Current Research Directions

Energy-efficient computing has become a crucial field of research, particularly with the increasing need for embedded machine learning algorithms in biomedical devices, wearables, and other applications that require real-time inference. The current reliance on power-hungry computational resources for neural network algorithms is not sustainable, and novel approaches are required to improve energy efficiency.

In this research project, the focus is on investigating embedded machine learning architectures, circuits, and system techniques that can reduce the energy consumption associated with fetching weights from memory, enable the reuse of weights and biases, and eliminate the explicit distinction between memory and compute during inference. These goals can be achieved by exploring analog neural networks and mixed-signal spiking neural networks.

One approach that has been previously investigated is using a crossbar array composed of floating-gate (FG) transistors for vector matrix multiplication (paper). The charge stored on the floating node, programmed using hot-electron injection process, is used as the weight of the vector-matrix multiplication. Crossbar arrays have proven to be highly energy-efficient and dense since they eliminate the need for an explicit distinction between memory and compute during inference.

Another direction that the research team is pursuing is investigating mixed-signal spiking neural networks. Non-volatile devices such as floating-gate transistors and resistive random-access memory (ReRAM) can be used for analog synapses and adaptive leaky-integrate and fire neurons. Hardware and software are co-designed by modeling these devices in Python and learning in the presence of non-linearities and variations (paper). This approach can enable more efficient computing by reducing the energy consumption associated with synaptic weights.

In conclusion, this research project aims to develop novel machine learning algorithms and hardware architectures that can efficiently infer data in real-time. By exploring analog neural networks and mixed-signal spiking neural networks, the research team hopes to reduce the power consumption associated with embedded machine learning platforms, making them more suitable for implantable and wearable devices.

Brain-machine interfaces (BMI) have emerged as a promising technology for increasing independence and improving the quality of life of patients with spinal cord injury (SCI). With around 17,700 new cases per year in the United States alone, there is a pressing need to develop efficient BMI systems that can read out neural signals and translate them into control signals for assistive devices.

In this research project, the focus is on developing BMI systems that can accurately decode neural signals. The initial work of this project involves recording from posterior parietal cortex of tetraplegic human research participants. This region has been shown to encode the goal of movements and is involved in sensorimotor integration and high-level motor planning. Here we  compared the effectiveness of machine learning algorithms, such as recurrent neural networks, with conventional decoding approaches like Kalman filter. Additionally, the research team is exploring various feature extraction techniques to increase the accuracy of the decoder over multiple days. The ultimate goal is to develop robust BMI decoders that can use stable neural data features and be implemented on an embedded or wearable platform [1,2]

One significant challenge in developing these BMI systems is designing and developing area- and power-efficient hardware for decoding kinematics. Currently, state-of-the-art decoders consume several watts of power, which is not suitable for real-time processing outside of a clinical setting. Thus, the research team is exploring novel architectures for computing these machine learning algorithms on embedded or wearable platforms, making these systems clinically relevant and more readily available to patients [3].

In conclusion, this research project focuses on developing a more efficient and effective BMI system to improve the quality of life of patients with SCI. The team is working on accurately decoding neural signals recorded from the posterior parietal cortex and comparing machine learning algorithms with conventional decoding approaches. They are also investigating various feature extraction techniques to increase the accuracy of the decoder over multiple days and designing area- and power-efficient hardware for decoding kinematics on embedded or wearable platforms. The ultimate goal is to develop a robust and clinically relevant BMI system that can improve the quality of life of patients with SCI. (See the paper for more details.)

The development of wearable devices for continuous monitoring of physiological signals has great potential to improve healthcare by enabling early detection and diagnosis of diseases. However, current state-of-the-art wearable devices rely heavily on cloud computers or remote servers to analyze the data they generate. This can result in large energy consumption and bandwidth requirements, which are set to grow with the increasing number of wearable devices.

This research project aims to address this challenge by investigating and developing circuits and systems that can efficiently process and analyze biological signals in real-time, without relying on cloud or remote servers. Specifically, the focus is on developing energy-efficient circuits and algorithms that can analyze the large amount of data generated by wearable devices for sensing, monitoring, and analyzing biological signals.

The ultimate goal of this project is to develop a wearable device that can perform continuous monitoring of physiological signals in real-time without relying on cloud or remote servers. This will enable patients and healthcare providers to monitor vital signs and other physiological signals in real-time, without the need for wired connections or bulky equipment. Our initial papers focuses on real-time signal processing using microcontroller [1,2] and mixed-signal circuits for signal processing [3,4,5]. 

In conclusion, this research project aims to investigate and develop energy-efficient circuits and systems for wearable devices that can perform continuous monitoring of physiological signals in real-time. By eliminating the need for cloud or remote servers, this technology has the potential to revolutionize healthcare by enabling early detection and diagnosis of diseases.