From a scientific/technical perspective, LiveMemories aimed at scaling up content extraction techniques towards very large scale extraction from multimedia sources, setting the scene for a Content Management Platform for Trentino; using this information to support new ways of linking, summarizing and classifying data in a new generation of digital memories which are `alive’ and user-centered; and to turn the creation of such memories into a communal web activity. Achieving these objectives made Trento a key player in the new Web Science Initiative, digital memories, and Web 2.0. But LiveMemories was also intended to have a social and cultural impact besides the scientific one: through the collection, analysis and preservation of digital memories of Trentino; by facilitating and encouraging the preservation of such community memories; and the fostering of new forms of community, and enrichment of our cultural and social heritage.
Our pick of the week by @dhairya_su47605
: "Scaling Laws for Precision" by @tanishqkumar07, Zachary Ankner, @bfspectorShiekh, @blake__bordelon, @Muennighoff, @mansiege, @CPehlevan, Christopher R´e, @AdtRaghunathan
📰
#Quantization #LLM #ScalingLaw
Pick of the week @fbk_mt
Super interesting paper on the limitations of quantization, demonstrating how post-training quantization scales poorly in data.
https://arxiv.org/abs/2411.04330
⭐ For our #PickOfTheWeek, this paper explores an important question for modern speech AI:
🎙️ Which Evaluation for Which Speech Model?
👥 Authors: @Maureendss , @EeshanDhekane
Speech foundation models are evolving rapidly, but evaluation practices are still fragmented.
🏝️ Yesterday at #LREC2026, Palma de Mallorca!
@lina_conti presented "Voice, Bias, and Coreference: An Interpretability Study of Gender in Speech Translation" at the poster session.
📄Paper:
💻Code: https://github.com/lina-conti/voice-bias-coreference
#SpeechTranslation #NLProc
How does the granularity of speech-text pairs impact SpeechLLM performance, and what is the optimal way to interleave tokens? Furthermore, what are the best practices for generating synthetic data to boost training?🧐