Hsin's research interests includes probabilistic neural computation in VLSI, neuromorphic engineering, and Neuro-engineering.
Visit the NeuroEngineering Lab(神經工程實驗室)for interesting neural activities and intelligent people in the NEL!!
Hsin has established the Neuro-Engineering Lab (NEL) since he joined the EE Dept. (http://nel003.ee.nthu.edu.tw/nel/index.html ). As shown by Fig.1, the NEL focuses on three main research areas: (1) novel neuro-electronic interfaces of various forms which enable electronic circuits to interact with neurons bi-directionally and reliably for a long term, (2)VLSI implementation of probabilistic models which utilises noise to form stochastic systems on-a-chip, suitable for recognising biomedical signals (e.g. neural signals) intelligently in implantable devices, (3)neuromorphic systems that facilitate advanced research in neuroscience. From the three research areas sprout two more activities, the biosensor and the analogue memory. Based on sensors employed at neuro-electronic interfaces, biosensors able to detect DNA concentration in real-time have been developed for disease diagnosis and drug discovery, through the collaboration with researchers in the National Health Research Institutes (NHRI) in Taiwan. On the other hand, the nonvolatile analogue memory embedded in standard CMOS technology have been developed for storing parameters values economically in intelligent stochastic systems.
Fig.1 The research
focuses of the Neuro-Engineering Lab
Probabilistic neural computation in VLSI
Probabilistic neural computation refers to brain-inspired algorithms that are based on computing units (called neurons) whose inputs merely decide their "output probabilities". Consisting of probabilistic neurons in parallel connection, probabilistic models can use stochasticity to generalise the natural variability of real data. Probabilistic models have thus been shown promising for recognising complex patterns (e.g. face images) and for modelling noisy biomedical signals. VLSI implementation of probabilistic models is attractive for at least the following reasons. Firstly, as VLSI technology continues to scale towards deep-sub-micron, intrinsic noise in transistors will become non-negligible and process variations will introduce more computational errors. For this challenge, VLSI circuits with probabilistic behaviour, i.e. probabilistic input-output relationship, suggest a useful approach to discouraging the propagation of noise and computational errors in VLSI circuits. Furthermore, probabilistic models provide a math tool for reasoning and constructing a system based on elementary circuits with probabilistic behaviour. A probabilistic VLSI system could then points towards a novel robust computational scheme in VLSI.
During Hsin's PhD research, Hsin has developed a continuous-valued probabilistic model, called the Continuous Restricted Boltzmann Machine (CRBM), which is able to adapt its "internal noise" effectively to code the variability of the continuous data it models. Not only has the CRBM been shown useful in modelling biomedical and micro-sensory data, but also has the CRBM been translated into a full system in mixed-mode VLSI (CRBMFULL). It has been demonstrated that continuous-valued probabilistic behaviour can be induced in VLSI effectively, by the use of noise, and that the noise-induced probabilistic behaviour in VLSI can be adapted towards modelling a probabilistic data distribution.
The immediate extension of this research will include low-power and programmable design of a CRBM-based system which suggests an intelligent embedded system in practical biomedical applications. As CRBM is a variant of the Diffusion Network, it is also interesting to explore the adaptation of probabilistic dynamics in VLSI circuits towards modelling distributions of data sequences. Moreover, as the probabilistic behaviour in VLSI currently relies on a noise injection from a noise generator, a robust “noisy computation” (A.F. Murray 2002) that utilizes rather than suppresses the intrinsic electronic noise would be especially interesting and useful in deep-sub-micron process where noise drastically increases (Fig.3b). This part of extensive research has been carried out in the "neural" activities in the Edinburgh University.
Neuromorphic Engineering
Neuromorphic engineering refers to hardware implementation that mimics the morphology of biological systems. As the models of biological cells or systems are described in terms of mathematical equations, neuromorphic engineering refers to translating these equations into hardware, and realising the artificial hardware analogy of biological systems. The hardware analogy can not only suggests alternative computational approach to real-world applications (e.g. spiking-neuron activities in ETH, Zurich), but also provides a tool to examine the reality of biological models in the real world (e.g. Babybots in LIRA-lab, Italy). The latter could lead to a better understanding of a real biological system. More examples of neuromorphic engineering include silicon cochlea, silicon retina, synapse transistors in Washington University, and the silicon growth cone. The challenges of neural engineering lie not only in achieving plausible emulation of a biological system, but also in that it is sometimes lack of a mathematical model to tell how to integrate neuromorphic hardware elements into a useful system, since biologists have not yet understood thoroughly how biological cells (e.g. neurons) develop and organise themselves into a system (e.g. our brain). But this is where research is needed! In this field, Hsin is especially interested in neuromorphic approach to developing circuits that can function as "signal converters" between analogue VLSI and biological neural systems.
Neuro-Engineering
Different research groups have defined "Neuro-Engineering" in different ways. In a more general definition given by the Dept. of Biomedical Engineering in the Johns Hopkins University, neuro-engineering refers to engineering responses to fundamental research issues and clinical problems in the neuroscience. Neuroengineering thus covers a wide field including recording and stimulating biological neurons, the development of human-machine interface, controllable modifications on the synaptic plasticity of biological neurons, etc. Hsin is particularly interested in implementing a hybrid silicon-neuron system that integrates electronic circuits with biological nervous systems. The hybrid system would also provide neuroscientists with a better micro-lab to explore our brain further. The development of hybrid systems could also lead to an alternative treatment for neural prostheses. Following better understanding of our brain, the hybrid system could further provide a platform to explore new computational power, such as cognition, induction, and imagination, which contemporary computers can not achieve.
Le Masson et al has demonstrated the first successful "conversation" between neuromorphic VLSI circuit and biological neuron (Hybrid thalamic circuits). A hybrid neural network not only reconstructs the same firing pattern as fully-biological. network, but also allows parameters to be selectively and individually tuned. Hsin’s first attempt in this research area will thus be trying to integrate the micro electrode arrays with the neuromorphic VLSI circuits. It will then be interesting to investigate whether the impaired function in biological systems could be compensated by the electronic system after learning, or vice versa?!