts & shot index** — A histogram-based cut detector (a lightweight stand-in for
+- **Shot cuts & shot index (2)** — A histogram-based cut detector (a lightweight stand-in for
   [TransNetV2](https://arxiv.org/abs/2008.04838)) flags edit boundaries and numbers the
   shots. Cuts are strong low-level visual-transient events and useful for modeling
   attention/event-onset responses.
 
-- **Faces & bodies** — [MTCNN](https://arxiv.org/abs/1604.02878) detects faces (count,
+- **Faces & bodies (7)** — [MTCNN](https://arxiv.org/abs/1604.02878) detects faces (count,
   detection confidence, largest-face size, face-present flag) and
   [Keypoint R-CNN](https://arxiv.org/abs/1703.06870) (from the Mask R-CNN family) detects
   human body pose (person count, pose confidence, body-present flag). Face presence/size
   is a primary regressor for face-selective cortex (FFA/OFA/STS); body/person presence
   drives EBA and the person-perception network.
 
-- **Saliency & depth** — A spectral-residual [saliency](https://ieeexplore.ieee.org/document/4270292)
+- **Saliency & depth (8)** — A spectral-residual [saliency](https://ieeexplore.ieee.org/document/4270292)
   model (Hou & Zhang 2007; stand-in for a deep video-saliency net) gives where the eye is
   likely drawn (mean/peak saliency, spread, salient-area fraction), and
   [Depth-Anything-V2](https://arxiv.org/abs/2406.09414) gives monocular depth (mean depth,
@@ -89,7 +110,7 @@ copies live in `PDFs/` (Dropbox-only).
 
 ### Audio — what the soundtrack sounds like
 
-- **Low-level acoustics** — [librosa](https://doi.org/10.25080/Majora-7b98e3ed-003)
+- **Low-level acoustics (34)** — [librosa](https://doi.org/10.25080/Majora-7b98e3ed-003)
   computes loudness (RMS), spectral shape (centroid, bandwidth, rolloff, flatness),
   zero-crossing rate, onset strength, a 13-value **MFCC** timbre vector, and a 12-value
   **chroma** (musical pitch-class) vector; pitch/F0 uses
@@ -112,7 +133,7 @@ copies live in `PDFs/` (Dropbox-only).
   pipeline keeps the 512-d embedding and 12 interpretable sound probes. The embedding is a
   general auditory-semantic feature space for encoding models.
 
-- **Speech** — [faster-whisper](https://arxiv.org/abs/2212.04356) (Whisper) transcribes
+- **Speech (5)** — [faster-whisper](https://arxiv.org/abs/2212.04356) (Whisper) transcribes
   speech and yields a speech-present flag and speaking rate (words/s);
   [audEERING wav2vec2](https://arxiv.org/abs/2203.07378) reads **vocal affect** directly
   from the voice as continuous valence, arousal, and dominance. Speech presence and
@@ -121,49 +142,50 @@ copies live in `PDFs/` (Dropbox-only).
 
 ### Language — what is said (from the transcript)
 
-- **Lexical semantic norms** — per word, human-rated norms are looked up: emotional
+- **Lexical semantic norms (5)** — per word, human-rated norms are looked up: emotional
   [valence, arousal, dominance](https://doi.org/10.3758/s13428-012-0314-x) (Warriner 2013),
   [concreteness](https://doi.org/10.3758/s13428-013-0403-5) (Brysbaert 2014), and
   [age of acquisition](https://doi.org/10.3758/s13428-012-0210-4) (Kuperman 2012). These
   are the classic word-property regressors for language and semantic fMRI.
 
-- **Word frequency & length** — [wordfreq](https://doi.org/10.5281/zenodo.7199437) gives
+- **Word frequency & length (2)** — [wordfreq](https://doi.org/10.5281/zenodo.7199437) gives
   each word's log frequency (Zipf scale) and its length. Frequency is a robust predictor
   of lexical-access difficulty and reading/listening time.
 
