ReLU and Addition-based Gated RNN

We replace the multiplication and sigmoid function of the conventional recurrent gate with addition and ReLU activation. This mechanism is designed to maintain long-term memory for sequence processing but at a reduced computational cost, thereby opening up for more efficient execution or larger models on restricted hardware. Recurrent Neural Networks (RNNs) with gating mechanisms such as LSTM and GRU have been widely successful in learning from sequential data due to their ability to capture long-term dependencies. Conventionally, the update based on current inputs and the previous state history is each multiplied with dynamic weights and combined to compute the next state. However, multiplication can be computationally expensive, especially for certain hardware architectures or alternative arithmetic systems such as homomorphic encryption. It is demonstrated that the novel gating mechanism can capture long-term dependencies for a standard synthetic sequence learning task while significantly reducing computational costs such that execution time is reduced by half on CPU and by one-third under encryption. Experimental results on handwritten text recognition tasks furthermore show that the proposed architecture can be trained to achieve comparable accuracy to conventional GRU and LSTM baselines. The gating mechanism introduced in this paper may enable privacy-preserving AI applications operating under homomorphic encryption by avoiding the multiplication of encrypted variables. It can also support quantization in (unencrypted) plaintext applications, with the potential for substantial performance gains since the addition-based formulation can avoid the expansion to double precision often required for multiplication.

Deep Perceptual Similarity is Adaptable to Ambiguous Contexts

The concept of image similarity is ambiguous, meaning that images that are considered similar in one context might not be in another. This ambiguity motivates the creation of metrics for specific contexts. This work explores the ability of the successful deep perceptual similarity (DPS) metrics to adapt to a given context. Recently, DPS metrics have emerged using the deep features of neural networks for comparing images. These metrics have been successful on datasets that leverage the average human perception in limited settings. But the question remains if they could be adapted to specific contexts of similarity. No single metric can suit all definitions of similarity and previous metrics have been rule-based which are labor intensive to rewrite for new contexts. DPS metrics, on the other hand, use neural networks which might be retrained for each context. However, retraining networks takes resources and might ruin performance on previous tasks. This work examines the adaptability of DPS metrics by training positive scalars for the deep features of pretrained CNNs to correctly measure similarity for different contexts. Evaluation is performed on contexts defined by randomly ordering six image distortions (e.g. rotation) by which should be considered more similar when applied to an image. This also gives insight into whether the features in the CNN is enough to discern different distortions without retraining. Finally, the trained metrics are evaluated on a perceptual similarity dataset to evaluate if adapting to an ordering affects their performance on established scenarios. The findings show that DPS metrics can be adapted with high performance. While the adapted metrics have difficulties with the same contexts as baselines, performance is improved in 99% of cases. Finally, it is shown that the adaption is not significantly detrimental to prior performance on perceptual similarity.

A Systematic Performance Analysis of Deep Perceptual Loss Networks Breaks Transfer Learning Conventions

Deep perceptual loss is a type of loss function in computer vision that aims to mimic human perception by using the deep features extracted from neural networks. In recent years the method has been applied to great effect on a host of interesting computer vision tasks, especially for tasks with image or image-like outputs. Many applications of the method use pretrained networks, often convolutional networks, for loss calculation. Despite the increased interest and broader use, more effort is needed toward exploring which networks to use for calculating deep perceptual loss and from which layers to extract the features. This work aims to rectify this by systematically evaluating a host of commonly used and readily available, pretrained networks for a number of different feature extraction points on four existing use cases of deep perceptual loss. The four use cases are implementations of previous works where the selected networks and extraction points are evaluated instead of the networks and extraction points used in the original work. The experimental tasks are dimensionality reduction, image segmentation, super-resolution, and perceptual similarity. The performance on these four tasks, attributes of the networks, and extraction points are then used as a basis for an in-depth analysis. This analysis uncovers essential information regarding which architectures provide superior performance for deep perceptual loss and how to choose an appropriate extraction point for a particular task and dataset. Furthermore, the work discusses the implications of the results for deep perceptual loss and the broader field of transfer learning. The results break commonly held assumptions in transfer learning, which imply that deep perceptual loss deviates from most transfer learning settings or that these assumptions need a thorough re-evaluation.

