literature_review.txt

1708.03880.pdf:  This paper explores the idea of using image quality assessment to guide the training of deep neural networks. The authors believe that this will lead to more reliable models that can deal with low quality images, which are often encountered in real-world scenarios. To test this idea, they propose a label smoothing method that takes into account the quality of training images. Experimental results show that the proposed method is effective in dealing with both low and high quality images. This paper investigates the use of distorted images for training deep neural networks, with the goal of improving the classification performance of low quality images. The authors first train the network with a mix of pristine and distorted data, then use an IQA-based label smoothing technique to fine-tune the network. Experimental results show that the proposed scheme can significantly improve the classification performance of both high and low quality images. The paper discusses a method for training a deep neural network to be robust and confident in classification tasks involving images of varying quality. The method, called IQA-based label smoothing, uses a measure of image quality (IQA) to determine the label distribution for each image. This allows the network to be more confident in its predictions for high quality images and less confident for low quality images, which is in line with human perception. The method is evaluated on the CIFAR-10 dataset and shows improved performance over other methods. The paper evaluates different training strategies for deep neural networks in order to improve their performance on distorted images. The authors find that training on a mixture of pristine and distorted images, with the label distribution maintained as a typical 0-1 distribution, is the most effective approach. This training strategy results in better classification performance on all three types of distortions (blur, noise, and JPEG compression) than training on pristine images alone. The paper explores the use of image quality assessment (IQA) to improve the performance of deep neural networks (DNNs) when training with distorted images. The paper proposes a method of IQA-based label smoothing to regularize the label distribution and improve the performance of the DNNs. The method is evaluated on three types of distortions: blur, noise, and JPEG compression, and results show that the proposed method outperforms the baseline method on all three types of distortions. The paper also shows that the proposed method is effective in reducing the overfitting problem when training with severely distorted images. The paper proposes a quality assessment based label smoothing approach for deep neural network learning. The authors claim that the proposed approach achieves high prediction accuracy across different distortion types and levels. The approach is based on the idea that distorted images should be included in the training process in order to learn a more robust representation. Quality assessment is used to regularize the label distribution of training samples. The performance of the proposed approach is evaluated on image classification tasks. Results show that the proposed approach outperforms the baseline approach.

2001.11355.pdf:  This paper discusses how to construct deep neural networks (DNNs) with a priori knowledge of wireless tasks in order to reduce training complexity. The authors show that two kinds of permutation invariant properties widely existed in wireless tasks can be harnessed to reduce the number of model parameters and hence the sample and computational complexity for training. They find special architecture of DNNs whose input-output relationships satisfy the properties, called permutation invariant DNN (PINN), and augment the data with the properties. By learning the impact of the scale of a wireless system, the size of the constructed PINNs can exibly adapt to the input data dimension. The paper takes predictive resource allocation and interference coordination as examples to show how the PINNs can be employed for learning the optimal policy with unsupervised and supervised learning. Simulations results demonstrate a dramatic gain of the proposed PINNs in terms of reducing training complexity. The paper discusses the potential of using deep learning to reduce training complexity for wireless tasks by harnessing permutation invariance properties. Two permutation invariant properties are introduced, and corresponding DNN architectures are proposed. It is shown how the PINNs can adapt to different input data