Scaling up training datasets and model parameters have benefited neural network-based language models, but also present challenges like distributed compute, input data bottlenecks and reproducibility of results. We introduce two simple and scalable software libraries that simplify these issues: t5x enables training large language models at scale, while seqio enables reproducible input and evaluation pipelines. These open-source libraries have been used to train models with hundreds of billions of parameters on multi-terabyte datasets. Configurations and instructions for T5-like and GPT-like models are also provided. The libraries can be found at https://github.com/google-research/t5x and https://github.com/google/seqio.

Contains fulltext : 161870.pdf (Publisher’s version ) (Open Access)

We introduce a logic for temporal beliefs and intentions based on Shoham's database perspective and we formalize his coherence conditions on beliefs and intentions. In order to do this we separate strong beliefs from weak beliefs. Strong beliefs are independent from intentions, while weak beliefs are obtained by adding intentions to strong beliefs and everything that follows from that. We provide AGM-style postulates for the revision of strong beliefs and intentions: strong belief revision may trigger intention revision, but intention revision may only trigger revision of weak beliefs. After revision, the strong beliefs are coherent with the intentions. We show in a representation theorem that a revision operator satisfying our postulates can be represented by a pre-order on interpretations of the beliefs, together with a selection function for the intentions.

We apply an existing formal framework for practical reasoning with arguments and evidence to the Goal-oriented Requirements Language (GRL), which is part of the User Requirements Notation (URN). This formal framework serves as a rationalization for elements in a GRL model: using attack relations between arguments we can automatically compute the acceptability status of elements in a GRL model, based on the acceptability status of their underlying arguments and the evidence. We integrate the formal framework into the GRL metamodel and we set out a research to further develop this framework.

In many logic-based BDI agent programming languages, plan selection involves inferencing over some underlying knowledge representation. While context-sensitive plan selection facilitates the development of flexible, declarative programs, the overhead of evaluating repeated queries to the agent's beliefs and goals can result in poor run time performance. In this paper we present an approach to multi-cycle query caching for logic-based BDI agent programming languages. We extend the abstract performance model presented in (Alechina et al. 2012) to quantify the costs and benefits of caching query results over multiple deliberation cycles. We also present results of experiments with prototype implementations of both single- and multi-cycle caching in three logic-based BDI agent platforms, which demonstrate that significant performance improvements are achievable in practice.