The increasing deployment of AI (artificial intelligence) on edge devices presents major challenges due to strict constraints on computation, memory, energy, and latency. Effective Edge AI systems thus require multi-objective optimization that balances accuracy, hardware efficiency, and reliability. The Horizon Twinning project AIDA4Edge tackles these challenges by developing methods for efficient and reliable AI on resource-constrained platforms. This paper presents key approaches explored within the project, including neural network quantization, hardware-aware neural architecture search, dynamic neural networks, and self-adaptive resilient AI architectures. Finally, these strategies are placed within a broader, biologically inspired paradigm, highlighting neuromorphic computing as a natural continuation of Edge AI efforts toward highly efficient and resilient intelligent systems.

The growing demand for deployment of Artificial Intelligence (AI) on resource-constrained edge devices has motivated extensive research on the design of efficient edge-compatible AI hardware accelerators. One of the most promising solutions are the self-adaptive AI accelerators, capable of optimizing in real time their performance and energy consumption according to application requirements. This work introduces the EU-funded project Twinning for Excellence in Adaptive Edge Artificial Intelligence (AIDA4Edge), aimed to advance the state-of-the-art in the design of adaptive neural network accelerators for edge applications. The main goal is to develop a novel hybrid self-adaptive neural network architecture combining spiking and artificial neural networks, and supporting runtime adaptation of network functionality, precision and reliability. Furthermore, we aim to enhance the neural network training by incorporating hardware and quantization constraints in an automated tuning engine.

Abstract Design of optimal and asymptotically optimal quantisation subject to the mean squared error (MSE) criterion is a complex issue, even in the case of uniform scalar quantisation (USQ). The reason is that the MSE distortion dependence on the key designing parameter of USQ for source densities with infinite supports are complex and limit analytical optimisation of USQs. This issue of USQ design has been addressed for some source densities derived from the generalised gamma density. However, to the best of our knowledge, USQ for the one‐sided Rayleigh density has not been studied in detail. This has prompted our research so that this study provides a detailed analysis of USQ for the one‐sided Rayleigh density and proposes an iterative algorithm for its asymptotically optimal design. To estimate signal to quantisation noise ratio, we derive an asymptotic formula having reasonable accuracy for rates higher than 3 bits/sample. Our analysis can be useful in digital‐to‐analogue and analogue‐to‐digital conversion in diversity systems, orthogonal frequency division multiplexing systems and medical image processing.

Abstract A novel 2‐bit adaptive delta modulation (ADM) algorithm is presented based on uniform scalar quantization and fractional linear prediction (FLP) for encoding the signals modelled by a Gaussian probability density function. The study focusses on two major areas: realization of a 2‐bit adaptive quantizer based on Q ‐function approximation that significantly facilitates quantizer design; and implementation of a recently introduced FLP approach with the memory of two samples, which replaces the first‐order linear prediction used in standard ADM algorithms and enables improved performance without increasing transmission costs. It furthermore represents the first implementation of FLP in signal encoding, therefore confirming its applicability in a real signal‐processing scenario. Based on the performance analysis conducted on a real speech signal, the proposed ADM algorithm with FLP is demonstrated to outperform other 2‐bit ADM baselines by a large margin for the gain in signal‐to‐noise ratio achieved over a wide dynamic range of input signals. The results of this research indicate that ADM with adaptive quantization based on Q‐function approximation and adaptive FLP represents a promising solution for encoding/compression of correlated time‐varying signals following the Gaussian distribution.

The main aim of the paper is to provide effective and accurate solutions for the calculation of the support region of the logarithmic companding quantizers. A new solution for the starting point of iterative methods will be proposed, that provide very accurate value of the support region only after one iteration of the iterative method. Based on this new starting point, an accurate closed-form approximate expression for the calculation of the support region will be derived, as a main contribution of the paper. Due to their numerous advantages (robustness, adjustability to the statistical distribution of the input signal), logarithmic quantizers are considered to be used in many topical applications, such as in receivers of 5G wireless systems or in neural networks for quantization of weights and activations.

This study deals with three‐level scalar quantisation of a Gaussian source, followed by Huffman encoder, especially suitable for signal compression purposes. The variance‐matched and variance‐mismatched versions of quantiser are considered. Simple approximate closed‐form expressions for performance evaluation in terms of signal‐to‐quantisation‐noise ratio and bit rate are derived using the corresponding Q ‐function approximations. The accuracy of the derived formulas is tested using the relative error as a measure of performance. It is shown that the derived formulas for variance‐matched (for a reference variance) and variance‐mismatched (in a wide range of input signal variances) provide higher accuracy in comparison to baselines using other available Q ‐function approximations or available approximate formulas for non‐uniform scalar quantisation of Gaussian source.

