Yusuf Meric Karadag (@attentionMatrix) header
Yusuf Meric Karadag (@attentionMatrix) profile picture

incoming msc cs @eth, bsc ceng @METU_ODTU, prev @ExplainableML

Joined May 2022
Very happy to share that my first conference paper, โ€œ๐—–๐—•๐—ซ-๐—•๐—ฒ๐—ป๐—ฐ๐—ต: ๐—” ๐—›๐˜‚๐—บ๐—ฎ๐—ป-๐—”๐—น๐—ถ๐—ด๐—ป๐—ฒ๐—ฑ ๐— ๐—Ÿ๐—Ÿ๐—  ๐—–๐—ผ๐˜‚๐—ป๐—ฐ๐—ถ๐—น ๐—ณ๐—ผ๐—ฟ ๐—•๐—ฒ๐—ป๐—ฐ๐—ต๐—บ๐—ฎ๐—ฟ๐—ธ๐—ถ๐—ป๐—ด ๐—–๐—ผ๐—ป๐—ฐ๐—ฒ๐—ฝ๐˜ ๐—•๐—ผ๐˜๐˜๐—น๐—ฒ๐—ป๐—ฒ๐—ฐ๐—ธ ๐— ๐—ผ๐—ฑ๐—ฒ๐—น ๐—˜๐˜…๐—ฝ๐—น๐—ฎ๐—ป๐—ฎ๐˜๐—ถ๐—ผ๐—ป๐˜€โ€ has been accepted to the ๐—ฒ๐—ซ๐—–๐—ฉ ๐—ช๐—ผ๐—ฟ๐—ธ๐˜€๐—ต๐—ผ๐—ฝ ๐—ฎ๐˜ ๐—˜๐—–๐—–๐—ฉ ๐Ÿฎ๐Ÿฌ๐Ÿฎ๐Ÿฒ! ๐ŸŽ‰ Concept Bottleneck Models are built to make predictions interpretable through human-understandable concepts, yet they are still mostly evaluated using downstream classification accuracy and a handful of qualitative examples. In CBX-Bench, we ask a simple question: ๐˜๐˜ฐ๐˜ธ ๐˜ค๐˜ข๐˜ฏ ๐˜ธ๐˜ฆ ๐˜ฒ๐˜ถ๐˜ข๐˜ฏ๐˜ต๐˜ช๐˜ต๐˜ข๐˜ต๐˜ช๐˜ท๐˜ฆ๐˜ญ๐˜บ ๐˜ฆ๐˜ท๐˜ข๐˜ญ๐˜ถ๐˜ข๐˜ต๐˜ฆ ๐˜ธ๐˜ฉ๐˜ฆ๐˜ต๐˜ฉ๐˜ฆ๐˜ณ ๐˜ข ๐˜Š๐˜‰๐˜” ๐˜ฆ๐˜น๐˜ฑ๐˜ญ๐˜ข๐˜ฏ๐˜ข๐˜ต๐˜ช๐˜ฐ๐˜ฏ ๐˜ช๐˜ด ๐˜ข๐˜ค๐˜ต๐˜ถ๐˜ข๐˜ญ๐˜ญ๐˜บ ๐˜ถ๐˜ด๐˜ฆ๐˜ง๐˜ถ๐˜ญ? We first collected 2,700 human judgments across CUB-200, ImageNet-100, and Places365, and used them to validate a council of multimodal LLMs for evaluating CBM explanations. The resulting council recovers over ๐Ÿณ๐Ÿฌ% ๐—ผ๐—ณ ๐—ต๐˜‚๐—บ๐—ฎ๐—ป ๐—ฝ๐—ฟ๐—ฒ๐—ณ๐—ฒ๐—ฟ๐—ฒ๐—ป๐—ฐ๐—ฒ ๐—ฟ๐—ฎ๐—ป๐—ธ๐—ถ๐—ป๐—ด๐˜€, reaching ๐Ÿด๐Ÿฏ% ๐˜„๐—ต๐—ฒ๐—ป ๐—ฎ๐—ป๐—ป๐—ผ๐˜๐—ฎ๐˜๐—ผ๐—ฟ๐˜€ ๐˜‚๐—ป๐—ฎ๐—ป๐—ถ๐—บ๐—ผ๐˜‚๐˜€๐—น๐˜† ๐—ฎ๐—ด๐—ฟ๐—ฒ๐—ฒ. We then use this evaluator to build a public benchmark and leaderboard for comparing CBM explanation quality. This project grew out of the Guided Research course I took during my final semester at METU. Iโ€™m very grateful to Dr. Emre Akbas for trusting us with this project, and to my co-authors Gรผlay Oklan, Seref Baris Cagliyan, and Umut ร–zdemir for all the work that went into it. Iโ€™m incredibly happy to have been part of such a welcoming and collaborative research group. It was a truly rewarding experience, with great mentorship and supervision throughout. ๐Ÿ“„ Paper: lnkd.in/dhdstRyx ๐Ÿ’ป Benchmark: lnkd.in/dHkGt_aQ Looking forward to presenting this work at ECCV 2026 in Malmรถ! Let me know if youโ€™ll be there too! #ECCV2026 #ExplainableAI #XAI #ComputerVision #ConceptBottleneckModels
3
1
27
1,277
Replying to @sebekeyedustuk
Gรผncel olarak bรถyle
1
3
269
4 yฤฑlฤฑn ardฤฑndan ODTรœ bilgisayar mรผhendisliฤŸinden 4.00 ortalama ile mezun oldum!๐Ÿฅณ๐ŸŽ‰ Buradan sonraki durak: ETH Zรผrichโ€™te yรผksek lisans. ๐Ÿ‡จ๐Ÿ‡ญ
894
158
275
17,790
2,633,562
VahลŸetin ortasฤฑnda ลŸafak vakti aรงmฤฑลŸ bir รงiรงek gibi ฤฑลŸฤฑldฤฑyorum
1
1
27
10,281
Mรผnih cennet cennet. Yazฤฑn ortasฤฑnda uzun kolluyla dolaลŸฤฑyorum, Tรผrkiye'de olsam รถlรผmรผ asfalttan kazฤฑrlardฤฑ. Yaz sฤฑcaklarฤฑnฤฑ teฤŸet geรงmek รงok iyi hissettiriyor
11
3,969
Devlet Devlet
devlet bazen yangฤฑnlarฤฑ en hฤฑzlฤฑ ลŸekilde kontrol altฤฑna almak yerine spotify playlistlerinizi stalklar
1
5
4,069
Emir kulu deฤŸil kendi halkฤฑnฤฑn kendi vicdanฤฑnฤฑn kulu
1
8
1,180
Bu da ee'den gรถรงme cengci
2
1
870