Quantum Gradient‐Based Methods for Learning Deformable Offsets

Mixed precision quantization of silicon optical neural network chip

The convolution neural network (CNN) is a classical neural network with advantages in image processing. The use of multiport optical interferometric linear structures in neural networks has recently attracted a great deal of attention. Here, we use three 3 × 3 reconfigurable optical processors, based on Mach-Zehnder interferometers (MZIs), to implement a two-layer CNN. To circumvent the random phase errors originating from the fabrication process, MZIs are calibrated before the classification experiment. The MNIST datasets and Fashion-MNIST datasets are used to verify the classification accuracy. The optical processor achieves 86.9% accuracy on the MNIST datasets and 79.3% accuracy on the Fashion-MNIST datasets. Experiments show that we can improve the classification accuracy by reducing phase errors of MZIs and photodetector (PD) noises. In the future, our work provides a way to embed the optical processor in CNN to compute matrix multiplication.

Exploring the Impact of Denoising Autoencoder Architectures on Image Retrieval

This research introduces least square adversarial autoencoder (LSAA)-an autoencoder that is able to reconstruct data and also generate data that has characteristics similar to data distribution from the prior distribution LSAA uses least square generative adversarial network loss function on its discriminator. LSAA minimizes Pearson χ <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> divergence between the latent variable distribution and the prior distribution. In this research, a Python program is developed to model LSAA by utilizing MNIST data set and FashionMNIST data set. The program is implemented using PyTarch. All of the programming activities are carried out in the cloud environment provided by the Tokopedia-Universitas Indonesia AI Center, using DGR-1 (GPU Tesla V100) as its computing resource. The experimental results show that the mean squared error of LSAA for MNIST data set and FasbionMNIST data set are 0.0080 and 0.0099, respectively. Furthermore, the Fréchet Inception Distance score of LSAA for MNIST data set and FashionMNIST data set are 11.1280 and 27.5737, respectively. These results indicate that the least square adversarial autoencoder is able to reconstruct the image properly and also able to generate images similar to the training samples.

The world is surrounded by a huge amount of data increasing day by day. The increase of data leads to the presence of high-dimensional data causing a challenge in data mining as it becomes costly in computation time and memory space. Moreover, high-dimensional data can affect the classification accuracy of machine learning algorithms because of the existence of redundant and irrelevant features, and that can cause the Curse of Dimensionality problem. Therefore, Dimensionality Reduction has been introduced to solve the Curse of Dimensionality problem by feature extraction techniques. Since there is a lack in the performance of some machine learning models because of high-dimensional data, this paper introduces two common dimensionality reduction methods, and conducts an empirical comparison between , Principal Component Analysis (PCA) and Auto-Encoders (AE) to study their effect in improving the performance of classification in high-dimensional data. The study uses three classification models, K-Nearest Neighbour(KNN),Support Vector Machine (SVM), and Random Forest (RF) to perform the classification in the MNIST, and Fashion-MNIST datasets. The results have been compared , analyzed, and show that AE has a better effect on improving performance of KNN, and SVM classifiers on MNIST dataset, as the SVM accuracy is improved from 94% to 97% when the AE used to reduce dimension with a percentage of 95%,90%, and 50%.Moreover, KNN accuracy has been improved by AE from 91% to 95% , when dimension is reduced with a percentage of 95%,90%, and 50%.On the other hand, the accuracy of the same classifier has been improved in Fashion-MNIST dataset from 81% to 83%, when dimension is reduced by AE with a percentage of 90% and 50%.

To improve the feature extraction ability and shorten the learning time, fuzzy removing redundancy restricted Boltzmann machine (F3RBM) is developed. The features extracted by F3RBM with unsupervised learning are imported into support vector machine (SVM) to establish F3RBM-SVM model, which achieves fast and high-precision automatic classification of different samples. To expand the feature extraction capability of restricted Boltzmann machine (RBM), the deterministic parameters of control model are replaced by fuzzy numbers in view of the superiority of fuzzy idea and the redundancy removal mechanism is introduced. Comparing the feature similarity of hidden units with the threshold value, if the similarity is greater than the threshold value, they are considered to be redundant units with the same features. The redundant units are removed to achieve further dimension reduction. Finally, the learning speed, feature extraction ability, and classification accuracy of different models are compared in MINIST handwritten dataset, Fashion MNIST dataset, and Olivetti Face dataset. The experimental results show that the feature extraction capability of FRBM and F3RBM is better than that of RBM. When there are a large number of hidden units, the learning speed of F3RBM is obviously faster than that of FRBM. The features extracted from F3RBM are imported into the SVM to build F3RBM-SVM model, which improves the classification accuracy and learning speed than general classifier. When adding other noises, F3RBM-SVM has better robustness than other models.

