We introduce a novel method for both partial and full optimization of the connections in deep differentiable logic gate networks (LGNs) and lookup table networks (LUTNs). Our training method utilizes a probability distribution over a set of connections per gate/lookup table (LUT) input pin, selecting the connection with highest merit, all whilst the optimal gate types or LUT-entries are learned in parallel. We show that the connection-optimized LGNs outperform standard fixed-connection LGNs on the Yin-Yang, MNIST Handwritten Digits and Fashion-MNIST benchmarks, while requiring only a fraction of the number of logic gates. We achieve 98.92

Deep Differentiable Logic Gate Networks (LGNs) and Lookup Table Networks (LUTNs) are demonstrated to be suitable for the automatic classification of electrocardiograms (ECGs) using the inter-patient paradigm. The methods are benchmarked using the MIT-BIH arrhythmia data set, achieving up to 94.28% accuracy and a $jκ$ index of 0.683 on a four-class classification problem. Our models use between 2.89k and 6.17k FLOPs, including preprocessing and readout, which is three to six orders of magnitude less compared to SOTA methods. A novel preprocessing method is utilized that attains superior performance compared to existing methods for both the mixed-patient and inter-patient paradigms. In addition, a novel method for training the Lookup Tables (LUTs) in LUTNs is devised that uses the Boolean equation of a multiplexer (MUX). Additionally, rate coding was utilized for the first time in these LGNs and LUTNs, enhancing the performance of LGNs. Furthermore, it is the first time that LGNs and LUTNs have been benchmarked on the MIT-BIH arrhythmia dataset using the inter-patient paradigm. Using an Artix 7 FPGA, between 2000 and 2990 LUTs were needed, and between 5 to 7 mW (i.e. 50 pJ to 70 pJ per inference) was estimated for running these models. The performance in terms of both accuracy and $jκ$-index is significantly higher compared to previous LGN results. These positive results suggest that one can utilize LGNs and LUTNs for the detection of arrhythmias at extremely low power and high speeds in heart implants or wearable devices, even for patients not included in the training set.

We introduce a novel method for partial optimization of the connections in Deep Differentiable Logic Gate Networks (LGNs). Our training method utilizes a probability distribution over a subset of connections per gate input, selecting the connection with highest merit, after which the gate-types are selected. We show that the connection-optimized LGNs outperform standard fixed-connection LGNs on the Yin-Yang, MNIST and Fashion-MNIST benchmarks, while requiring only a fraction of the number of logic gates. When training all connections, we demonstrate that 8000 simple logic gates are sufficient to achieve over 98

An important bottleneck in present-day neuromorphic hardware is its reliance on synaptic addition, which limits the achievable degree of parallelization and thus processing throughput. We present a network of monostable multivibrator timers, whose synaptic inputs are simply OR-ed together, thus mitigating the synaptic addition bottleneck. Monostable multivibrators are simple timers which are easily implemented using counters in digital hardware and can be interpreted as non biologically-inspired spiking neurons. We show how fully binarized event-driven recurrent networks of monostable multivibrators can be trained to solve classification tasks. Our training algorithm resolves temporally overlapping input events. We demonstrate our approach on the MNIST handwritten digits, Google Soli radar gestures, IBM DVS128 gestures and Yin-Yang classification tasks. The estimated energy consumption for the MNIST handwritten digits task, excluding the final linear readout layer, is 855pJ per inference for a test accuracy of 98.61% for a reconfigurable network of 500 units, when mapped to the TSMC HPC+ 28nm process.

Well-known approaches to person identification (PID) from afar are based on face recognition or gate detection using a camera. The former is privacy-violating while both suffer from camera poor performance in low/intense light and harsh weather conditions. Mm-wave radars, on the other hand, are proven to be a privacy-friendly robust solution in many applications, including object detection/classification. In this paper, we propose two approaches to radar-based PID using micro-Doppler (MD) analysis and inverse synthetic aperture radar (ISAR). The proposed approaches are evaluated in PID within a family of five members. The results show the great potential of the mm-wave technology in PID applications.