Linear prediction is extensively used in modeling, compression, coding, and generation of speech signal. Various formulations of linear prediction are available, both in time and frequency domain, which start from different assumptions but result in the same solution. In this letter, we propose a novel, generalized formulation of the optimal low-order linear prediction using the fractional (non-integer) derivatives. The proposed fractional derivative formulation allows for the definition of predictor with versatile behavior based on the order of fractional derivative. We derive the closed-form expressions of the optimal fractional linear predictor with restricted memory, and prove that the optimal first-order and the optimal second-order linear predictors are only its special cases. Furthermore, we empirically prove that the optimal order of fractional derivative can be approximated by the inverse of the predictor memory, and thus, it is a priori known. Therefore, the complexity is reduced by optimizing and transferring only one predictor coefficient, i.e., one parameter less in comparison to the second-order linear predictor, at the same level of performance.

This paper considers the design of robust logarithmic μ-law companding quantizers for the use in analog-to-digital converters in communication system receivers. Quantizers are designed for signals with the Gaussian distribution, since signals at the receivers of communication systems can be well modeled by this type of distribution. In order to reduce energy consumption, low-resolution quantizers are considered (up to 5 bits per sample). The main advantage of these quantizers is a high robustness - they can provide approximately constant SNR in a wide range of signal power (this is very important since the signal power at receivers can vary in wide range due to the fading and other transmission effects). The logarithmic μ-law companding quantizers eliminate the need of using AGC (automatic gain control), which reduces the implementation complexity and increases the speed of the analog-to-digital converters (ADC) due to the absence of AGC delay. Numerical results show that the proposed model achieves a good performance, better than a uniform quantizer, especially in a wide range of signal power. The proposed low-bit ADCs can be used in MIMO and 5G massive MIMO systems, where due to high operating frequencies and a large number of receiving channels (and consequently a large number of ADCs), the reduction of ADC complexity and energy consumption becomes a significant goal.

In this study, a very simple but successful method for designing the pulse‐amplitude modulation (PAM) constellations with non‐equiprobable symbols and not equally spaced amplitudes is presented. The method, inspired by companding quantisation, introduces a constellation design function having a feature that maps the non‐uniformly distributed constellations points into uniform ones. In particular, two classes of piecewise‐linear constellation design function having two segments are provided. In PAM constellation designed by the proposed method, the set of constellation points is divided into two uniform subsets. Each of these subsets is characterised by a certain probability of constellation points and by a constant distance between adjacent points. For this model of PAM constellation, the formulas that enable the elegant assessment of the signal energy and the symbol error probability are derived. Results obtained analytically and verified through simulations show that, in terms of the symbol error probability, the proposed PAM constellations outperforms conventional uniform PAM constellation as well as the other models of piecewise‐uniform PAM constellations.

This study proposes a novel model of hybrid quantiser composed of a uniform scalar quantiser and a non‐uniform optimal companding scalar quantiser, both designed for a Gaussian source. We examine whether by appropriately designing a novel forward adaptive hybrid quantiser with Golomb–Rice code, one can achieve more sophisticated compression and a higher signal to quantisation noise ratio compared with the uniform quantiser with Golomb–Rice code. We observe which value of the bit rate should be chosen to provide high‐quality quantisation. It is shown that the authors compression model can satisfy G.712 recommendation for high‐quality quantisation achieving the compression of 1.68 bit/sample over the G.711 quantiser. In addition, for the average bit rate of 6.32 bit/sample their hybrid quantiser outperforms the uniform quantiser for 1.32 dB. The presented performances of the forward adaptive hybrid quantiser indicate that it should be of theoretical and practical significance in quantisation of the Gaussian source signals.

In this letter, asymptotic formulas are derived for signal to quantization noise ratio (SQNR) of two unrestricted polar quantizers (UPQs) for the case where designed-for and applied-to Gaussian sources have different variances. It is shown that the limitation in the variance mismatch analysis observed in one of the UPQs can be overcome by the constrained application of asymptotic approximations and by optimizing the last magnitude reconstruction level subsequently. The derived formulas are a useful analytical tool for determining the effect of variance mismatch on SQNR in unrestricted polar quantization for Gaussian source.

An optimized vector-quantization-inspired signal constellation design (OVQ-SCD) suitable for multidimensional optical transport is proposed, in which signal constellation radii transformation function is optimized and near-uniform distribution of points is achieved. The proposed OVQ-SCD is used in a tandem with a hybrid multidimensional coded-modulation scheme employing Slepian sequences as electrical discrete-time basis functions, orthogonal prolate spheroidal wave functions as impulse responses of optical filters in orthogonal-division multiplexing, and spatial modes as optical continuous-time basis functions. It has been shown that the proposed multidimensional coded-modulation schemes based on OVQ-SCDs outperform corresponding counterparts and can be used to enable beyond 10 Pb/s serial optical transport over spatial division multiplexing (SDM) fibers as well as beyond 1 Pb/s transport over SMFs. The proposed OVQ-SCD-based hybrid multidimensional coded modulation scheme can simultaneously solve the problems related to the limited bandwidth of information-infrastructure, high energy consumption, and heterogeneity of network segments; while enabling elastic and dynamic bandwidth allocation.