A Comparison of Temporal Encoders for Neuromorphic Keyword Spotting with Few Neurons

With the expansion of AI-powered virtual assistants, there is a need for low-power keyword spotting systems providing a "wake-up" mechanism for subsequent computationally expensive speech recognition. One promising approach is the use of neuromorphic sensors and spiking neural networks (SNNs) implemented in neuromorphic processors for sparse event-driven sensing. However, this requires resource-efficient SNN mechanisms for temporal encoding, which need to consider that these systems process information in a streaming manner, with physical time being an intrinsic property of their operation. In this work, two candidate neurocomputational elements for temporal encoding and feature extraction in SNNs described in recent literature - the spiking time-difference encoder (TDE) and disynaptic excitatory-inhibitory (E-I) elements - are comparatively investigated in a keyword-spotting task on formants computed from spoken digits in the TIDIGITS dataset. While both encoders improve performance over direct classification of the formant features in the training data, enabling a complete binary classification with a logistic regression model, they show no clear improvements on the test set. Resource-efficient keyword spotting applications may benefit from the use of these encoders, but further work on methods for learning the time constants and weights is required to investigate their full potential.

Integration of Neuromorphic AI in Event-Driven Distributed Digitized Systems: Concepts and Research Directions

Increasing complexity and data-generation rates in cyber-physical systems and the industrial Internet of things are calling for a corresponding increase in AI capabilities at the resource-constrained edges of the Internet. Meanwhile, the resource requirements of digital computing and deep learning are growing exponentially, in an unsustainable manner. One possible way to bridge this gap is the adoption of resource-efficient brain-inspired "neuromorphic" processing and sensing devices, which use event-driven, asynchronous, dynamic neurosynaptic elements with colocated memory for distributed processing and machine learning. However, since neuromorphic systems are fundamentally different from conventional von Neumann computers and clock-driven sensor systems, several challenges are posed to large-scale adoption and integration of neuromorphic devices into the existing distributed digital-computational infrastructure. Here, we describe the current landscape of neuromorphic computing, focusing on characteristics that pose integration challenges. Based on this analysis, we propose a microservice-based framework for neuromorphic systems integration, consisting of a neuromorphic-system proxy, which provides virtualization and communication capabilities required in distributed systems of systems, in combination with a declarative programming approach offering engineering-process abstraction. We also present concepts that could serve as a basis for the realization of this framework, and identify directions for further research required to enable large-scale system integration of neuromorphic devices.

Identifying and Mitigating Flaws of Deep Perceptual Similarity Metrics

Measuring the similarity of images is a fundamental problem to computer vision for which no universal solution exists. While simple metrics such as the pixel-wise L2-norm have been shown to have significant flaws, they remain popular. One group of recent state-of-the-art metrics that mitigates some of those flaws are Deep Perceptual Similarity (DPS) metrics, where the similarity is evaluated as the distance in the deep features of neural networks. However, DPS metrics themselves have been less thoroughly examined for their benefits and, especially, their flaws. This work investigates the most common DPS metric, where deep features are compared by spatial position, along with metrics comparing the averaged and sorted deep features. The metrics are analyzed in-depth to understand the strengths and weaknesses of the metrics by using images designed specifically to challenge them. This work contributes with new insights into the flaws of DPS, and further suggests improvements to the metrics. An implementation of this work is available online: https://github.com/guspih/deep_perceptual_similarity_analysis/

Technical Language Supervision for Intelligent Fault Diagnosis in Process Industry

In the process industry, condition monitoring systems with automated fault diagnosis methods assisthuman experts and thereby improve maintenance efficiency, process sustainability, and workplace safety.Improving the automated fault diagnosis methods using data and machine learning-based models is a centralaspect of intelligent fault diagnosis (IFD). A major challenge in IFD is to develop realistic datasets withaccurate labels needed to train and validate models, and to transfer models trained with labeled lab datato heterogeneous process industry environments. However, fault descriptions and work-orders written bydomain experts are increasingly digitized in modern condition monitoring systems, for example in the contextof rotating equipment monitoring. Thus, domain-specific knowledge about fault characteristics and severitiesexists as technical language annotations in industrial datasets. Furthermore, recent advances in naturallanguage processing enable weakly supervised model optimization using natural language annotations, mostnotably in the form ofnatural language supervision(NLS). This creates a timely opportunity to developtechnical language supervision(TLS) solutions for IFD systems grounded in industrial data, for exampleas a complement to pre-training with lab data to address problems like overfitting and inaccurate out-of-sample generalisation. We surveyed the literature and identify a considerable improvement in the maturityof NLS over the last two years, facilitating applications beyond natural language; a rapid development ofweak supervision methods; and transfer learning as a current trend in IFD which can benefit from thesedevelopments. Finally, we describe a framework for integration of TLS in IFD which is inspired by recentNLS innovations.

