ETH Zurich develops algorithm to make AI responses more reliable
Cette publication existe aussi en Français
Researchers at the Institute for Machine Learning in the Department of Computer Science at ETH Zurich have developed a method to reduce the uncertainty of AI response engines.
The primary issue with powerful AI response engines is that they provide perfect answers and obvious nonsense with equal ease. One of the significant challenges lies in how the large language models (LLMs) that underlie AI handle uncertainty. Until now, it has been difficult to determine whether LLMs designed for text processing and generation base their responses on a solid data foundation or on uncertain ground.
“Our algorithm can enrich the general language model of the AI with additional data from the relevant subject area of a question. In combination with the specific question, we can then extract from the depths of the model and the enrichment data precisely those connections that are most likely to generate a correct answer,” explains Jonas Hübotter from the Learning and Adaptive Systems Group, who developed the new method as part of his PhD studies.
“The method is particularly suitable for companies, scientists or other users who want to use general AI in a specialised field that is only covered partially or not at all by the AI training data,” adds Andreas Krause, head of the research group and Director of the ETH AI Centre.
Users can feed their locally stored data into an LLM like Llama. The so-called SIFT algorithm (Selecting Informative data for Fine-Tuning), developed by ETH computer scientists, can then use the additional data provided to select specific information that is most closely related to the question.
The algorithm utilises the structure according to which the language information is organised in the LLM to find related information. The models divide the language information in their training data into word parts. The semantic and syntactic relationships between the word parts are then arranged as connecting arrows — known in the field as vectors — in a multidimensional space. The dimensions of space, which can number in the thousands, arise from the relationship parameters that the LLM independently identifies during training using the general data.
Relational arrows pointing in the same direction in this vector space indicate a strong correlation. The larger the angle between two vectors, the less the two information units relate to one another. The SIFT algorithm developed by ETH researchers now uses the direction of the relationship vector of the input query (prompt) to identify those information relationships that are closely related to the question but simultaneously complement each other in terms of content. “The angle between the vectors corresponds to the relevance of the content, and we can use the angles to select specific data that reduces uncertainty,” explains Hübotter.
By contrast, the most common method used to date for selecting the information suitable for the answer, the nearest neighbour method, tends to accumulate redundant information that is widely available.
For example, to answer the two-part question “How old is Roger Federer and how many children does he have?” the nearest neighbour method considers similar information such as “Roger Federer is 43 years old” and “Roger Federer’s birthday is 8 August 1981” to be equally relevant. Information about his children, appropriate for the second part of the question, is sometimes missing. It is overlaid by birth date information, which occurs much more frequently in the AI training data. The SIFT algorithm, however, considers the extent to which the pieces of information included complement each other, i.e. whether the information vectors point in different directions. This allows relevant information to be identified for both aspects of the question.
However, targeted information selection not only enhances the quality of responses but can also help reduce the ever-increasing computing power required by AI applications. By indirectly measuring uncertainty, the model can determine how much more data is needed to provide a sufficiently reliable answer. Consequently, the computational overhead required by an LLM can be systematically adjusted according to the complexity of the question and the availability of relevant information.
Since SIFT continuously adapts the weighting of the arrow directions to its calculations during data retrieval, the enriched model becomes increasingly reliable the more it is used. This is known as test-time training and can be used to achieve the same output performance with smaller models. “In tests with standard data sets, we used SIFT tuning to outperform even the best current AI models with models up to 40 times smaller,” emphasises Hübotter.
Additional applications for the SIFT algorithm are opening up regarding data evaluation, as Krause explains: “We can track which enrichment data SIFT selects. They are closely related to the question and, therefore, particularly relevant to this subject area. For example, this could be used in medicine to investigate which laboratory analyses or measurement values are significant for a specific diagnosis and which are less so.”
https://doi.org/10.48550/arXiv.2410.08020
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
