Expert Group Meeting – Natural Language Processing in Action
Our next Expert Meeting on Monday, 21 August 2023, 17:00-18:30, will focus on various aspects and applications of Large Language Models.
It will take place at the ZHAW premises in Lagerstrasse 45, 8004 Zurich in room ZL O3.01 on the third floor. The meeting will be followed by an apéro. Online participation is also possible.
We will then send you a calendar invitation which includes online participation details.
The following presentations are confirmed for the meeting:
Kim Engels, Converto AG – Large Language Models for Cross-Media Marketing
In this talk, Kim briefly presents some examples of how he and his team at Converto use AI and LLM to improve or speed up their projects.
Besides approaches such as text generation for newsletters, there are also variants such as code generation within the team as well as the use of self-developed solutions to create 3D models for customer campaigns.
Florian Tramér, ETH – Are Aligned Neural Networks Adversarially Aligned?
Large language models are now tuned to align with the goals of their creators, namely to be “helpful and harmless.” These models should respond helpfully to user questions, but refuse to answer requests that could cause harm. However, adversarial users can construct inputs which circumvent attempts at alignment. In this talk, we’ll discuss to what extent these models remain aligned, even when interacting with an adversarial user who constructs worst-case inputs (adversarial examples). We’ll see that existing optimization attacks are insufficiently powerful to reliably attack aligned text models, except when these models are multimodal (i.e., they can process both text and images). In that case, we show these models can be easily attacked, i.e., induced to perform arbitrary un-aligned behavior through adversarial perturbation of the input image.
Alex Paramythis, Contexity AG – Adapting Large Language Models for Customer Request Handling
With the rise of Generative Large Language Models (LLM), companies are looking into the many opportunities proffered by this new technology. One area of particular interest is the automated handling of customer requests (e.g., received through email, chat, social media, etc.) using the institutional knowledge at hand. In such a context, LLMs may need to be trained on, or have access to, privileged, non-public information in the company’s knowledge base. This, in turn, entails that the models need to be prepared within, and served from, a company’s own infrastructure to prevent information leakage — a requirement that points in the direction of commercially friendly open-source models. In this talk we will present our work on generation of responses to customer requests using the IGEL (a BLOOM based model), FLAN-UL2, and Falcon LLMs. For the first two models we will also report on our attempts to fine-tune the models before use, with a variety of training data.