For edge intelligence applications, this work proposes a tiny spike encoding network embedded with high-speed on-chip encoding capability, which applies the proposed dual-mode Integrate & Fire (IF) neuron model to support different coding schemes. The proposed encoding network was prototyped on a Zynq-7020 FPGA device, with an on-chip encoding speed as high as 2127 frame/s, while dissipating only 69 mW under a 250 MHz clock frequency. Our spiking neural network hardware encoder adopts an eight-core architecture for parallel computing to improve processing speed, supporting three well-known coding schemes, i.e. rate coding, burst coding and time-to-firstspike (TTFS) coding. To verify the performance of different coding schemes realized by our hardware encoder, the widely used MNIST and Fashion-MNIST datasets were selected as benchmark. After encoding, all spiking addressevent representation (AER) data were sent to a two-layer fully connected network with BP-STDP learning rule for training and inference, which was completed on PC software. Finally, three coding schemes (rate coding, burst coding and time-to-first-spike coding) all achieved comparably high classification accuracies on MNIST and Fashion-MNIST datasets (95.87%, 92.23% and 88.22% on MNIST, 83.79%, 82.71% and 73.79% on Fashion-MNIST, respectively).

Catastrophic forgetting in neural networks is a common problem, in which neural networks lose information from previous tasks after training on new tasks. Although adopting a regularization method that preferentially retains the parameters important to the previous task to avoid catastrophic forgetting has a positive effect; existing regularization methods cause the gradient to be near zero because the loss is at the local minimum. To solve this problem, we propose a new continuous learning method with Bayesian parameter updating and weight memory (CL-BPUWM). First, a parameter updating method based on the Bayes criterion is proposed to allow the neural network to gradually obtain new knowledge. The diagonal of the Fisher information matrix is then introduced to significantly minimize computation and increase parameter updating efficiency. Second, we suggest calculating the importance weight by observing how changes in each network parameter affect the model prediction output. In the process of model parameter updating, the Fisher information matrix and the sensitivity of the network are used as the quadratic penalty terms of the loss function. Finally, we apply dropout regularization to reduce model overfitting during training and to improve model generalizability. CL-BPUWM performs very well in continuous learning for classification tasks on CIFAR-100 dataset, CIFAR-10 dataset, and MNIST dataset. On CIFAR-100 dataset, it is 0.8%, 1.03% and 0.75% higher than the best performing regularization method (EWC) in three task partitions. On CIFAR-10 dataset, it is 2.25% higher than the regularization method (EWC) and 0.7% higher than the scaled method (GR). It is 0.66% higher than the regularization method (EWC) on the MNIST dataset. When the CL-BPUWM method was combined with the brain-inspired replay model under the CIFAR-100 and CIFAR-10 datasets, the classification accuracy was 2.35% and 5.38% higher than that of the baseline method, BI-R + SI.

This article presents a resistive random access memory (ReRAM)-based convolutional neural network (CNN) accelerator with a new analog layer normalization (ALN) technique. The proposed ALN can be used to effectively reduce the effect of the conductance variation in ReRAM devices by normalizing the outputs of the vector-matrix multiplication (VMM) in the charge domain. The ALN achieves high energy and hardware efficiencies because it directly processes the normalization of the VMM outputs without storing their values in memory and is merged into the neuron circuit of the accelerator. To verify the effect of the ALN through experiments, a VMM accelerator that consists of two 25 × 25 sized ReRAM arrays and peripheral circuits with ALN is used for a convolution layer with digital signal processing in a field programmable gate array. The MNIST dataset is used to train and inference a CNN employing two VMM accelerators that work as convolution layers in a pipelined manner. Despite the conductance variation of the ReRAM devices, the ALN successfully stabilizes the output distribution of the convolution layer, which improves the classification accuracy of the network. A final classification accuracy for the MNIST and Fashion-MNIST datasets of 96.2% and 83.1% is achieved, respectively, with an energy efficiency of 9.94 tera-operations per second per Watt.

Nowadays, Convolutional Neural Network (CNN) based image recognition is a popular research direction. This study uses the Fashion-Mnist dataset, which is more challenging than the Mnist dataset. aims to add Long short-term memory (LSTM) to the structure of CNN to create a hybrid model of CNN and LSTM, called CNN+LSTM model. This model is used to complete and optimize the image classification problem on Fashion-Mnist dataset. The final image classification accuracy of the obtained model is 91.36%, which still needs to be improved, but the accuracy results are better compared to the accuracy of other models.