dimension. The proposed PINNs are demonstrated to reduce the sample and computational complexities compared to the non-structural FC-DNN with same performance. This paper discusses the concept of permutation invariance in the context of resource allocation tasks for wireless networks. The paper provides sufficent and necessary conditions for a function to be permutation invariant, and demonstrates how to construct Deep Neural Network (DNN) architectures to learn permutation invariant functions. This is relevant for resource allocation tasks in wireless networks because the order of users or transmitters/receivers should not impact the solution. The paper discusses how to reduce the number of parameters in a neural network by introducing parameter sharing among blocks. The paper first considers a one-dimensional permutation invariant function, and shows how to construct a neural network with 1D permutation invariance property. The paper then extends the result to a two-dimensional permutation invariant function. The paper discusses how to use a permutation-invariant neural network (PINN) to learn the optimal predictive resource allocation (PRA) policy for a cellular network. The PINN is trained using unsupervised learning, and the network is able to adapt to changes in the number of users over time. The PINN is tested on two case studies, predictive resource allocation and interference coordination, and is shown to outperform traditional methods. This paper presents a method for reducing the computational complexity of resource allocation in a wireless network. The authors formulate the resource allocation problem as a convex optimization problem, which can be solved by interior-point methods. However, the computational complexity of this approach scales with O(KTf)3:5, which is prohibitive. To reduce the on-line computational complexity, the authors train a DNN to learn the optimal resource allocation plan. To avoid the computational complexity in generating labels, the DNN is trained with unsupervised learning. The authors show that the DNN can be used to approximate the optimal solution of the resource allocation problem with reduced computational complexity. The paper presents a method for learning resource allocation policies using deep neural networks (DNNs). The policies are learned by training the DNNs to minimize a cost function that is the empirical form of a Lagrangian function. The policies are tested on two tasks: resource allocation for a mobile data network, and interference coordination for a wireless network. The results show that the learned policies outperform existing methods. The paper discusses a resource allocation method that uses a deep neural network (DNN) to predict the resources needed by users in a cellular network. The DNN is trained using samples generated by a simulated system. The results show that the DNN can accurately predict the resource needs of users, even when the number of users in the system varies. The DNN is also more efficient than a traditional resource allocation method, requiring fewer parameters and less training time. This paper presents a study on the sample and computational complexity of three different DNNs for learning PRA. The three DNNs are PINN-2D- Adp-K, PINN-2D and FC-DNN. The study found that PINN-2D-Adp-K achieves the best performance while PINN-2D has the lowest training complexity. The paper explores the use of deep neural networks (DNNs) for resource allocation and interference coordination in wireless networks. The authors propose a new type of DNN, called a permutation-invariant neural network (PINN), which is designed to be more efficient than traditional DNNs. PINNs are constructed by sharing weights among permutation-invariant blocks, and the authors demonstrate how they can be adapted to different scales of wireless systems. The authors also show how the permutation invariance property can be used for dataset augmentation, which can reduce the time required to generate labels for supervised learning. Simulation results show that the proposed PINNs can achieve the same performance as traditional DNNs with significantly less computational complexity. This paper discusses the use of deep learning for wireless tasks. The authors prove that a function is permutation invariant if the output does not change when the input is permuted. They then show that a function is permutation invariant if it has the form of a deep neural network. The authors use deep learning to optimize wireless resource management. They show that deep learning can be used to improve the performance of wireless systems.

