←
Frieda Born48 posts
Frieda Born avatar

Frieda Born

@FriedaBorn

PhD Student @TUBerlin @bifoldberlin @mpib_berlin | MSc Neuroscience | Examining life through working memory research and nightly philosophy of mind discussions

◷ Joined March 2021▧ 7 Photos and videos
Frieda Born avatar
Frieda Born@FriedaBorn·Apr 16

Can context-sensitivity improve human–machine alignment in VFMs? We show that incorporating context improves modeling of human similarity decisions. 🎉Accepted to ReAlign @iclr_conf! Now on @arxiv: arxiv.org/abs/2604.13883 Joint first-authored work with amazing Tom Neuhäuser. 1/5

Post attachment
Frieda Born avatar
Frieda Born@FriedaBorn·Apr 16

Context shifts which features matter. The same object can be represented differently depending on what it is compared to, effectively reshaping the embedding space to better match human judgments. 4/5

Frieda Born avatar
Frieda Born@FriedaBorn·Apr 16

Standard ML models miss this: they represent inputs as fixed points in embedding space, independent of context. We show that incorporating context and modeling dynamic representations leads to substantial improvements (~15%) in predicting human choices across multiple VFMs. 3/5

Post attachment
Frieda Born avatar
Frieda Born@FriedaBorn·Apr 16

Humans don't perceive objects in isolation — context shapes how we compare things. Context can come from many sources, such as previous experience, goals, or the surrounding items in an environment. 2/5

Frieda Born avatar
Frieda Born@FriedaBorn·Feb 18

🎉Happy to share that this work is now published in @CommsPsychol. 🔗You can find the final version here: nature.com/articles/s44271-0…

Why do some memories fade in seconds, while others stay with us for life? Working Memory (WM) holds info for just moments, but certain bits manage to stick around and make it into Long-Term Memory (LTM). In our new ⚡️preprint⚡️, we examined what helps these memories stick. 1/

Post attachment
Frieda Born avatar
Frieda Born@FriedaBorn·13 Nov 2025

🥳 Our paper on aligning neural networks with human conceptual structure is out in @Nature ! Honored to be part of this amazing team. It was a lot of fun! Thank you @lukas_mut , @AndrewLampinen ,@mc_mozer, Klaus Greff, Bernhard Spitzer & all other collaborators!

What aspects of human knowledge do vision models like CLIP fail to capture, and how can we improve them? We suggest models miss key global organization; aligning them makes them more robust. Check out @lukas_mut's work, finally out (in @Nature!?) + our new blogpost! 1/4

Post attachment
Sander van Bree avatar
⟳ Frieda Born retweeted
Sander van Bree@sandervanbree·25 Oct 2024

During my PhD, I noticed it can be challenging to decode long-term memory (LTM) contents from EEG activity. We were struck by research showing that transient visual "pings" can boost working memory classification. So, we evaluated this technique for LTM: direct.mit.edu/imag/article/…

Lukas Muttenthaler avatar
⟳ Frieda Born retweeted
Lukas Muttenthaler@lukas_mut·19 Oct 2024

🚨The Levels dataset that we collected as part of the AligNet effort is now publicly available: gin.g-node.org/fborn/Dataset… Levels is a human similarity judgment dataset across three abstraction/granularity levels meant for evaluating the alignment of vision foundation models w/ 🧠

What aspects of human knowledge are vision models missing, and can we align them with human knowledge to improve their performance and robustness on cognitive and ML tasks? Excited to share this new work led by @lukas_mut! 1/10

Post attachment
Lukas Muttenthaler avatar
⟳ Frieda Born retweeted
Lukas Muttenthaler@lukas_mut·15 Oct 2024

Which tasks benefit the most from aligning vision models with human perceptual judgments? In our most recent NeurIPS paper we pin down the downstream tasks where perceptual alignment yields the strongest increases in performance. More in the thread below 👇

What happens when models see the world as humans do? In our #NeurIPS2024 paper we show that aligning to human perceptual preferences can *improve* general-purpose representations! 📝: arxiv.org/abs/2410.10817 🌐: percep-align.github.io 💻: github.com/ssundaram21/dream… (1/n)

Post attachment
Frieda Born avatar
Frieda Born@FriedaBorn·15 Oct 2024

The WM-testing effects, particularly for deprioritized info, show intriguing parallels with classic “retrieval practice” effects found in episodic (long-term) memory research. Want to dive deeper into this? 7/

Frieda Born avatar
Frieda Born@FriedaBorn·15 Oct 2024

Also interesting, our results showed that LTM benefits occurred (only) after recall-like WM tests (continuous report), not after simple WM comparisons against a probe. This suggests, the benefits were driven by active retrieval and/or self-generation processes. 6/

Post attachment
Frieda Born avatar
Frieda Born@FriedaBorn·15 Oct 2024

Interestingly, although deprioritization reduced immediate WM test accuracy as expected, it paradoxically improved subsequent LTM recall—revealing a stronger WM-testing effect for deprioritized (unattended) WM contents (WM Test 2)! 5/

Post attachment
Frieda Born avatar
Frieda Born@FriedaBorn·15 Oct 2024

We ran 3 experiments testing how different WM priority manipulations and WM testing formats influenced later LTM recall. Both prioritized and deprioritized info benefited from WM retrieval. 4/