Spatiotemporal Spike-Pattern Selectivity in Single Mixed-Signal Neurons with Balanced Synapses

Realizing the potential of mixed-signal neuromorphic processors for ultra-low-power inference and learning requires efficient use of their inhomogeneous analog circuitry as well as sparse, time-based information encoding and processing. Here, we investigate spike-timing-based spatiotemporal receptive fields of output-neurons in the Spatiotemporal Correlator (STC) network, for which we used excitatory-inhibitory balanced disynaptic inputs instead of dedicated axonal or neuronal delays. We present hardware-in-the-loop experiments with a mixed-signal DYNAP-SE neuromorphic processor, in which five-dimensional receptive fields of hardware neurons were mapped by randomly sampling input spike-patterns from a uniform distribution. We find that, when the balanced disynaptic elements are randomly programmed, some of the neurons display distinct receptive fields. Furthermore, we demonstrate how a neuron was tuned to detect a particular spatiotemporal feature, to which it initially was non-selective, by activating a different subset of the inhomogeneous analog synaptic circuits. The energy dissipation of the balanced synaptic elements is one order of magnitude lower per lateral connection (0.65 nJ vs 9.3 nJ per spike) than former delay-based neuromorphic hardware implementations. Thus, we show how the inhomogeneous synaptic circuits could be utilized for resource-efficient implementation of STC network layers, in a way that enables synapse-address reprogramming as a discrete mechanism for feature tuning.

Pretraining Image Encoders without Reconstruction via Feature Prediction Loss

This work investigates three different loss functions for autoencoder-based pretraining of image encoders: The commonly used reconstruction loss, the more recently introduced perceptual similarity loss, and a feature prediction loss proposed here; the latter turning out to be the most efficient choice. Former work shows that predictions based on embeddings generated by image autoencoders can be improved by training with perceptual loss. So far the autoencoders trained with perceptual loss networks implemented an explicit comparison of the original and reconstructed images using the loss network. However, given such a loss network we show that there is no need for the timeconsuming task of decoding the entire image. Instead, we propose to decode the features of the loss network, hence the name "feature prediction loss". To evaluate this method we compare six different procedures for training image encoders based on pixel-wise, perceptual similarity, and feature prediction loss. The embedding-based prediction results show that encoders trained with feature prediction loss is as good or better than those trained with the other two losses. Additionally, the encoder is significantly faster to train using feature prediction loss in comparison to the other losses. The method implementation used in this work is available online: https://github.com/guspih/Perceptual-Autoencoders

Synaptic Integration of Spatiotemporal Features with a Dynamic Neuromorphic Processor

Spiking neurons can perform spatiotemporal feature detection by nonlinear synaptic and dendritic integration of presynaptic spike patterns. Multicompartment models of nonlinear dendrites and related neuromorphic circuit designs enable faithful imitation of such dynamic integration processes, but these approaches are also associated with a relatively high computing cost or circuit size. Here, we investigate synaptic integration of spatiotemporal spike patterns with multiple dynamic synapses on point-neurons in the DYNAP-SE neuromorphic processor, which can offer a complementary resource-efficient, albeit less flexible, approach to feature detection. We investigate how previously proposed excitatory--inhibitory pairs of dynamic synapses can be combined to integrate multiple inputs, and we generalize that concept to a case in which one inhibitory synapse is combined with multiple excitatory synapses. We characterize the resulting delayed excitatory postsynaptic potentials (EPSPs) by measuring and analyzing the membrane potentials of the neuromorphic neuronal circuits. We find that biologically relevant EPSP delays, with variability of order 10 milliseconds per neuron, can be realized in the proposed manner by selecting different synapse combinations, thanks to device mismatch. Based on these results, we demonstrate that a single point-neuron with dynamic synapses in the DYNAP-SE can respond selectively to presynaptic spikes with a particular spatiotemporal structure, which enables, for instance, visual feature tuning of single neurons.