Linear Transformers and State Space Models have emerged as efficient alternatives to softmax Transformers for causal sequence modeling, enabling parallel training via matrix multiplication and efficient RNN-style inference. However, despite their success in causal tasks, no unified framework exists for applying Linear Transformers to bidirectional sequence modeling. We introduce LION, the first framework to systematically extend Linear Transformers to the bidirectional setting. LION generalizes three core representations commonly used in the causal case—full Linear Attention , bidirectional RNN, and chunkwise parallel form—to the bidirectional setting. These forms are theoretically equivalent and enable models to exploit the strengths of each during training and inference. We prove that a broad class of Linear Transformers can be extended using LION and validate our framework via three core examples based on the choice of decay type: LION-LIT, the bidirectional extension of [25]; LION-D, based on [44]; and LION-S, a variant using selective decay [34, 13]. Across standard bidirectional tasks, LION enables models to match or exceed the performance of softmax Transformers, while offering significantly faster training and more efficient inference than existing State Space Models.

EEG-based seizure detection models face challenges in terms of inference speed and memory efficiency, limiting their real-time implementation in clinical devices. This paper introduces a novel graph-based residual state update mechanism (REST) for real-time EEG signal analysis in applications such as epileptic seizure detection. By leveraging a combination of graph neural networks and recurrent structures, REST efficiently captures both non-Euclidean geometry and temporal dependencies within EEG data. Our model demonstrates high accuracy in both seizure detection and classification tasks. Notably, REST achieves a remarkable 9-fold acceleration in inference speed compared to state-of-the-art models, while simultaneously demanding substantially less memory than the smallest model employed for this task. These attributes position REST as a promising candidate for real-time implementation in clinical devices, such as Responsive Neurostimulation or seizure alert systems.

The success of self-supervised learning in computer vision and natural language processing has motivated pretraining methods on tabular data. However, most existing tabular self-supervised learning models fail to leverage information across multiple data tables and cannot generalize to new tables. In this work, we introduce XTab, a framework for cross-table pretraining of tabular transformers on datasets from various domains. We address the challenge of inconsistent column types and quantities among tables by utilizing independent featurizers and using federated learning to pretrain the shared component. Tested on 84 tabular prediction tasks from the OpenML-AutoML Benchmark (AMLB), we show that (1) XTab consistently boosts the generalizability, learning speed, and performance of multiple tabular transformers, (2) by pretraining FT-Transformer via XTab, we achieve superior performance than other state-of-the-art tabular deep learning models on various tasks such as regression, binary, and multiclass classification.

Decision trees have been widely used as classifiers in many machine learning applications thanks to their lightweight and interpretable decision process. This paper introduces Tree in Tree decision graph (TnT), a framework that extends the conventional decision tree to a more generic and powerful directed acyclic graph. TnT constructs decision graphs by recursively growing decision trees inside the internal or leaf nodes instead of greedy training. The time complexity of TnT is linear to the number of nodes in the graph, therefore it can construct decision graphs on large datasets. Compared to decision trees, we show that TnT achieves better classification performance with reduced model size, both as a stand-alone classifier and as a base-estimator in bagging/AdaBoost ensembles. Our proposed model is a novel, more efficient and accurate alternative to the widely-used decision trees.

Early detection of mental fatigue and changes in vigilance could be used to initiate neurostimulation to treat patients suffering from brain injury and mental disorders. In this study, we analyzed electrocorticography (ECoG) signals chronically recorded from two non-human primates (NHPs) as they performed a cognitively demanding task over extended periods of time. We employed a set of biomarkers to identify mental fatigue and a gradient boosting classifier to predict the performance outcome, seconds prior to the actual behavior response. An average F1 score of 75.4% ± 8.4% and 86.4%±6.6% was obtained for the two studied NHPs. Our preliminary results demonstrate the feasibility of detecting mental fatigue in healthy primates that could be used for closed-loop control of neurostimulation therapy.

Deep Brain Stimulation (DBS) is an established therapy for advanced Parkinson's disease (PD). Recent studies have applied the closed-loop control (adaptive DBS or aDBS) using feedback from local field potential (LFP) signals. However, current aDBS practices focus on simple feedback like beta band power and thresholding, without optimized control or classification algorithms. In this work, we study the capacity of several classifiers including automatic shrinkage linear discriminant analysis (LDA) to predict motor impairment. We use 20 features extracted from both monopolar and bipolar LFPs in 12 PD patients. In our best setting, we achieve a median accuracy of 70.2%, sensitivity of 81.2% and prediction lead time of 0.1 s across patients. By including relevant features other than beta power, a 13.6% improvement in accuracy is achieved. Moreover, the Hjorth parameters and high-frequency oscillation (HFO) features perform best according to the Analysis of Variance (ANOVA) p-value and classifier weights. These results suggest a great potential to improve current aDBS system for PD, by implementing a classifier with multiple features.

