Autonomous brain-computer interfaces (BCIs) are devices designed to record, process, and interpret neural activity, enabling individuals with neurodegenerative diseases or spinal cord injuries to regain their ability to interact with their environment without assistance. Over the past two decades, BCI technology has advanced remarkably, largely due to improvements in neural recording interfaces, such as high-density micro-electrode arrays (MEAs). These innovations allow researchers to collect more comprehensive neural data, enhancing our understanding of neural dynamics and their interaction with human physiology. However, BCI technology still faces three major challenges that hinder its progression toward practical applications.

1. Data Overload: Modern MEAs, with their increased number of recording channels, generate vast amounts of data. Consequently, the power consumption required to acquire, process, and transmit this data becomes a significant barrier, especially given tissue-safe design constraints. This dissertation investigates three effective approaches to reduce data rates, and thus power consumption: spike detection, spike sorting, and neural signal compression, all of which leverage relevant signal features for specific applications.2. Neural decoding versatility: No single neural decoding algorithm suits all applications or users. Therefore, a versatile BCI system must be capable of executing a range of neural decoding algorithms. This dissertation examines the efficient design and implementation of two processor architectures supporting various neural decoding schemes. One processor uses fine-grained sequential processing to implement arbitrary machine-learning-based neural decoding models, while the other employs biologically plausible neuron models to realize a spiking neural network. 3. Autonomous user engagement: Traditional BCIs require users to engage with the system during predefined time periods. This dissertation explores two methods for estimating a user's intention to engage with a BCI application: one using high-frequency neural spikes and the other using low-frequency local field potentials. A hybrid (multi-signal) asynchronous BCI is designed, implemented, and verified, combining both neural signal types to optimize intention estimation and improve neural decoding performance.

The approaches discussed in this dissertation present practical strategies to advance BCI technology by reducing power consumption, enabling flexible and robust neural decoding, and incorporating various neural signal features for efficient user intention estimation.