(Registration ClOSED) HKML S07E01 - Meetup @Premialab

***ATTENTION***: You need a confirmation email from HKML to be able to attend the event. The final list of attendees will be at the discretion of the host.

Registration link:
https://forms.gle/hNS221QrZRaSLfv98

We are delighted to announce that our next meetup will be in cooperation with Premialab.

Special thanks to our sponsor for F&B.

Speaker 1: Gautier Marti
Founder of HKML & quant researcher

Serious, Sassy, or Sad? Teaching Machines to Read the Room (From Speech Embeddings)

Disentangling Speech Embeddings, we introduced a simple yet effective method for removing linguistic content (what is being said) from speech embeddings. Concretely, we trained a linear model to predict raw speech embeddings from the corresponding text embeddings and used the residuals (differences) as a proxy for non-linguistic vocal features. We demonstrated that the resulting “residual” embeddings retained enough vocal cue information to cluster sentences by speaker identity rather than semantic content — suggesting that these embeddings preserve vocal tone and style independently of the words spoken.
Now, we take this idea one step further. This time, we explore whether these residual embeddings can help classify how something is said—focusing on vocal tone and speaking style.

Speaker 2: Chung Wang Wong
AI Researcher @ Qube Research & Technologies

Chung Wang Wong will share advances in the hard-core methodologies of natural language processing with applications in the finance industry.

Title of the talk:
Training LLMs for text-adventure games

Large Language Models (LLMs) have become essential components in various autonomous agent systems. Recent advancements leverage fine-tuning with expert trajectories to improve agent performance in complex decision-making tasks. This talk will explore recent fine-tuning algorithms for LLMs, providing a comparative analysis of several state-of-the-art methods. The discussion will be grounded in empirical evaluations of standard text-adventure game datasets, highlighting key insights, and future research directions.

Drink networking after the event:
https://maps.app.goo.gl/covYWB5itASvYPZV8