Auditory-Visual Speech Association (AVISA)

@avsp.bsky.social

1198 Followers 973 Following

The official(ish) account of the Auditory-VIsual Speech Association (AVISA) AV 👄 👓 speech references, but mostly what interests me avisa.loria.fr

A scene looking down from above of boxes of food

Lines 7 - 9 of the Abstract
"commentary, I argue that multiple sources of evidence, including failures to replicate key findings, demonstrate that emantic processing of action words and action-sentences is unlikely to rely on"

A collection of logos from AVSP conferences over the years

Conclusion
In summary, I have argued that a bilateral temporo–frontal network subserves auditorysentence comprehension. Although syntactic and semantic information are processed predominately by the left hemisphere, processing of prosodic information occurs predominantly in the right hemisphere. Temporal regions support identification processes, with syntactic processes involving the left anterior STG, semantic processes recruiting the left MTG and prosodic processes involving the right posterior STG. Frontal regions, by contrast, support the formation of relationships, with syntactic relationships involving BA 44 and the frontal opercular cortex, and semantic relationships recruiting BA 45/47. These different areas within the network must be activated and coordinated to achieve auditory sentence comprehension. The timing of the syntactic processes of structure building precedes, and are initially independent of, semantic processes, although both interact during a later processing phase. Prosodic processes influence syntactic processes, however, the exact timing of this is a subject for future research.

Graph that shows Centered cosine similarity. Mean/STD spread-band plots - Illustrates Three phases:

Encoding (layers 0–5). Wild oscillation. The model is doing heavy lifting to normalise radically different surface forms. The model knows it’s reading English or Chinese, and the representations reflect that.

Reasoning (layers ~10–45). The red line dominates. These are the cross-language same-content pairs. The model has converged on a representation where content identity matters far more than language identity. Same-language different-content pairs (green) are much lower. And cross-language different-content (blue) has negative cosine similarity. This is the format-agnostic reasoning space that Part 1 hypothesised.

Decoding (layers ~45–64). The latenst re-differentiates. The model is preparing to emit tokens, and it needs to commit to a specific language and format. Cross-language similarity drops. The representations become surface-specific again.

This is the three-phase architecture, directly visible in a single forward pass. The encoding region is narrow, just a handful of layers. The reasoning region is quite broad, the bulk of the transformer stack. And the decoding region is where everything unravels back into language-specific token predictions.

Modiolus modiolus - Horse mussel

Am Psychol. 2025 Dec;80(9):1350-1352. doi: 10.1037/amp0001601IF: 7.2 Q1 .

ABSTRACT

The Awards for Distinguished Scientific Contributions are presented to persons who, in the opinion of the Committee on Scientific Awards, have made distinguished theoretical or empirical contributions to basic research in psychology. The 2025 recipients of the APA Distinguished Scientific Contribution Awards were recognized by the 2024 Board of Scientific Affairs and selected by the 2024 Committee on Scientific Awards. For pioneering research on gesture in language creation, language production, language learning, and cognition, Susan Goldin-Meadow is a 2025 award winner. Susan Goldin-Meadow has studied gesture in deaf children who do not receive structured language input, and she has found that the gesture systems that these children generate-called homesign-are highly structured and language-like. She has studied the gestures people produce along with speech, and she has shown that such gestures both reveal and influence thinking. Her work has shaped scientific understanding of human development, learning, and communication. She is a brilliant scientific communicator, a generous leader in the scientific community, and a dedicated and inspiring mentor.

This is a nearly final" draft of our analysis of the geometric ability of linear--separation machines. The book will be published by the MIT Press, fairly soon, and will contain some additional material on learning and on relations between serial and parallel machines. We are anxious to collect: 1. simple errata, typos, etc., 2. sentences or sections that are not very clear, 3. statements that are clear but false, 4. hard-:to-understand notations and illustrations. Please make a list of everything you find along these lines. Mathematicians have drifted into a set of conventions for obscurity in exposition, that make it very hard to understand their basically simple ideas. We are trying to resist this in many ways. Therefore we will be especially grateful for every suggestion that will make the book easier to read. But we need your comments around the middle of April. Diligence will be rewarded, somehow or other. Send comments to M. Minsky Room 817 545 Tech Square.

Graduation vs. reality meme - gradate throws cap in air - 2 weeks later at 3:00 AM trying to remember the code

T-shirt with Mundus sine caesaribus written on it - what Bluesky CEO Jay Graber wore at SXSW 2025

neighborhood information across three levels (orthographic, phonological, and phonographic) across 40 languages.

