Haotian Ma

PhD Candidate @WisconsinCS. Founder & Maintainer of OpenSAE. Interpretability & AI Safety.

[Haotian Ma @AtheavenMa]

🎉#NeurIPS2026: We can identify important regions and spectral channels. But which channels matter in each region?

We introduce Spectral-Spatial Interpretation (SSI) to explain spectral and spatial information jointly.

🤔 The core problem

In multispectral and hyperspectral images, the same channel can contribute differently across spatial regions.

Most attribution methods can produce spatial maps or channel importance scores, but aggregating either dimension can obscure how the model combines spectral and spatial information.

⚙️ How SSI works

SSI uses Shapley values to explain each spatial location while preserving its surrounding image context.

→ Divide the image into spatial cells.

→ Keep all other cells at their observed values.

→ Within each cell, use Shapley values to distribute the score change among its spectral channels.

This produces a signed attribution for every spatial cell and channel, allowing us to examine where individual channels support or oppose a prediction.

🔍 Why this matters

The key is conditioning on the actual surrounding image rather than averaging over different configurations of the external context.

Within each cell, the channel attributions sum exactly to the score difference caused by masking that cell.

SSI therefore provides a local explanation of spectral contributions while preserving the spatial context in which they occur.

📈 Main results

Across CUB-200, EuroSAT, and HLS Burn Scars, SSI achieves the best reported mean joint deletion and insertion AUC among IG, DeepSHAP, KernelSHAP, and RISE under both mean and zero masking baselines.

On HLS Burn Scars with the mean baseline:

→ Joint deletion AUC ↓: 0.2096 → 0.1545

→ Average spatial mIoU ↑: 0.2584 → 0.3555

Both comparisons use the strongest non-SSI method for the respective metric.

We also released an interactive Playground with real examples across six datasets, covering RGB, multispectral, and hyperspectral imagery. You can explore real model explanations and see how different channels contribute across an image.

📄 Paper: https://t.co/Eb1QOfUlFC

💻 Code: https://t.co/1oCWk4X4Pp

🎛️ Playground: https://t.co/WxJtrqVR3O

🤗 Hugging Face: https://t.co/IzasT6guFr

Grateful to @Candice08063476 and @PhilTownsnd for the collaboration!

[Haotian Ma @AtheavenMa]

@hthieblot Intelligence as currency.

[Haotian Ma @AtheavenMa]

@SuJinyan6 谢谢！感觉如果能拿到一个offer确实会好很多

[Haotian Ma @AtheavenMa]

Hi Ran! I'm a PhD candidate at UW-Madison, working on interpretability. I'm also an epic fantasy writer and finished my first book last year. I used AI quite a bit for visual design, which made me believe that AI can be a bridge to a meaningful life. I really resonate with your ideas and would love to connect!

[Haotian Ma @AtheavenMa]

@KobeissiLetter I forget who said.... model training costs are not real costs.

[Haotian Ma @AtheavenMa]

@QuantumTransf 确实啊，组织agent不是那么容易的事情

[Haotian Ma @AtheavenMa]

@Fei2411 我猜是因为训练方式和侧重点不同。如果你用cursor经常看它执行问题的思路的话，会发现如果它不需要思考的话，还是很得力的(所以决策权要别的来

[Haotian Ma @AtheavenMa]

@TrainCodeMan Yes maybe 3-5 years later. But human still need some weak and safe AI for their personal goal

[Haotian Ma @AtheavenMa]

At this stage, building a harness should be only about three things:
1. Building things no one has built yet
2. Finding better ways to organize agents
3. Learning as much as possible, staying in close touch with the community, and getting feedback

Optimizing a harness for a particular task cuts the value of the work by 90%.

Optimizing it to get good results on that particular task cuts it by 99%.

And if you need the agent to decide what a "good result" even means, you're left with maybe 0.01% of the value, and yet it still feels like I'm getting work done.

(until all harnesses can be transferred efficiently…

[Haotian Ma @AtheavenMa]

@nickcammarata @Sauers_ It doesn’t surprise me. There are so many features that can make great effect. But if one day, without feature steering or intervention, a model says something strange…… Then it would definitely raise my concerns.

[Haotian Ma @AtheavenMa]

@TrainCodeMan Hi! I build harness for long horizon research. I believe human and AI should work very closely.

[Haotian Ma @AtheavenMa]

@nrbsld237 @Samsung_RA Hi Taeyun, happy to connect! I’m also working on agent memory~

[Haotian Ma @AtheavenMa]

@poteto @poteto , If you build a community and let people contribute, then many researchers, engineers and creators would like to their best! Please, let us know how to contribute.

We can do broad practices beyond large company.

[Haotian Ma @AtheavenMa]

🤔For publishing research papers, which one do you prefer?

A. strong agent + YOLO run
B. strong harness + Human in the loop
C. 50% agent + 50% human
D. I can combine them very vell ( 1+1 > 2)

[Haotian Ma @AtheavenMa]

@SharonYixuanLi At UW–Madison, you get to see how excellent researchers like Sharon think about research. I once sat in on her course and gained so many valuable insights and ideas. Perhaps this is the certainty I can find among so much uncertainty.

[Haotian Ma @AtheavenMa]

@SharonYixuanLi At UW–Madison, you get to see how excellent researchers like Sharon think about research. I once sat in on her course and gained so many valuable insights and ideas. Perhaps this is the certainty I can find among so much uncertainty.

[Sharon Li @SharonYixuanLi]

I know many PhD students are questioning what their paths mean as AI accelerates scientific discovery. That anxiety deserves to be taken seriously.

I wrote some reflections in a recent blog post. My students found them helpful, so I’m sharing them more widely.

I hope it offers a little light as we navigate this strange time. Please share it with someone who might need it.

https://t.co/wr1upnki47

[Haotian Ma @AtheavenMa]

@itz_komal01 Because we didn’t make good visualizations yet

[Haotian Ma @AtheavenMa]

@summer_ikiki_ Maybe weekly or even daily.

[Haotian Ma @AtheavenMa]

@harris_p10 Can’t agree more