2002.03847.pdf:  This paper proposes a novel logic learning pipeline that combines the advantages of neural networks and logic circuits. The pipeline first trains a neural network on a classification task, and then translates this, first to random forests, and then to AND-Inverter logic. The authors show that this pipeline maintains greater accuracy than naive translations to logic, and minimizes the logic such that it is more interpretable and has decreased hardware cost. The paper demonstrates the utility of the pipeline on two datasets: one from hospitals around the world seeking medical care for gastrointestinal bleeding symptoms, and one from the Veterans Aging Cohort Study. The authors show that the pipeline is able to accurately classify patients in both datasets, and that the resulting logic is more interpretable and easier to verify than logic resulting from other methods. This paper discusses the trade-off between memorization and learning with LogicNet, a network of concatenated look-up tables with multiple layers. The paper also describes how decision trees can be used to improve the accuracy of machine learning models. Finally, the paper describes how advances in deep learning are affecting the field of medicine and health care. The paper discusses a novel pipeline for translating neural networks into logic gates, which can be used for industrial on-chip machine learning designs. The pipeline is inspired by previous work on gradient-boosted decision trees and logic synthesis. The goal of the pipeline is to reduce the number of basic gates required for building a logic circuit, while still maintaining accuracy. The paper discusses how the pipeline can be used for medical data to provide rich summaries of both an entire model and individual features. This paper discusses the use of logic to create a framework for training a neural network. The paper begins by discussing the use of a SAT solver to generate input vectors for testing the logic of the system. The paper then describes the use of a random forest to train the neural network. The paper describes the use of a quantization scheme to reduce the precision of the weights and activations of the neural network. The paper then describes the use of a translation to create a logic circuit from the trained neural network. The paper describes the use of an And-Inverter-Graph to measure the size and complexity of the logic circuit. The paper describes the use of a simulation to gain insights on the performance and generalization capability of the logic circuit. The paper evaluates three different pipelines for translating a neural network into an arithmetic circuit. The first pipeline uses a quantization scheme and trains random forests on the quantized activations. The second pipeline is inspired by FPGA architectures and uses look-up tables instead of random forests. The third pipeline is a "straightforward" translation of the neural network into an arithmetic circuit. The results show that the third pipeline outperforms the other two pipelines in terms of accuracy. The paper presents a novel framework for translating trained neural networks into random forests and then into logic gate representations. The motivation is to combine the learnability of neural networks with the verifiability and implementability of logic. The framework is evaluated on several datasets, and the results show that it outperforms the traditional approach of translating neural networks into arithmetic logic in terms of accuracy and hardware complexity. The paper explores the idea of using machine learning to interpret black box models. The authors argue that this could have a high impact on how the aspect of interpretability in black box models is thought. They provide a novel setup that needs to be further explored, but that can put research forth in the intersection of machine learning and logic circuits. The paper discusses intrinsic methods to detect overfitting in Boolean satisability (SAT) solvers. The authors use a graph neural network (GNN) to learn a SAT solver from single-bit supervision. They then compare the performance of the GNN-based SAT solver to that of a traditional SAT solver. The results show that the GNN-based SAT solver outperforms the traditional SAT solver in terms of accuracy and runtime. This paper introduces a novel pipeline for translating neural networks into logic. The pipeline includes a number of steps for training and translating the neural network into logic. The first step is to train the neural network on the activations of the neural network. The second step is to translate the neural network into logic. The third step is to train the random forest on the activations of the neural network. The fourth step is to translate the random forest into logic. The fifth step is to concatenate the modules to recreate the entire network. The sixth step is to translate the network into an AIG. The paper discusses the use of inverse weighting to prevent class imbalance when training a random forest, and the use of bitwise training to simplify the logic of the forest. The argmax function is used to determine the final class prediction of the Forest. The paper discusses the different ways of training a random forest, and how this affects the logic implementation of the forest. In particular, it compares the bitwise and none-bitwise training methods, and argues that the latter is preferable. The bitwise training method leads to a more complex logic implementation, due to the need to sum up class probabilities across multiple decision trees. The none-bitwise method is simpler, as it only requires a comparator and multiplexer at each branch of the decision