Wireless implantable devices capable of monitoring the electrical activity of the brain are becoming an important tool for understanding and potentially treating mental diseases such as epilepsy and depression. While such devices exist, it is still necessary to address several challenges to make them more practical in terms of area and power dissipation. In this work, we apply Learning Based Compressive Subsampling (LBCS) to tackle the power and area trade-offs in neural wireless devices. To this end, we propose a low-power and area-efficient system for neural signal acquisition which yields state-of-art compression rates up to 64x with high reconstruction quality, as demonstrated on two human iEEG datasets. This new fully digital architecture handles one neural acquisition channel, with an area of 210x210μm in 90nm CMOS technology, and a power dissipation of only 1μW.

The jitter and the phase noise of ring oscillators utilizing subthreshold source-coupled logic (STSCL) style are analyzed in this paper. Closed-form equations are derived to predict the jitter and phase noise caused by white and flicker noise. Measurement results of a test chip fabricated in a standard CMOS 90 nm technology are presented to validate these expressions. The performed analysis shows that jitter in STSCL-based ring oscillator is independent of technology parameters, as opposed to its CMOS counterparts that depend on supply voltage and parameters of technology. Based on measured results, noise on current control line can dominate the total jitter of the oscillator. Design guidelines are proposed to limit the jitter effect of ring oscillators using STSCL logic. The proposed STSCL-based ring oscillator achieves an average RMS jitter as low as 0.24 % of the oscillation period at a 1.08 MHz/μA energy efficiency, which demonstrates its suitability for ultra-low-power applications.

Ultra-low power circuit design techniques have enabled rapid progress in biosignal acquisition. The design of a multi-channel biosignal recording system is a challenging task, considering the low amplitude of neural signals and limited power budget for an implantable system. The front-end low-noise amplifier is a critical component with respect to overall power consumption and noise of such system. In this paper, we present a new design of III-V Heterojunction TFET (HTFET)-based neural amplifier employing a telescopic operational transconductance amplifier (OTA) for multi-channel neural spike recording. Exploiting the unique device characteristics of HTFETs, our simulation shows that the proposed amplifier exhibits a midband gain of 39 dB, a gain bandwidth of 12 Hz-2.1 kHz, and an input-referred noise of 6.27 μVrms, consuming 5 nW of power at a 0.5 V supply voltage. Using the proposed HTFET amplifier, a noise efficiency factor (NEF) of 0.64 is achieved, which is significantly lower than the CMOS-based theoretical limit. Design tradeoffs related to gain, power and noise requirements are investigated, based on a comprehensive electrical noise model of HTFET and compared with the baseline Si FinFET design.

This paper presents a novel approach to acquire multichannel wireless intracranial neural data based on a compressive sensing scheme. The designed circuits are extremely compact and low-power which confirms the relevance of the proposed approach for multichannel high-density neural interfaces. The proposed compression model enables the acquisition system to record from a large number of channels by reducing the transmission power per channel. Our main contributions are the twofold. First, a CMOS compressive sensing system to realize multichannel intracranial neural recording is described. Second, we explain a joint sparse decoding algorithm to recover the multichannel neural data. The idea has been implemented at system as well as circuit levels. The simulation results reveal that the multichannel intracranial neural data can be acquired by compression ratios as high as four.

High-density wireless intracranial neural recording is a promising technology enabling the autonomous diagnosis and therapy of brain diseases. Increasing the number of recording channels is accompanied by the increased amount of data resulting in an unacceptable transmission power. A comprehensive study of possible compressed sensing methods in the context of neural signals has been done, and the compression of signals originating from different channels in the spatial domain has been implemented at the system and circuit levels. Results of the simulations in a UMC 0.18μm CMOS technology and subsequent reconstructions show the possibility of compressing with ratios as high as 2.6 with a recovery SNR of at least 10dB using extremely compact and low-power circuits. The power efficiency and limited area per channel confirm the relevance of the proposed approach for multi-channel high-density neural interfaces.