BEGIN:VCALENDAR VERSION:2.0 PRODID:-//tcsai.github.io//Seminar//EN CALSCALE:GREGORIAN METHOD:PUBLISH X-WR-CALNAME:Seminar | Tilburg Computational Linguistics & Psycholinguistic s X-WR-TIMEZONE:Europe/Amsterdam REFRESH-INTERVAL;VALUE=DURATION:PT1H X-PUBLISHED-TTL:PT1H BEGIN:VTIMEZONE TZID:Europe/Amsterdam BEGIN:DAYLIGHT TZOFFSETFROM:+0100 TZOFFSETTO:+0200 TZNAME:CEST DTSTART:19700329T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:+0200 TZOFFSETTO:+0100 TZNAME:CET DTSTART:19701025T030000 RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:20260917-fd976a49@tcsai.github.io DTSTAMP:20261007T145756Z LAST-MODIFIED:20261007T145756Z SEQUENCE:1 DTSTART;VALUE=DATE:20260917 DTEND;VALUE=DATE:20260918 SUMMARY:Seminar: Gaofei Shen – Representational Differences Explanations for spoken language models. DESCRIPTION:Representational similarity analysis tells us whether two space s look alike. Neither describes the structure of the difference itself. In this talk I present the work in progress from my research visit to the Un iversity of Edinburgh on Representational Difference Explanations (RDX\; K ondapaneni et al.\, 2025)\, a method from computer vision that clusters th e stimuli which two models represent differently. We adapt RDX to speech a nd apply it in layerwise comparisons between wav2vec2-base and the ASR-fin e-tuned wav2vec2\, HuBERT-base\, an AudioSet-trained wav2vec2\, Whisper-sm all\, and a random-weight wav2vec2 as control\, on the same TIMIT dataset. Different model pairs produce different layerwise profiles in terms of ho w the organization of their internal representations differs. Alongside a curated inventory of audio- and text-based features\, we also found that s ome of these automatically discovered clusters do not correspond to any la bel in that inventory. To address this\, we are attempting to use an LLM-a ssisted method to help label the clusters with statistical cross-checks.\n \nhttps://tcsai.github.io/language-computation/seminar.html URL:https://tcsai.github.io/language-computation/seminar.html END:VEVENT BEGIN:VEVENT UID:20261008-50a6fc78@tcsai.github.io DTSTAMP:20261007T145756Z LAST-MODIFIED:20261007T145756Z SEQUENCE:1 DTSTART;TZID=Europe/Amsterdam:20261008T123000 DTEND;TZID=Europe/Amsterdam:20261008T133000 SUMMARY:Seminar: Hosein Mohebbi – Efficient and Interpretable Models of S poken Language with Adaptive Context Compression LOCATION:D 152 A\, Tilburg University DESCRIPTION:This talk explores adaptive compute for audio context compressi on. We introduce a method to progressively shorten the sequence of acousti c frames as they pass through successive model layers. By dynamically merg ing redundant temporal frames\, our approach physically reduces the contex t length at inference time\, yielding tangible latency improvements. A maj or focus of the talk will be the techniques used to guarantee strict train -inference consistency despite physically merging and removing these frame s. I will share evaluations on OpenAI's Whisper model across ASR and Speec h Translation tasks\, showing that adaptive context compression consistent ly outperforms static baselines like LayerDrop and Distillation\, maintain ing or even improving original model performance\, while eliminating 50– 70% of candidate frames and yielding significant inference speedups.\n\nht tps://tcsai.github.io/language-computation/seminar.html URL:https://tcsai.github.io/language-computation/seminar.html END:VEVENT BEGIN:VEVENT UID:20261029-b2dbd4f2@tcsai.github.io DTSTAMP:20261007T145756Z LAST-MODIFIED:20261007T145756Z SEQUENCE:1 DTSTART;TZID=Europe/Amsterdam:20261029T123000 DTEND;TZID=Europe/Amsterdam:20261029T133000 SUMMARY:Seminar: Bruno Nicenboim – All models are wrong\, and so is model comparison LOCATION:D 152 A\, Tilburg University DESCRIPTION:A central goal of cognitive science\, and of any field that for malizes its theories\, is to determine which theory best and most accurate ly explains a set of phenomena. Researchers often implement competing theo ries as computational models and use model comparison techniques to evalua te the merits of the models as proxies for the theories.