Historical report; not the current benchmark. This document retains original claims and scores, some of which were corrected or withdrawn. Use the current study and research record for the final interpretation.
Professor Feedback Analysis (annotated PDF) + Persistence in the New Paper
Annotated document: Quintanilha and Virnelli - Dual-Channel Contrastive Encoding of Sperm Whale Coda.pdf (7 pages, dated May 8 2026).
Important: this is a third version of the paper — a human-rewritten draft whose prose differs from both (a) the committed main.tex ("Beyond WhAM…") and (b) our new rewrite ("Two Channels, One Representation…"). So some marks target prose that no longer exists; the conceptual concerns, however, carry over and are what matter. The professor was also reading the old individual-ID-centric thesis (DCCE "+83% over WhAM"), which we have since reframed.
Legend for status: RESOLVED (fixed in new paper) · PARTIAL (improved, needs a touch) · PERSISTS (still an issue) · DECISION (needs your call) · PROSE-ONLY (was specific to the marked draft).
1. Complete catalog of the professor's annotations
Page 1 (highlights)
- Yellow: "whether biological inductive bias can outperform raw data scale"; "vowel patterns"; "VampNet-based"; "generative modeling"; "being able to identify the certain type of clicks released [3]" (awkward phrasing); "around 10 times larger [4]" (factual error).
- Red: "the central question:" (redundant with "raising the question"); "our Phase 0 EDA … spectral centroid and ICI features" with "ICI" boxed; "adding a 'vowel-like' channel"; "current state, which is WhAM," (awkward).
Page 2 (heavy handwriting)
- "no labels exist." underlined + big "?", with boxed note: "I thought you know the 22 coda types of the data from DSWP. Or are the labels from the Sharma et al.?" → contradiction: claims no labels but names 22 coda types / 3 units.
- "the team" highlighted → "which team?" (should be "we").
- "where does the codaNUM come from?";
codaNUM2018underlined + "?" → identifier never defined. - Table 1: "IDN-labeled" column circled + "?", and "↓ what is that?" on "762 IDN-labeled codas across 12 individuals" → IDN concept + the 762/12 figures unexplained; why A/D fully labeled but F only 200.
- Vowel-labels note: "Using a CNN for a diff task? meaning we are NOT doing a vowel classification task due to lack of GT?" → wants the spectral-encoder motivation explicit.
- Underlines: "3 social units", "22 coda types", "noise-flagged", "Second/temporal confound", "Third/class imbalance".
Page 3 (heavy handwriting)
- Top margin: "we should clearly define the tasks in main text prior to Fig. 3."
- Rhythm Encoder: "how did you obtain the ICI vector?"
- §3.2: "Contrastive learning should be explained, how and why."
- Fig. 4 caption: "we need to clearly define terminologies, this confuses me, especially" (on "speaker identity (social unit)").
- Right col: "I thought most data are from the same social unit" → doesn't follow how same-unit positive pairs work when Unit F dominates.
- Table 2: "maybe a clearer term like contrastive" on the "Cross-channel" column.
- Task list: "which task is the contrastive part helping?"
- §4.1: "Why record year? what does it show?"
- §4.1.1: full-red highlight + "duplicate" → the "We extract intermediate representations…" sentence is copy-pasted twice.
Page 4
- Red checkmarks (approval) on the "DCCE inherently less susceptible" paragraph and the §4.1.3 Recording-Year Confound heading.
- Table 3: circles on 0.895 / Coda-Type column / 0.834 + "explain a bit."
- §4.2.2: "terminology"; "speaker-identity separation" underlined.
Page 5
- Fig. 7 (UMAP): "maybe in appendix if no space."
- §4.3 Synthetic Augmentation: "Need more clarification. In this case, do we still keep this section?"
Page 6–7
- No handwriting. (Observed independently: References 13 & 14 are duplicate Goldwasser entries; two Beguš entries, one a GitHub placeholder.)