A conceptually simple way to classify images is to directly compare test-set data and training-set data. The accuracy of this approach is limited by the method of comparison used, and by the extent to which the training-set data cover configuration space. Here we show that this coverage can be substantially increased using coarse-graining (replacing groups of images by their centroids) and stochastic sampling (using distinct sets of centroids in combination). We use the MNIST and Fashion-MNIST data sets to show that a principled coarse-graining algorithm can convert training images into fewer image centroids without loss of accuracy of classification of test-set images by nearest-neighbor classification. Distinct batches of centroids can be used in combination as a means of stochastically sampling configuration space, and can classify test-set data more accurately than can the unaltered training set. On the MNIST and Fashion-MNIST data sets this approach converts nearest-neighbor classification from a mid-ranking- to an upper-ranking member of the set of classical machine-learning techniques.

A scalable (<130 nm) resistive switching memristor that features both filamentary and interfacial switching aimed at neuromorphic computing is developed in this study. The typically perceived noise or volatility was effectively harnessed as a controlled mechanism for interfacial switching. The multilayer structure for the proposed memristor enhances switching stability by curbing ionic overmigration and mitigating leakage paths. Furthermore, the memristors showcased their reliability by demonstrating more than 15 M cycles in the filamentary mode and 1 M pulses in the interfacial mode. Additionally, retention tests at 85 °C for 104 s confirmed the stability across different states, affirming its reliability as a nonvolatile CMOS-compatible element. While many studies validate performance solely on the MNIST data set, this work also evaluates more complex data sets, demonstrating the robustness of the demonstrated memristor in supervised learning. Specifically, supervised learning simulations on MNIST and fashion MNIST data sets indicated a high learning rate with <4% deviations from numerical training, while offline inference trained on CIFAR-10 and CIFAR-100 data sets revealed <2.5% and <7% deviations caused by programing error accumulation, even with increased memristor counts for these highly complex data sets. Unsupervised learning via spike-timing-dependent plasticity further highlights the potential of the developed memristor in bridging artificial and biological paradigms, offering a significant advance toward efficient and biologically inspired computing architectures.

Nonparametric dimensionality reduction techniques, such as t-distributed Stochastic Neighbor Embedding (t-SNE) and uniform manifold approximation and projection (UMAP), are proficient in providing visualizations for data sets of fixed sizes. However, they cannot incrementally map and insert new data points into an already provided data visualization. We present self-organizing nebulous growths (SONG), a parametric nonlinear dimensionality reduction technique that supports incremental data visualization, i.e., incremental addition of new data while preserving the structure of the existing visualization. In addition, SONG is capable of handling new data increments, no matter whether they are similar or heterogeneous to the already observed data distribution. We test SONG on a variety of real and simulated data sets. The results show that SONG is superior to Parametric t-SNE, t-SNE, and UMAP in incremental data visualization. Especially, for heterogeneous increments, SONG improves over Parametric t-SNE by 14.98% on the Fashion MNIST data set and 49.73% on the MNIST data set regarding the cluster quality measured by the adjusted mutual information scores. On similar or homogeneous increments, the improvements are 8.36% and 42.26%, respectively. Furthermore, even when the abovementioned data sets are presented all at once, SONG performs better or comparable to UMAP and superior to t-SNE. We also demonstrate that the algorithmic foundations of SONG render it more tolerant to noise compared with UMAP and t-SNE, thus providing greater utility for data with high variance, high mixing of clusters, or noise.

Connectome, the complete wiring diagram of the nervous system of an organism, is the biological substrate of the mind. While biological neural networks are crucial to the understanding of neural computation mechanisms, recent artificial neural networks (ANNs) have been developed independently from the study of real neural networks. Computational scientists are searching for various ANN architectures to improve machine learning since the architectures are associated with the accuracy of ANNs. A recent study used the hermaphrodite Caenorhabditis elegans (C. elegans) connectome for image classification tasks, where the edge directions were changed to construct a directed acyclic graph (DAG). In this study, we used the whole-animal connectomes of C. elegans hermaphrodite and male to construct a DAG that preserves the chief information flow in the connectomes and trained them for image classification of MNIST and fashion-MNIST datasets. The connectome-inspired neural networks exhibited over 99.5% and 92.6% of accuracy for MNIST and fashion-MNIST datasets, respectively, which increased from the previous study. Together, we conclude that realistic biological neural networks provide the basis of a plausible ANN architecture. This study suggests that biological networks can provide new inspiration to improve artificial intelligences (AIs).

We demonstrate significant improvements in the inference accuracy of diffractive optical neural networks and report that a five-layer, phase-only (or amplitude/phase) modulation diffractive network can achieve 97.18% (97.81%) and 89.13% (89.32%) blind-testing accuracy for MNIST and Fashion-MNIST datasets, respectively. Moreover, the integration of diffractive neural networks with electronic deep neural networks is investigated. Using a single fully-connected layer on the electronic part and a five-layer, phase-only diffractive neural network at the optical front-end, we achieved blind-testing accuracies of 98.71% and 90.04% for MNIST and Fashion-MNIST datasets, respectively, despite a >7.8-fold reduction in the number of pixels at the opto-electronic sensor-array.