This work proposes a novel approach for hand gesture recognition using an inexpensive, low-resolution (24 x 32) thermal sensor processed by a Spiking Neural Network (SNN) followed by Sparse Segmentation and feature-based gesture classification via Robust Principal Component Analysis (R-PCA). Compared to the use of standard RGB cameras, the proposed system is insensitive to lighting variations while being significantly less expensive compared to high-frequency radars, time-of-flight cameras and high-resolution thermal sensors previously used in literature. Crucially, this paper shows that the innovative use of the recently proposed Monostable Multivibrator (MMV) neural networks as a new class of SNN achieves more than one order of magnitude smaller memory and compute complexity compared to deep learning approaches, while reaching a top gesture recognition accuracy of 93.9% using a 5-class thermal camera dataset acquired in a car cabin, within an automotive context. Our dataset is released for helping future research.

Monostable multivibrators are simple timers which are easily implemented using counters in digital hardware and can be interpreted as non-biologically inspired spiking neurons. We show how fully binarized event-driven recurrent networks of monostable multivibrators can be trained to solve classification tasks. We mitigate an important bottleneck in neuromorphic hardware concepts by circumventing synaptic addition within the network. Here rather, input signals to a neuron are simply OR-ed together. Temporally overlapping input events are resolved at the neuron level. We demonstrate our approach on the MNIST handwritten digits, Google Soli radar gestures, IBM DVS128 gestures and Yin-Yang classification tasks, all with excellent results. The estimated energy consumption for the MNIST handwritten digits task, excluding the final readout layer, is 855pJ per inference for a test accuracy of 98.61% for a reconfigurable network of 500 units that was mapped to a 28nm process.

This work studies how brain-inspired neural ensembles equipped with local Hebbian plasticity can perform active inference (AIF) in order to control dynamical agents. A generative model capturing the environment dynamics is learned by a network composed of two distinct Hebbian ensembles: a posterior network, which infers latent states given the observations, and a state transition network, which predicts the next expected latent state given current state-action pairs. Experimental studies are conducted using the Mountain Car environment from the OpenAI gym suite, to study the effect of the various Hebbian network parameters on the task performance. It is shown that the proposed Hebbian AIF approach outperforms the use of Q-learning, while not requiring any replay buffer, as in typical reinforcement learning systems. These results motivate further investigations of Hebbian learning for the design of AIF networks that can learn environment dynamics without the need for revisiting past buffered experiences.

Event-based neural networks are currently being explored as efficient solutions for performing AI tasks at the extreme edge. To fully exploit their potential, event-based neural networks coupled to adequate preprocessing must be investigated. Within this context, we demonstrate a 4-b-weight spiking neural network (SNN) for radar gesture recognition, achieving a state-of-the-art 93% accuracy within only four processing time steps while using only one convolutional layer and two fully connected layers. This solution consumes very little energy and area if implemented in event-based hardware, which makes it suited for embedded extreme-edge applications. In addition, we demonstrate the importance of signal preprocessing for achieving this high recognition accuracy in SNNs compared to deep neural networks (DNNs) with the same network topology and training strategy. We show that efficient preprocessing prior to the neural network is drastically more important for SNNs compared to DNNs. We also demonstrate, for the first time, that the preprocessing parameters can affect SNNs and DNNs in antagonistic ways, prohibiting the generalization of conclusions drawn from DNN design to SNNs. We demonstrate our findings by comparing the gesture recognition accuracy achieved with our SNN to a DNN with the same architecture and similar training. Unlike previously proposed neural networks for radar processing, this work enables ultralow-power radar-based gesture recognition for extreme-edge devices.

In this article, two implementations of a radar-based human-robot interface are presented. These implementations represent two classes of inference approaches that are investigated in the radar group at imec. The first class exploits traditional machine learning classification techniques. The second class uses spiking neural networks. The machine learning classification system presented in this article supports nine gestures and achieves a gesture classification accuracy of 93%. This compares to an accuracy of 98% for our spiking neural network system operating on four gestures. Based on public data sets, this accuracy remains, while exceeding published state of the art results. Misclassification is however significant, which is still precluding safety critical interactions when using a single radar sensor. As proof-of-concept, a discrete control of a robot will be demonstrated by means of radar-based gesture recognition using five gestures. We present the main concepts of this demonstrator. For pre-validation, we use emulation of the gesture recall statistics and timing characteristics to model the radar part.

