Search

The basal ganglia are part of the neural circuit involved in inhibitory control. Gradual, consistent changes in firing rate (ramps) have been found in the basal ganglia during 'hold' periods of different behavioural tasks involving inhibitory control. However, the properties of these ramps, such as startpoints, endpoints, amplitude, length and single-trial properties, have not yet been characterised in detail. In order to obtain these properties accurately, we tested and adapted various methods from the wind-ramp and changepoint literature. Specifically, we used pruned exact linear time (PELT) to detect changepoints for ramp detection. Compared to other algorithms, the PELT method was found to have the best performance. Utilising PELT, we then analysed a large data set of electrophysiological recordings in rats performing behavioural tasks to determine the properties of firing rate ramps during hold periods. First, we established that these ramps exist during the hold period in a stop-signal task. Second, we demonstrate that these ramp properties vary widely across the basal ganglia. Third, at the single-trial level, we show that neurons with detected ramps could have different underlying spiking patterns, for example, stepping or ramping. Finally, we investigated the entrainment of neurons in the basal ganglia and found that different waveforms can influence entrainment properties at gamma frequencies.

The emergence and development of information technologies have given rise to an influx of digital data. This data, spanning unstructured text documents, semi-structured tables, web pages, and structured databases, is widely disseminated over the Internet or privately stored within computer systems. With great convenience, the explosion and complexity of such data also pose a great challenge for efficient machine and human consumption. There is an increasing need for intelligent data systems that can effectively process massive amounts of digital data from heterogeneous sources and make them easily accessible to users of all backgrounds.The burgeoning of artificial intelligence techniques in recent years, especially machine learning, has shed some light on how we could design the next-generation intelligent data system. On the one hand, machine learning models have shown great potential in understanding various types of data (e.g., images and texts) by learning distributed representations. On the other hand, AI techniques have achieved great success in assisting the user with information acquisition and decision-making tasks. In this dissertation, we focus on bringing machine intelligence to different layers of the data ecosystem: At the low-level, we study knowledge discovery from heterogeneous data that extracts clean, actionable knowledge from raw data to serve different stakeholders and user communities. At the high-level, we build natural language interfaces (NLIs) that allow users to query massive heterogeneous data using natural language.Specifically, in Part II, we will discuss solutions to automatically harness knowledge from heterogeneous data sources. Existing methods mainly focus on unstructured data such as plain text or images. In contrast, data in reality usually comes with varied structures (e.g., tables, graph data, and HTML webpages). Moreover, machine learning models are known to be data-hungry. While massive unlabeled raw data are relatively easy to get, high-quality human annotated data is not. In light of these limitations, we focus on (i) leveraging data from heterogeneous sources, especially semi-structured and structured data in addition to plain text, (ii) self-supervised pretraining methods, which pretrains the model with massive unlabeled data, allowing efficient adaption to different domains and applications.In part III, we will then show our research on building natural language interfaces (NLIs), which allow users to express their needs, either querying the data or conducting data analysis, in natural language and directly return the results. Despite the impressive performance of pretrained language models in natural language understanding, they are still weak at conducting complex reasoning over the input. This is of particular importance for NLI systems to resolve complicated user queries. In addition, as the intermediate interface between the user, machine, and data, NLI requires a joint understanding of natural language, machine language, and backend data of varied types. I study two representative types of natural language interfaces: Text2SQL, which translates user natural language queries to executable SQL programs; and Question Answering (QA), which extracts the answer from the data given the user question.In the final chapter, we will conclude the dissertation and outline potential areas for future exploration. We will first highlight the key contributions of our work to date, then suggest prospective research pathways that extend our work along the two directions. In essence, our research aims to mitigate the burden of manual data processing and exploration, ultimately unlocking the full potential of the digital era. This entails the development of more versatile and data-efficient models and the provision of Natural Language Interfaces (NLIs) that exhibit greater reliability, transparency, and trustworthiness.