tree. The paper presents a method for translating real-valued neural networks into arithmetic circuits. The results of the translation are evaluated using a random forest logic. The results show that the accuracy of the translated neural networks is preserved. The paper explores the use of random forests for logic synthesis. The authors compare the performance of random forests against other logic synthesis methods, and find that random forests outperform other methods in terms of accuracy and synthesis time. The paper presents the results of a random forest logic on the "V ACS Model 1" setup for a quantization scheme of 7 total bits and 3 fractional bits. The results show that the accuracy of the Top-1 and Top-N predictions is 49.83% and 48.62%, respectively. The AIG nodes and levels are also presented. The paper provides the results of a study on the use of LogicNets as an intermediate step in the creation of logic circuits. The study found that LogicNets can learn to represent the input data well, and that the resulting logic circuits can be accurate. The paper presents the results of a logic optimization technique called LogicNet on two different setups. The first setup is called "GIB Model 1" and uses 18 total bits and 10 fractional bits. The second setup is called "GIB Model 2" and uses 6 total bits and 4 fractional bits. For both setups, the LogicNet logic optimization technique achieved good results in terms of accuracy and AIG node count. However, the "GIB Model 2" setup achieved better results than the "GIB Model 1" setup, especially in terms of accuracy. This paper presents the results of a study on the use of a quantization scheme for digital logic circuits. The study found that a quantization scheme of 12 total bits and 8 fractional bits yielded the best results for accuracy and node count. The study also found that a depth of 3 and a width of 2 was the optimal configuration for the quantization scheme. The paper presents the results of a logic synthesis tool called LogicNet on two different hardware setups, "V ACS Model 1" and "V ACS Model 2". For each setup, the tool was used to synthesize a circuit with a quantization scheme of 32 total bits and 16 fractional bits. The results show that LogicNet was able to successfully synthesize the circuits with a high degree of accuracy. This paper discusses the use of logic modules for interpretable neural networks. AIGs are used to generate Boolean equations which can be interpreted by humans. It is shown that this method can be used to generate reports for each node in a neural network, allowing for the analysis of which inputs are most important for each node. This method can also be used to generate reports for intermediate layers or nodes in a neural network. The paper discusses how running SAT-solvers on logic functions can help in system verification. The example used shows that missing dependencies on features can be found which can help in curating data sets. The computing times for MiniSAT to find satisfying arguments for multiple settings of bloated and reduced logic are provided. The paper presents the results of a comparison between three different machine learning models - LogicNet, Arithmetic Circuit, and Random Forest - applied to the task of image classification. The LogicNet model outperformed the other two models in terms of accuracy, with a depth of 4 and a width of 100. However, the Arithmetic Circuit model was faster to train, with a depth of 8 and a width of 4. The Random Forest model was the slowest to train, with a depth of 5 and a width of 6.

2109.01614.pdf:  The paper examines the role of microscopic rules in determining the collective dynamics of a neuronal network. The authors compare the dynamics of two cellular automaton models commonly used to simulate large-scale neuronal networks. They find that the two models exhibit very different collective dynamics as a function of network topology. The results highlight the importance of paying attention to the microscopic rules chosen to model inter-neuronal interactions in large-scale numerical simulations. This paper looks at two different neural network models and compares their effectiveness in terms of mapping the human brain connectome. The first model proposed by Greenberg & Hastings [29] uses a neuronal interaction rule that depends on the state of its presynaptic neighbors, while the second model introduced by Kinouchi & Copelli [15] uses a rule that maintains a prescribed branching of activity on the target neurons. The paper shows that the first model exhibits continuous or discontinuous phase transitions depending on the network topology, while the second model is completely insensitive to it. The paper examines the role of criticality in biological systems, specifically in neural networks. The authors use numerical simulations to study how different activation rules can affect the mapping of real brain connectomes. They find that not all neural models are the same, and that special care should be taken to understand the biases and limitations introduced by different models. The paper discusses two models of brain activity, the Greenberg-Hastings (GH) model and the Kinouchi-Copelli (KC) model. The GH model is a cellular automaton with three states (quiescent, active, and refractory), while the KC model is a branching process with n states. The two models are compared in terms of their ability to reproduce criticality in brain activity. It is found that the KC model is more successful in this regard. In particular, the KC model is able to reproduce the power-law scaling of the distribution of active neurons, while the GH model does not. The paper explores the dynamical repertoire