This letter performs an asymptotic analysis of the switched uniform polar quantizer (SUPQ) composed of$k$asymptotically optimal unrestricted uniform polar quantizers designed for the memoryless Gaussian source. The closed-form formulas are derived for signal to quantization noise ratio (SQNR) and the number of phase levels of the quantizers constituting the SUPQ. It is studied how SQNR depends on the variance mismatch and the number of quantizers$k$. It is shown that with a log-uniform distribution of the variances for which the quantizers constituting the SUPQ are designed one can reduce the variance range of average-taking of SQNR.

In this letter, in order to outperform the existing method for unrestricted polar quantizer (UPQ) design in terms of signal to quantization noise ratio, the asymptotic approximations of Rayleigh distributed function are applied to all magnitude regions of the UPQ, except to the last one. Given the constraint, the UPQ is designed to provide the minimum of the asymptotic mean-squared error distortion for the Gaussian source of unit variance. The effects of this constraint are studied for different bit rates R. The accuracy of the derived formulas is assessed and the reasonable accuracy is observed for R ≥ 2.5 bit/sample.

This study proposes a quantiser, named the combined quantiser, composed of two forward adaptive μ-law quantisers (FAμQs), one designed to quantise signals having restricted amplitude dynamics and the other designed for unrestricted signals. The combined quantiser performs frame-by-frame analysis of the input signal, according to which, for each frame procession, it selects one of the two disposable FAμQs. In order to exploit the higher density of the restricted FAμQ's quantisation levels, the restricted FAμQ is selected when all amplitudes of samples within a frame belong to the restricted FAμQ's support region. Otherwise, the unrestricted FAμQ having the same number of quantisation levels, but a wider support region is selected. To provide a more frequent selection of the restricted FAμQ than the unrestricted one, the disposable quantisers’ support region thresholds are determined to satisfy the total distortion minimisation criterion. The authors show that the combined quantiser designed for the Laplacian source provides gain in the signal to quantisation noise ratio and in the compression over the G.711 quantiser as well as satisfies G.712 Recommendation for high-quality quantisation. This indicates the possibility of practical application of the combined quantiser in the contemporary transmission of Laplacian source signals.

Previous studies of nonlinear prediction of speech have been mostly focused on short-term prediction. This paper presents long-term nonlinear prediction based on second-order Volterra filters. It will be shown that the presented predictor can outperform conventional linear prediction techniques in terms of prediction gain and “whiter” residuals.

In this study, the authors present an efficient technique for compression and coding of memoryless Laplacian sources, which uses variable-length code (VLC). That technique is based on the combination of two companding quantisers in the first case and three companding quantisers in the second case. These quantisers have disjoint support regions, different number of representation levels and different compressor functions. The closed-form expressions are obtained for the distortion, average bit rate and signal to quantisation noise ratio (SQNR). The presented numerical results point out the effects of rate-distortion (R-D) optimisation on the system performances. Since our model assumes the general case of Laplacian distribution, it has wide applications like the coding of speech and images. It is shown that the difference of SQNR of our model and classical companding quantiser based model is 2.8 dB for two quantisers and 4.2 dB in three quantisers model. The authors have also made a comparison between our model, combination of the optimal uniform quantiser and Huffmann lossless coder and combination of optimal companding quantiser and simple lossless coder.

In this paper a new model for compression of Laplacian source is given. This model consists of hybrid quantizer whose output levels are coded with Golomb-Rice code. Hybrid quantizer is combination of uniform and nonuniform quantizer, and it can be considered as generalized quantizer, whose special cases are uniform and nonuniformquantizers. We propose new generalized optimal compression function for companding quantizers. Hybrid quantizer has better performances (smaller bit-rate and complexity for the same quality) than both uniform and nonuniformquantizers, because it joins their good characteristics. Also, hybrid quantizer allows great flexibility, because there are many combinations of number of levels in uniform part and in nonuniformpart, which give similar quality. Each of these combinations has different bit-rate and complexity, so we have freedom to choose combination which is the most appropriate for our application, in regard to quality, bit-rate and complexity. We do not have such freedom of choice when we use uniform or nonuniform quantizers. Until now, it has been thought that uniform quantizer is the most appropriate to use with lossless code, but in this paper we show that combination of hybrid quantizer and lossless code gives better performances. As lossless code we use Golomb-Rice code because it is especially suitable for Laplacian source since it gives average bit-rate very close to the entropy and it is easier for implementation than Huffman code. Golomb-Rice code is used in many modern compression standards. Our model can be used for compression of all signals with Laplacian distribution.