2. Persistence mapping (does it still apply to the new paper?)
| # | Concern | Status | Action |
|---|---|---|---|
| 1 | "no labels exist" contradicts naming 22 coda types | PERSISTS — main.tex L112 still says "no labels exist" |
Change to "no labels ship with the DSWP audio release"; state up front that all labels are recovered from Sharma's DominicaCodas.csv (so the 22 types are consistent, not contradictory) |
| 2 | "10 times larger" factual error | RESOLVED — now "≥143 combinations, roughly doubling information" (L92) | none |
| 3 | codaNUM2018 undefined |
PERSISTS — used L119 without definition | Add one line: "codaNUM2018 is the coda-index column in DominicaCodas.csv" |
| 4 | IDN-labeled unclear; why A/D full, F partial; 762/12 | PARTIAL — caption explains IDN=0 = "not photo-identified, mostly Unit F"; could define the IDN acronym and the 12-individual subset explicitly | Define "IDN (individual identification number)"; one line on the 762/12 subset |
| 5 | How was the ICI vector obtained? | PARTIAL — says "pre-computed ICI sequences" (from Sharma) but not crisply | State explicitly: the ICI vector is the pre-computed inter-click-interval column from DominicaCodas.csv, zero-padded to length 9 |
| 6 | Define the 3 tasks before the architecture figure | PARTIAL — tasks named in abstract/intro; no crisp "we study 3 tasks" up front | Add one sentence early in §Method/§Setup enumerating the 3 tasks + class counts before the TikZ figure |
| 7 | "Cross-channel" term unclear | PERSISTS (term kept) | Either rename or add a one-line definition at first use |
| 8 | Terminology (speaker/identity/social unit) inconsistent | MOSTLY RESOLVED — consistent "social unit / individual identity / dialect"; only 2 stray "speaker" uses (L270, L439) | Replace the 2 "speaker" mentions for consistency |
| 9 | Contrastive learning: explain how & why | IMPROVED — NT-Xent + cross-channel pairs explained w/ figure | Add one sentence on why contrastive (open-set identity / no fixed classes) |
| 10 | CNN for a different task, not vowel classification | RESOLVED — explicit note (L121) that lack of vowel GT motivated a self-supervised CNN, not vowel classification | none |
| 11 | Which task does the contrastive part help? | RESOLVED — §Ablations now states it helps social unit (+0.222) but barely ID (+0.009) | none (this is now a strength) |
| 12 | Why recording year? what does it show? | RESOLVED/PARTIAL — framed as a confound diagnostic and the confound is quantified | Add half a sentence up front stating why we probe it |
| 13 | "explain a bit" — interpret the results table | RESOLVED — tiered table + interpretation text added | none |
| 14 | Cross-channel pairing when most data is one unit | PERSISTS — not directly addressed | Add a sentence: WeightedRandomSampler balances units per batch; positive pairs are same-unit while negatives are other units, so the objective is not degenerate under F-dominance |
| 15 | "the team" → "we" | RESOLVED — new paper uses "we" throughout (0 "the team") | none |
| 16 | Awkward phrasings (p.1) | PROSE-ONLY / RESOLVED — fresh prose in the rewrite | proofread pass |
| 17 | Duplicate "We extract…" paragraph | RESOLVED — appears once | none |
| 18 | Duplicate references 13/14; placeholder Beguš | RESOLVED — rho.bib has one Goldwasser entry; Beguš entry fixed to the real Open Mind 2025 paper |
none |
| 19 | UMAP figure → appendix if no space | RESOLVED — UMAP figure removed; two-channel scatter is the main figure | none (or re-add to appendix if wanted) |
| 20 | Synthetic augmentation: keep the section? | DECISION — kept as a short section; now tangential to the two-channel thesis | Decide: keep short, move to appendix, or cut |
| 21 | Recording-year confound (approved) | KEPT — professor liked it | keep |
3. Bottom line
- The rewrite already resolves most of the professor's marks — the two big factual/writing errors (10× larger; duplicate paragraph), the "the team" voice, the UMAP space issue, the "which task does contrastive help" question (now a headline result), and the CNN-vs-vowel-classification motivation.
- Genuinely still-open items (worth fixing): #1 the "no labels exist" contradiction, #3 codaNUM2018 definition, #5 ICI-vector provenance, #7 the "cross-channel" term, #14 the same-unit-pairing-under-imbalance concern, and #6 a crisp task definition up front. These are small, targeted edits.
- One decision: #20 whether the synthetic-augmentation section stays.
- Caveat: the professor has not seen the new two-channel framing. It answers several of his implicit questions, but he will likely have fresh comments on the reframed thesis.