-- **Surprisal & next-word entropy** — [GPT-2](https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf)
+- **Surprisal & next-word entropy (2)** — [GPT-2](https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf)
   provides each word's **surprisal** (how unexpected it was, in bits) and the entropy of
   the model's next-word prediction. Word-by-word surprisal from language models is one of
   the most reliable predictors of language-network and N400-type responses in
   naturalistic-comprehension fMRI/MEG.
 
-- **Syntax** — [spaCy](https://spacy.io) parses each utterance for content-word fraction,
+- **Syntax (5)** — [spaCy](https://spacy.io) parses each utterance for content-word fraction,
   noun/verb fraction, parse-tree depth, and mean dependency distance — proxies for
   syntactic complexity and processing load.
 
 ### Social — who is present and how they relate
 
-- Derived from the video by the [Qwen2.5-VL](https://arxiv.org/abs/2502.13923) vision-
-  language model (setting, interaction type, social dominance) plus face/pose detectors
+- **Social relations (4)** — derived from the video by the [Qwen2.5-VL](https://arxiv.org/abs/2502.13923)
+  vision-language model (interaction type, social dominance) plus face/pose detectors
   ([MTCNN](https://arxiv.org/abs/1604.02878) agent count;
   [Keypoint R-CNN](https://arxiv.org/abs/1703.06870) closest-pair distance). These target
   the social-perception network (STS, TPJ, mentalizing regions) that tracks the number of
   agents, their proximity, and the type of interaction.
 
 ### Situation — where/when; event structure
 
-- **Scene & setting** — the [Qwen2.5-VL](https://arxiv.org/abs/2502.13923) VLM produces a
+- **Scene & setting (3)** — the [Qwen2.5-VL](https://arxiv.org/abs/2502.13923) VLM produces a
   free-text scene description, an indoor/outdoor label, and a coarse setting label per
   second — high-level context for scene/place-selective cortex.
-- **Event boundaries** — [GSBS](https://doi.org/10.1016/j.neuroimage.2021.118085)
+- **Event boundaries (2)** — [GSBS](https://doi.org/10.1016/j.neuroimage.2021.118085)
   (Greedy State Boundary Search; Geerligs 2021) segments the annotation time series into
   discrete "events," emitting boundary events and a running event number. This directly
   operationalizes event-segmentation theory and models the boundary responses seen in
   hippocampus and default-mode regions during naturalistic viewing.
 
 ### Affect — depicted emotion across channels
 
-Emotion is read from four independent channels so it can be cross-validated by modality:
+Emotion is read from several independent channels/modalities so it can be cross-validated
+(facial, textual, and holistic-scene here; vocal affect lives under Audio → Speech):
 
 - **EmoNet — image emotion (20)** — [EmoNet](https://www.science.org/doi/10.1126/sciadv.aaw4358)
   (Kragel et al. 2019, *Science Advances*) is an AlexNet-based CNN that maps a video frame
@@ -172,7 +194,7 @@ Emotion is read from four independent channels so it can be cross-validated by m
   human ratings and brain responses, making it a purpose-built regressor for visual
   emotion-category coding in the brain.
 
-- **Facial affect (8 + valence/arousal)** —
+- **Facial affect (10: 8 expressions + valence + arousal)** —
   [HSEmotion](https://arxiv.org/abs/2403.11590) (Savchenko) reads faces detected in the
   video and classifies expression into **8 categories** (anger, contempt, disgust, fear,
   happiness, neutral, sadness, surprise) plus continuous facial valence and arousal —
@@ -184,7 +206,7 @@ Emotion is read from four independent channels so it can be cross-validated by m
   RoBERTa scores **3-way sentiment** (negative/neutral/positive) with a continuous polarity.
   These capture the emotional content of what is said.
 
-- **VLM depicted emotion** — [Qwen2.5-VL](https://arxiv.org/abs/2502.13923) also judges the
+- **VLM depicted emotion (3)** — [Qwen2.5-VL](https://arxiv.org/abs/2502.13923) also judges the
   overall depicted emotion, valence, and arousal of the scene as a holistic cross-check on
   the specialized channels.