\n\nModel comparis on has important limitations when all candidate models are misspecified\, a situation that is virtually guaranteed in real-world applications but is often downplayed or plainly ignored. I illustrate these problems with a s imulation. I generated response times and choices from a known race proces s and fit three kinds of misspecified models to them: (1) models structura lly similar to the "true" model but with wrong peripheral (or auxiliary) a ssumptions\, (2) a model that assumes a qualitatively different mechanism\ , and (3) theory-agnostic statistical models. I then ranked the models wit h Bayes factors and cross-validation. The simulation shows that the rankin gs depended on which aspect of the data was modeled\, which peripheral ass umptions were implemented\, and which comparison method was used. The best predictions did not always come from the model closest to the truth. In a real application\, moreover\, we would not know which situation we were i n\, whether we had all the relevant data\, or how sensitive our comparison was to peripheral assumptions. The case study uses Bayesian cognitive mod els\, but the problem affects any model comparison\, whether it relies on Bayes factors\, cross-validation\, information criteria\, or held-out benc hmarks. All of these tools rank models according to statistical criteria t hat do not\, by themselves\, establish which theory is closer to the truth . Turning model rankings into claims about theories therefore requires arg uments that go beyond statistics. I close with suggestions for using model comparison as a diagnostic tool: separating core from peripheral assumpti ons\, choosing informative data and severe tests\, and building shared ben chmarks so that evidence accumulates across studies.\n\nhttps://tcsai.gith ub.io/language-computation/seminar.html URL:https://tcsai.github.io/language-computation/seminar.html END:VEVENT BEGIN:VEVENT UID:20261112-c350b275@tcsai.github.io DTSTAMP:20261007T145756Z LAST-MODIFIED:20261007T145756Z SEQUENCE:1 DTSTART;TZID=Europe/Amsterdam:20261112T124500 DTEND;TZID=Europe/Amsterdam:20261112T134500 SUMMARY:Seminar: Sara Østergaard LOCATION:MKZ 221\, Tilburg University DESCRIPTION:https://tcsai.github.io/language-computation/seminar.html URL:https://tcsai.github.io/language-computation/seminar.html END:VEVENT BEGIN:VEVENT UID:20261126-f0b7be34@tcsai.github.io DTSTAMP:20261007T145756Z LAST-MODIFIED:20261007T145756Z SEQUENCE:1 DTSTART;TZID=Europe/Amsterdam:20261126T123000 DTEND;TZID=Europe/Amsterdam:20261126T133000 SUMMARY:Seminar: Thomas Lieber – Piecewise additive methods LOCATION:D 152 A\, Tilburg University DESCRIPTION:https://tcsai.github.io/language-computation/seminar.html URL:https://tcsai.github.io/language-computation/seminar.html END:VEVENT BEGIN:VEVENT UID:20261210-2829a3e9@tcsai.github.io DTSTAMP:20261007T145756Z LAST-MODIFIED:20261007T145756Z SEQUENCE:1 DTSTART;TZID=Europe/Amsterdam:20261210T123000 DTEND;TZID=Europe/Amsterdam:20261210T133000 SUMMARY:Seminar: Grzegorz Chrupała – The rise and evolution of a referen tial code in populations of bee-like agents LOCATION:D 152 A\, Tilburg University DESCRIPTION:Communication typically relies on a shared code\, and any chang e to it must be coordinated between senders and receivers to avoid a break down of communication. The honeybee waggle dance illustrates this problem: species with horizontal combs point directly at a food source\, while spe cies with vertical combs cannot point directly and instead reference the d ance to gravity\, decoded against the position of the sun. We model the ri se of the first of these codes and its evolutionary transition to the seco nd in populations of bee-like agents\, with selection acting at the level of colonies. In a horizontal-comb model\, we find that direct pointing evo lves readily when food is moderately hard to find by random search alone\, whether because sites are few and large or many and small. Communication fails to evolve when food is too sparse to spark dances or so abundant tha t it is found without signaling. Adding an exogenous benefit for vertical combs\, we then find that the transition to the gravity-referenced code is driven mainly by the magnitude of this benefit and by the mutation scale\ , with the coupling between sender and receiver mutations playing a furthe r role when the mutation scale is low. Given a favorable confluence of the se factors\, the transition proceeds reliably and without a breakdown of c ommunication. Outside that confluence it remains possible\, though rarer\, across a much wider range of settings.\n\nhttps://tcsai.github.io/languag e-computation/seminar.html URL:https://tcsai.github.io/language-computation/seminar.html END:VEVENT BEGIN:VEVENT UID:20270121-413b4e9d@tcsai.github.io DTSTAMP:20261008T164303Z LAST-MODIFIED:20261008T164303Z SEQUENCE:1 DTSTART;TZID=Europe/Amsterdam:20270121T123000 DTEND;TZID=Europe/Amsterdam:20270121T133000 SUMMARY:Seminar: Albert Gatt – How do VLMs reason about relations? Mechan istic explorations of spatial and temporal reasoning in vision-language mo dels. LOCATION:D 152 A\, Tilburg University DESCRIPTION:https://tcsai.github.io/language-computation/seminar.html URL:https://tcsai.github.io/language-computation/seminar.html END:VEVENT END:VCALENDAR