ABSTRACT

Siblings of children with developmental language disorder (DLD) often have weaker language skills compared to peers with typical development (TD). However, whether their language-relevant audiovisual skills are also atypical is unknown. Study 1 examined whether siblings use information about a talker's mouth shape during phonemic processing as children with TD do. Study 2 examined siblings' ability to match auditory words with observed word articulations. Only children with TD showed a significant MMN to audiovisual phonemic violations, suggesting that, just like in children with DLD, lip shape does not modulate phonemic processing in siblings. Children with DLD and siblings were also less accurate than children with TD at detecting audiovisual word mismatches. The N400 amplitude in children with TD was significantly larger than in children with DLD and marginally larger than in siblings. Phonemic and lexical representations in siblings lack audiovisual details, which may contribute to poor language development.

The ant central complex (CX) and lateral complex (highlighted in colour) are embedded in the centre of the ant brain (top left) and located in close proximity to the mushroom bodies (MBs). Data from flies and moths, as well as other insects, have crystallised the role of the CX as an integrative centre to compare the head direction and goal direction of the insect and to instruct the motor system to make corrective turns. The head direction representation is based predominantly on visual information from the optic lobes (OL). The lateral accessory lobes (LALs) are known as premotor control centres that receive CX output and contain reciprocally connected neurons that, at least in moths110, drive oscillatory steering behaviour. The mushroom bodies (MB) are the likely location for storing view memories that contribute to defining the momentary goal of the insect. Images from the Insect Brain Database148, data from103. (B) Conceptual model of how view learning and corrective steering could be phase locked to zigzagging behaviour110. Left: hypothetical neural activity of lateral accessory lobe (LAL) neurons during ant zigzagging behaviour. Activity patterns are inspired by data from silkworm (Bombyx mori) neurons that control the tracking of pheromone plumes via zigzagging behaviour114. In addition to neurons regularly alternating between up and down states (indicating right/left), other cells show phasic excitation when the states switch (blue). Such neurons in an ant would be highly suited to drive the phase locking of corrective turns. Right: suggested role of dopaminergic feedback from the LAL to the fan-shaped body (FB) and MB to phase lock the occurrence of saccade-like turns and view learning to the turning points of behavioural oscillations, as signalled by hypothetical phasic LAL neurons. Neuroanatomical images © Insect Brain Database (CC BY 4.0)

A mixture of old memes

Claude gives Tal's idea a favourable rating :)

Cartoon shows a person very concerned with the safety of another - punchline "that's reviewer one"

cartoon face covered with musical notes

See https://arcprize.org/blog/oai-o3-pub-breakthrough and https://simonwillison.net/2024/Dec/20/openai-o3-breakthrough/

Shows the painting "Ulysses and the Sirens" (1891) by John William Waterhouse (1849-1917) with the sirens labelled as LLMs

Illustrations (hand drawn) of something meant to be neurons : )

Plot of how long (secs) it took to find a city name in the BlueSky stream - The cities: New York Paris London Rome Tokyo Los Angeles Shanghai Moscow Singapore Chicago Athens Berlin Cairo Sydney Bangkok Delhi Seoul

A screen grab of a browser window showing a feed from the Blue Sky WebSocket -

Shows a man with an statue of an angel (blurred) in the background

EEG data was recorded using 128 channels from 30 healthy subjects using audio, visual and AV stimuli in ‘object detection task’. Electrical source imaging was performed using s-LORETA across seven frequency bands (lower alpha 1, lower alpha 2, upper alpha, beta, delta, gamma, theta) during AV versus unimodal conditions across 66 gyri.

Picture of A. mexicanus and Astyanax aeneus

Socially guided early vocal learning hypothesis. Shows a depiction of the hypothesis about the interacting factors allowing the evolution of socially guided vocal learning during early infancy in humans and the hypothesis for a case of convergent evolution of the same process in marmosets.

Cartoon depicting a boy talking to a dog

A person and a parrot rhythmic dancing to music

http://firehose3d.theo.io/

A cartoon like image of a person wearing an EEG cap and pressing keys on a computer keyboard - there is a monitor with some wavy lines and a load speaker

Image prompt: jazz pianist rhythm - ensemble synchronization - complexity; Bill Evans; Bud Powell; Kenny Barron; Oscar Peterson

Is this a Moment? - Meme from robot/butterfly

The power of human language and thought arises from systematic compositionality—the algebraic ability to understand and produce novel combinations from known components. Fodor and Pylyshyn1 famously argued that artificial neural networks lack this capacity and are therefore not viable models of the mind. Neural networks have advanced considerably in the years since, yet the systematicity challenge persists. Here we successfully address Fodor and Pylyshyn’s challenge by providing evidence that neural networks can achieve human-like systematicity when optimized for their compositional skills. To do so, we introduce the meta-learning for compositionality (MLC) approach for guiding training through a dynamic stream of compositional tasks. To compare humans and machines, we conducted human behavioural experiments using an instruction learning paradigm. After considering seven different models, we found that, in contrast to perfectly systematic but rigid probabilistic symbolic models, and pe

The MARCS Institute, Western Sydney University is advertising Chair i Human Neuroscience based at the Westmead Innovation Quarter. Research human cognition across the life-span

Stick figure pointing to the sky - says "yep, just blue no birds