of two neuron models - the GH model and the KC model - under a range of neuronal dynamics and network topologies. The authors use an order parameter (fS) and the connected autocorrelation of the order parameter (AC(1)) to identify and classify the dynamical regimes observed. They find that the GH model undergoes a discontinuous transition, while the KC model exhibits a continuous transition. In addition, the GH model shows a large increase in the variability of the order parameter near the transition, while the KC model does not. These observations point to important dynamical differences between the two models. The paper explores the dynamical behavior of two models, the GH model and the KC model, as a function of network topology. The authors find that both models exhibit a continuous phase transition for a wide range of parameters, but that the GH model also shows a discontinuous phase transition for high values of degree and disorder. The results suggest that the spiking patterns observed in the two models are similar for a wide range of control parameter values. The paper examines the effects of network topology on two simplified models of neuronal activation, the GH model and the KC model. The GH model shows a variety of phase transitions, including both continuous and discontinuous transitions, while the KC model only shows a continuous phase transition. The autocorrelation coefficient AC(1) is used to track the dynamics and identify phase transitions. The key difference between the two models is in the rule that determines how the activity propagates from a given neuron to its connected neighbors. The GH model mimics a discrete integrate-and-fire process taking place in real neurons, while the KC model is a probabilistic contagion-like process. The KC model is insensitive to the network topology. The paper discusses the use of the KC model to study the dynamics of excitable systems. The authors argue that the model is not realistic enough to be used to study real-world systems, as it does not take into account the number and state of post-synaptic neurons. They also argue that the model is not suitable for studying the dynamics of non-random topologies, as it does not accurately capture the behavior of real-world systems. The paper describes the implementation of a model for neuronal activity, called the Greenberg-Hastings (GH) model. The model is implemented in both Fortran 90 and Python 3, and the code is available on GitHub. The model is designed to simulate the activity of neurons in a network, and the results of the simulation are saved in a file. The paper also describes the implementation of a second model, called the Kinouchi-Copelli (KC) model, which is similar to the GH model but with some slight variations. The code for the KC model is also available on GitHub. The paper explores two alternative implementations of the KC model for simulating neuronal activity. The first implementation, KCm, is faster but less accurate, while the second implementation, KCb, is more accurate but slower. The paper also introduces a new implementation, KCl, which is intermediate in terms of accuracy and speed. The paper demonstrates that the KCl implementation is able to reproduce the same results as the KCm and KCb implementations, but is faster due to the fact that only active neurons need to be updated. The paper also shows that the KCl implementation is able to reproduce the same results as the KCm and KCb implementations, but is faster due to the fact that only active neurons need to be updated. The paper also shows that the KCl implementation is able to reproduce the same results as the KCm and KCb implementations, but is faster due to the fact that only active neurons need to be updated.

2207.10723.pdf:  The paper discusses the design of an efficient and scalable CNN-based SoC-FPGA accelerator that takes pre-trained weights with a 16-bit fixed-point quantization and target hardware specification to generate an optimized template. The template analyzes the computational workload, data dependency, and external memory bandwidth and utilizes loop tiling transformation along with dataflow modeling to convert convolutional and fully connected layers into vector multiplication between input and output feature maps, which resulted in a single compute unit on-chip. Furthermore, the accelerator was examined among AlexNet, VGG 16, and LeNet networks and ran at 200-MHz with a peak performance of 230 GOP/s depending on ZYNQ boards and state-space exploration of different compute unit configurations during simulation and synthesis. Lastly, the proposed methodology was benchmarked against the previous development on Ultra96 for higher performance measurement. The paper discusses a proposed methodology for designing an accelerator for convolutional and fully connected layers in a CNN. The methodology uses loop tiling to improve performance and resource utilization. The design is benchmarked against previous designs and shown to improve performance. The paper discusses an algorithm for convolutional neural networks that is based on a window operation. The algorithm is designed to be more efficient than previous methods, and is able to achieve higher performance while using less resources. The paper includes results from simulations that show the improved performance of the proposed algorithm. The paper discusses various ways to optimize FPGA-based accelerator design for deep convolutional neural networks (CNNs). The authors compare and contrast different approaches, including data compression, HLS, and model compression. They conclude that data compression is the most effective way to reduce the size of the CNN accelerator.