We show how anomalous time reversal of stimuli and their associated responses can exist in very small connectionist models. These networks are built from dynamical toy model neurons which adhere to a minimal set of biologically plausible properties. The appearance of a "ghost" response, temporally and spatially located in between responses caused by actual stimuli, as in the phi phenomenon, is demonstrated in a similar small network, where it is caused by priming and long-distance feedforward paths. We then demonstrate that the color phi phenomenon can be present in an echo state network, a recurrent neural network, without explicitly training for the presence of the effect, such that it emerges as an artifact of the dynamical processing. Our results suggest that the color phi phenomenon might simply be a feature of the inherent dynamical and nonlinear sensory processing in the brain and in and of itself is not related to consciousness.

Box 1. Empirical Evidence for Early Competence Just as animals and plants are getting smarter every year (they do not, but our detection methods have become more sensitive), there is now substantial evidence that infants are capable of much more than we assumed was the case even just 10 years ago. Infants as young as 5 months old are capable of implicit theory of mind - that is, the ability to correctly predict others' actions based on their goals [6]. Thus, although explicit ToM may only develop much later, even young infants are capable of correctly inferring other people's mental states in specific contexts. Likewise, just as is the case for animals (see [7] for a recent special issue), preverbal infants as young as 12 months old are capable of limited forms of metacognition [8]. Furthermore, Kouider et al. [9] showed how neural markers associated with consciousness are already present, though in a delayed form, in infants. There is also evidence that fetuses have social interactions, as an in utero study of the movement kinematics of twins has intriguingly suggested [10]. Finally, self-awareness itself is far from a monolithic concept and likely develops gradually in different directions [11,12]. These findings all suggest that core cognitive abilities come into existence much earlier than typically assumed.

As radar sensors are being miniaturized, there is a growing interest for using them in indoor sensing applications such as indoor drone obstacle avoidance. In those novel scenarios, radars must perform well in dense scenes with a large number of neighboring scatterers. Central to radar performance is the detection algorithm used to separate targets from the background noise and clutter. Traditionally, most radar systems use conventional constant false alarm rate (CFAR) detectors but their performance degrades in indoor scenarios with many reflectors. Inspired by the advances in nonlinear target detection, In this article, we propose a novel high performance, yet low-complexity target detector and we experimentally validate our algorithm on a dataset acquired using a radar mounted on a drone. We experimentally show that our proposed algorithm drastically outperforms ordered statistics CFAR (OS-CFAR) (standard detector used in automotive systems) for our specific task of indoor drone navigation with more than 19% higher probability of detection for a given probability of false alarm. We also benchmark our proposed detector against a number of recently proposed multitarget CFAR detectors and show an improvement of 16% in probability of detection compared to censored harmonic averaging CFAR, with even larger improvements compared to both outlier-robust CFAR and truncated statistics log-normal CFAR in our particular indoor scenario. To the best of authors’ knowledge, this article improves the state-of-the-art for high-performance yet low-complexity radar detection in critical indoor sensing applications.

Consciousness remains a formidable challenge. Different theories of consciousness have proposed vastly different mechanisms to account for phenomenal experience. Here, appealing to aspects of global workspace theory, higher-order theories, social theories, and predictive processing, we introduce a novel framework: the self-organizing metarerpresentational account (SOMA), in which consciousness is viewed as something that the brain learns to do. By this account, the brain continuously and unconsciously learns to redescribe its own activity to itself, so developing systems of metarepresentations that qualify target first-order representations. Thus, experiences only occur in experiencers that have learned to know they possess certain first-order states and that have learned to care more about certain states than about others. In this sense, consciousness is the brain's (unconscious, embodied, enactive, nonconceptual) theory about itself.