Search

There are two common ways in which developers are incorporating proprietary and domain-specific data when building applications of Large Language Models (LLMs): Retrieval-Augmented Generation (RAG) and Fine-Tuning. RAG augments the prompt with the external data, while fine-Tuning incorporates the additional knowledge into the model itself. However, the pros and cons of both approaches are not well understood. In this paper, we propose a pipeline for fine-tuning and RAG, and present the tradeoffs of both for multiple popular LLMs, including Llama2-13B, GPT-3.5, and GPT-4. Our pipeline consists of multiple stages, including extracting information from PDFs, generating questions and answers, using them for fine-tuning, and leveraging GPT-4 for evaluating the results. We propose metrics to assess the performance of different stages of the RAG and fine-Tuning pipeline. We conduct an in-depth study on an agricultural dataset. Agriculture as an industry has not seen much penetration of AI, and we study a potentially disruptive application - what if we could provide location-specific insights to a farmer? Our results show the effectiveness of our dataset generation pipeline in capturing geographic-specific knowledge, and the quantitative and qualitative benefits of RAG and fine-tuning. We see an accuracy increase of over 6 p.p. when fine-tuning the model and this is cumulative with RAG, which increases accuracy by 5 p.p. further. In one particular experiment, we also demonstrate that the fine-tuned model leverages information from across geographies to answer specific questions, increasing answer similarity from 47% to 72%. Overall, the results point to how systems built using LLMs can be adapted to respond and incorporate knowledge across a dimension that is critical for a specific industry, paving the way for further applications of LLMs in other industrial domains.

Content-Centric Networking (CCN) is an internetworking paradigm that offers an alternative to today's IP-based Internet Architecture. Instead of focusing on hosts and their locations, CCN emphasizes addressable named content. By decoupling content from its location, CCN allows opportunistic in-network content caching, thus enabling better network utilization, at least for scalable content distribution. However, in order to be considered seriously, CCN must support basic security services, including content authenticity, integrity, confidentiality, authorization and access control. Current approaches rely on content producers to perform authorization and access control. This general approach has several disadvantages. First, consumer privacy vis-a-vis producers is not preserved. Second, identity management and access control impose high computational overhead on producers. Also, unnecessary repeated authentication and access control decisions must be made for each content request. These issues motivate our design of KRB-CCN - a complete authorization and access control system for private CCNs. Inspired by Kerberos in IP-based networks, KRB-CCN involves distinct authentication and authorization authorities. By doing so, KRB-CCN obviates the need for producers to make consumer authentication and access control decisions. KRB-CCN preserves consumer privacy since producers are unaware of consumer identities. Producers are also not required to keep any hard state and only need to perform two symmetric key operations to guarantee that sensitive content is confidentially delivered only to authenticated and authorized consumers. Most importantly, unlike prior designs, KRB-CCN leaves the network (i.e., CCN routers) out of any authorization, access control or confidentiality issues. We describe KRB-CCN design and implementation, analyze its security, and report on its performance.

In this work we propose, implement, and evaluate GRM, a novel mobility model that accounts for the role of group meeting dynamics and regularity in human mobility. Specifically, we show that existing mobility models for humans do not capture the regularity of human group meetings which is present in real mobility traces. Next, we characterize the statistical properties of such group meetings in real mobility traces and design GRM accordingly. We show that GRM maintains the typical pairwise contact properties of real traces, such as contact duration and inter-contact time distributions. In addition, GRM accounts for the role of group mobility, presenting group meetings regularity and social communities' structure. Finally, we evaluate state-of-art social-aware protocols for opportunistic routing using a synthetic contact trace generated by our model. The results show that the behavior of such protocols in our model is similar to their behavior in real mobility traces.

Techniques such as clusterization, neural networks and decision making usually rely on algorithms that are not well suited to deal with missing values. However, real world data frequently contains such cases. The simplest solution is to either substitute them by a best guess value or completely disregard the missing values. Unfortunately, both approaches can lead to biased results. In this paper, we propose a technique for dealing with missing values in heterogeneous data using imputation based on the k-nearest neighbors algorithm. It can handle real (which we refer to as crisp henceforward), interval and fuzzy data. The effectiveness of the algorithm is tested on several datasets and the numerical results are promising.

In the next generation cellular networks, device-to-device (D2D) communication is already considered a fundamental feature. A problem of multi-hop D2D networks is on how to define forwarding algorithms that achieve, at the same time, high delivery ratio and low network overhead. In this paper we aim to understand group meetings' properties by looking at their structure and regularity with the final goal of applying such knowledge in the design of a forwarding algorithm for D2D multi-hop networks. We introduce a forwarding protocol, namely GROUPS-NET, which is aware of social group meetings and their evolution over time. Our algorithm is parameter-calibration free and does not require any knowledge about the social network structure of the system. In particular, different from the state of the art algorithms, GROUPS-NET does not need communities detection, which is a complex and expensive task. We validate our algorithm using different publicly available data-sources. In real large scale scenarios, our algorithm achieves approximately the same delivery ratio of the state-of-art solution with up to 40% less network overhead., Comment: arXiv admin note: text overlap with arXiv:1512.04828

This paper proposes an efficient method based on the steered-response power (SRP) technique for sound source localization using microphone arrays: the volumetric SRP (V-SRP). As compared to the SRP, by deploying a sparser volumetric grid, the V-SRP achieves a significant reduction of the computational complexity without sacrificing the accuracy of the location estimates. By appending a fine search step to the V-SRP, its refined version (RV-SRP) improves on the compromise between complexity and accuracy. Experiments conducted in both simulated- and real-data scenarios demonstrate the benefits of the proposed approaches. Specifically, the RV-SRP is shown to outperform the SRP in accuracy at a computational cost of about ten times lower., Comment: 14 pages, 9 figures, 5 tables