acceptodds
Withdrawn from ICLR 2027

What Improves When a Microscopy Encoder Is Updated?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Microscopy encoders support scientific measurements through classifiers, segmentation decoders and measurement rules. We study which update gains survive the choice to co-train a head, leave it unchanged, or fit it afresh to frozen features. In matched dense-encoder forks across ten seeds, absolute task weighting leads normalization by 4.92 percentage points of culture balanced accuracy after 512 updates through co-trained heads and 4.22 through unchanged source heads; fresh-head contrasts remain unresolved. The preference is therefore measurable as compatibility with an existing classifier. The co-trained–fresh interaction persists under fresh-optimizer and separate-clipping controls, but equal-budget per-arm learning-rate tuning removes the early absolute advantage. Later heads diverge in recipe ranking without meeting the prespecified criterion for a co-trained head hiding resolved accuracy loss. DINOv2-small does not reproduce the early preference at its default rates. SCOPE, a completed 4B-parameter microscopy model, supplies complementary positive evidence: an unchanged classifier improves from 0.6235 to 0.9798 on terminal features, and fresh fits also improve on the previously supervised culture task. On 23 held-out Allen plates, brightfield structure Dice is 0.312, exceeding a 37M-parameter scratch control by 0.056; most of this capability predates continuation. Spatial controls show how decoder capacity and feature access change the update contrast. On held-out larvae, improved counts coexist with reduced strict object overlap, while a validation-based retention gate has no consistent advantage on external reporter assays. Most follow-ups reuse inspected panels. The study makes encoder progress interpretable through the complete use: which readout changes, which labels and optimization it receives, and which biological measurement the resulting procedure supports.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
representation learning, multitask learning, representation evaluation, encoder adaptation, microscopy, bioimage analysis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_f8QqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper asks how a microscopy encoder could be improved. It looks like the paper motivates the idea of how updated can change what the model can do. The paper is not clear at all so it is difficult to provide summary. The paper does the evaluation on SCOPE, a 4 billion parameter model which is not cited or clearly well explained.

Critical strengths

None

Critical weaknesses
  1. The paper is poorly written especially the abstract which never explains the problem at a high level and clearly state the findings. There is no related work section so the paper does not even attempt to showcase what research question they are trying to answer and how it fits into current contributions of the field.
  2. There is no clear conclusion or what the community can learns provided by the paper at the end. It is just a bunch of analysis.
  3. The SCOPE model is not clearly explained and provided details for. It is an integral part of the results but it is hard to parse what it was pre-trained with and what task does it tackle? This weakness makes half of the paper's results hard to interpret and read.
  4. Is the scope of the paper in segmentation of microscopy images then why is DINOv2 being compared? Instead, the comparisons should be against microscopy specific foundation models for segmentation.
Questions for authors
  1. What is the research question that the authors try to answer? What is the motivation for solving this problem?
  2. Can the author(s) provide more details about the SCOPE model?
  3. What is the takeaway of this research paper that the reader should comprehend?
  4. Is the paper focusing on segmentation related tasks (Figure 1 shown) or is it related to microscopic image-based representation learning as they use DINOv2 as well?
Reviewer_WseBOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates what kind of encoder updates, namely using fixed features with a new prediction head, using fixed features with existing prediciton head, or co-training both the encoder and the head is best for downstream performance.

The authors have used 3 types of encoders for this study. First a "dense" encoder, second a DINOv2-small, and third an encoder named SCOPE. The evaluation protocols are classifiying cell culture, segmentation on brightfield images and few others which are difficult to decipher.

The paper shows the trend that overall performance reaches a point of diminishing returns when the pre-trained encoder is stronger, regardless of the update. For weaker pre-trained checkpoints such as the "dense" models, the type of update matters. Co-training works best there.

Critical strengths

Experimental procedure: The authors have conducted a large set of experiments to validate their results. The results seem to validate existing knowledge on general domains, wherever the pre-trained model is strong, with less parameters to be updated.

Critical weaknesses

Unfortunately, the paper is impossible to read. There is no coherence between sentences, there are sentences without context popping up in the middle of paragraphs, and overall there is no story. If the paper was decently written, the reader might be able to gain some insights. However, at this stage of the paper, it is very difficult.

For example, most of the terminology have not been defined in text, and this would make it impossible for anyone within the general machine learning / AI domain (as is the case for a reader in this type of conferences), or even within the medical AI domain to understand this paper.

Finally, I want to mention that the the low confidence is not due to a lack of familiarity with the topic, but rather to difficulties in understanding the written content.

Questions for authors

No Questions.

Reviewer_H2TXOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates how different encoder update strategies affect downstream performance in microscopy image analysis. It compares absolute task weighting and normalization under different classifier head configurations, including co-trained, unchanged, and freshly trained heads. Experiments across multiple microscopy tasks and encoder architectures suggest that the benefits of encoder updates depend on classifier compatibility, optimization settings, and downstream evaluation procedures.

Critical strengths

The paper investigates an interesting problem concerning the relationship between encoder updates and downstream performance in microscopy image analysis, considering different classifier configurations and evaluation tasks.

Critical weaknesses
  1. The proposed loss function is complicated, but most of its terms are not sufficiently explained. The motivation and specific contribution of each term remain unclear, making it difficult to understand the optimization objective.

  2. The relationship between Equation (1) and Equation (3) is unclear. In addition, the meanings of several terms and mathematical symbols in Equation (3) are not properly defined or explained.

  3. The overall writing and presentation are difficult to understand. The technical descriptions are complicated, and the organization of the methodology lacks clarity, making it challenging to follow the main ideas and contributions.

Questions for authors
  1. What is the specific role of each term in the overall loss function, and how do these terms contribute to the optimization process?

  2. What is the mathematical relationship between Equation (1) and Equation (3)?

  3. Could the authors clearly define all variables, symbols, and terms appearing in Equation (1) and (3)?

  4. Could the authors provide a clearer and more structured explanation of the overall methodology and its main contributions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.