Neural networks (NNs) have shown growing impacts on daily life. Two major categories of NNs include well-recognized artificial deep neural networks (DNNs) and biological spiking neural networks (SNNs). While DNNs achieve increasing success in a wide spectrum of applications such as computer vision and natural language processing, SNNs lead people to discover the unveiled brain functions further. Leveraging these models in the biomedical field regarding both real-world diagnoses applications and cutting-edge neuroscience research is of enormous importance. DNNs are widely adopted in bio-signal classification (e.g. electrocardiogram, ECG and electromyography, EMG) and regression (e.g. calcium imaging signal). They typically outperform conventional methods, which require a vast investment of expert knowledge and human labor. However, the data-driven nature of these DNNs models also strongly contradicts the characteristics of bio-signals, which are scarce in labeled samples and share substantial inter-sample variability. This hinders the performance and efficiency of DNNs models for bio-signal classification and regression. Specifically, these models need frequent re-training with additionally annotated subject-specific data, which is costly to obtain and requires manual intervention, to avoid performance drops upon usage. On the other hand, SNNs are capable of simulating bio-realistic neurons and brain structures faithfully. They possess rich potentials in validating neural circuits (e.g. nucleus to cortex feedbacks) and interpreting brain functions (e.g. sensorimotor and associative learning functions of the cerebellum). Nevertheless, SNN models and studies regarding the cerebellar feedback circuits are still lacking.

To address these issues, this thesis is presented, which consists of two parts. The first part is on the design of DNN for bio-signals (including ECG and calcium imaging signal), while the second part is on modeling the cerebellar feedback system with SNN. We first optimized the model design of DNNs for arrhythmia classification from ECG, which is one of the most critical diagnoses in medical routines. We resolved the problems of class imbalance and insufficiency of labeled samples by employing generative adversarial networks (GANs), a special kind of DNN model, which served the purposes of both data augmentation and classification here. We also considered a particular pseudo labeling approach, so as to introduce subject-specific information into the DNN model, where inter-sample variability could be largely reduced. Our proposed GANs based system (ACE-GAN) is fully automatic and performs on similar levels as some expert-assisted methods. In particular, the F₁ score of supraventricular ectopic beats (SVEB or S beats) has been improved by up to 10% over the top-performing automatic systems. Moreover, detection of high sensitivity for both SVEB (87%) and ventricular ectopic beats (VEB or V beats, 93%) has been achieved, which is of great value for practical diagnoses. These results suggest that our optimized automatic system can be a promising tool for high throughput clinical screening practice.

We also modeled DNNs for spike inference from calcium imaging data, one of the fundamental regression tasks to facilitate neuroscience research. We dealt with the significant variability in calcium signal dynamics by modeling a 1-dimensional U-Net with proper optimization and regularization. By translating raw calcium traces into spike-rates in a sequence-to-sequence manner, the resultant predictions were accurate in both time and amplitude. On this basis, we proposed the ENS² system with generalization ability for un-seen calcium imaging signals, where re-training or re-calibration is not required upon usage. It consistently outperforms state-of-the-art algorithms in both spike-rate and spike-event predictions with reduced computational loads. We applied our proposed system to in vivo animal experiments, and testified that our system would improve the interpretation of neuronal orientation selectivity for neuroscience research. We also addressed factors that hindered the performance of our system, where valuable insights were provided for future data preparation to develop DNNs for spike inference.

In the second part, we utilized large-scaled and complex SNN in studies of cerebellar functions. We built SNNs (with neuron number in the order of 10⁴) to model the nucleo-ponto-cortical (NPC) excitatory feedback pathway in the cerebellum, a circuit that had recently been found but remains unclear on its mechanisms. We reproduced the neuronal activities as well as animal behaviors from delay eye-blink conditioning (DEC) experiments with the SNNs. Through a series of in-depth simulations and ablation studies, the plausible mechanisms and possible structures of the NPC feedback circuits in cerebellum-mediated associative learning were revealed. The results suggested that this validated circuit, including at least three differently scaled sub-pathways, could speed up the acquisition of conditioned responses in DEC. Moreover, these three pathways were found to play different roles in synergy.

In summary, we introduced particular modeling methods of DNNs for their applications in bio-signals classification and regression. They overcame challenges in biomedical fields and could serve as automatic tools for daily uses reliably. Moreover, we presented a cerebellum model with feedback circuits based on SNNs, which provided clues about various cerebellar functions for future neuroscience research.