In this paper, we present results of processing Dynamic Vision Sensor (DVS) recordings of visual patterns with a retinal model based on foveal-pit inspired Difference of Gaussian (DoG) filters. A DVS sensor was stimulated with varying number of vertical white and black bars of different spatial frequencies moving horizontally at a constant velocity. The output spikes generated by the DVS sensor were applied as input to a set of DoG filters inspired by the receptive field structure of the primate visual pathway. In particular, these filters mimic the receptive fields of the midget and parasol ganglion cells (spiking neurons of the retina) that sub-serve the photo-receptors of the foveal-pit. The features extracted with the foveal-pit model are used for further classification using a spiking convolutional neural network trained with a backpropagation variant adapted for spiking neural networks.

Recent researches show that deep learning model is susceptible to backdoor attacks where the backdoor embedded in the model will be triggered when a backdoor instance arrives. In this paper, a novel backdoor detection method based on adversarial examples is proposed. The proposed method leverages intentional adversarial perturbations to detect whether the image contains a trigger, which can be applied in two scenarios (sanitize the training set in training stage and detect the backdoor instances in inference stage). Specifically, given an untrusted image, the adversarial perturbation is added to the input image intentionally, if the prediction of model on the perturbed image is consistent with that on the unperturbed image, the input image will be considered as a backdoor instance. The proposed adversarial perturbation based method requires low computational resources and maintains the visual quality of the images. Experimental results show that, the proposed defense method reduces the backdoor attack success rates from 99.47%, 99.77% and 97.89% to 0.37%, 0.24% and 0.09% on Fashion-MNIST, CIFAR-10 and GTSRB datasets, respectively. Besides, the proposed method maintains the visual quality of the image as the added perturbation is very small. In addition, for attacks under different settings (trigger transparency, trigger size and trigger pattern), the false acceptance rates of the proposed method are as low as 1.2%, 0.3% and 0.04% on Fashion-MNIST, CIFAR-10 and GTSRB datasets, respectively, which demonstrates that the proposed method can achieve high defense performance against backdoor attacks under different attack settings.

With the advent of Artificial Intelligence (AI)-empowered communications, industry, academia, and standardization organizations are progressing on the definition of mechanisms and procedures to address the increasing complexity of future 5G and beyond communications. In this context, the International Telecommunication Union (ITU) organized the first AI for 5G Challenge to bring industry and academia together to introduce and solve representative problems related to the application of Machine Learning (ML) to networks. In this paper, we present the results gathered from Problem Statement~13 (PS-013), organized by Universitat Pompeu Fabra (UPF), which primary goal was predicting the performance of next-generation Wireless Local Area Networks (WLANs) applying Channel Bonding (CB) techniques. In particular, we overview the ML models proposed by participants (including Artificial Neural Networks, Graph Neural Networks, Random Forest regression, and gradient boosting) and analyze their performance on an open dataset generated using the IEEE 802.11ax-oriented Komondor network simulator. The accuracy achieved by the proposed methods demonstrates the suitability of ML for predicting the performance of WLANs. Moreover, we discuss the importance of abstracting WLAN interactions to achieve better results, and we argue that there is certainly room for improvement in throughput prediction through ML.

Studying the implicit regularization effect of the nonlinear training dynamics of neural networks (NNs) is important for understanding why over-parameterized neural networks often generalize well on real dataset. Empirically, existing works have shown that weights of NNs condense on isolated orientations with a small initialization. The condensation dynamics implies that NNs can learn features from the training data with a network configuration effectively equivalent to a much smaller network during the training. In this work, we show that the multiple roots of activation function at origin is a key factor to understanding the condensation at the initial stage of training. Our experiments suggest that the maximal number of condensed orientations is twice of the multiplicity. Our theoretical analysis confirms experiments for two cases, one is for the activation function of multiplicity one and the other is for the one-dimensional input. This work makes a step towards understanding how small initialization implicitly leads NNs to condensation at initial stage of training, which lays a solid foundation for the future study of the nonlinear dynamics of NNs and its implicit regularization effect at a later stage of training.

Current out-of-distribution detection approaches usually present special requirements (e.g., collecting outlier data and hyperparameter validation) and produce side effects (classification accuracy drop and slow/inefficient inferences). Recently, entropic out-of-distribution detection has been proposed as a seamless approach (i.e., a solution that avoids all the previously mentioned drawbacks). The entropic out-of-distribution detection solution comprises the IsoMax loss for training and the entropic score for out-of-distribution detection. The IsoMax loss works as a SoftMax loss drop-in replacement because swapping the SoftMax loss with the IsoMax loss requires no changes in the model's architecture or training procedures/hyperparameters. In this paper, we propose to perform what we call an isometrization of the distances used in the IsoMax loss. Additionally, we propose to replace the entropic score with the minimum distance score. Our experiments showed that these simple modifications increase out-of-distribution detection performance while keeping the solution seamless.

We have presented a Spiking Convolutional Neural Network (SCNN) that incorporates retinal foveal-pit inspired Difference of Gaussian filters and rank-order encoding. The model is trained using a variant of the backpropagation algorithm adapted to work with spiking neurons, as implemented in the Nengo library. We have evaluated the performance of our model on two publicly available datasets - one for digit recognition task, and the other for vehicle recognition task. The network has achieved up to 90% accuracy, where loss is calculated using the cross-entropy function. This is an improvement over around 57% accuracy obtained with the alternate approach of performing the classification without any kind of neural filtering. Overall, our proof-of-concept study indicates that introducing biologically plausible filtering in existing SCNN architecture will work well with noisy input images such as those in our vehicle recognition task. Based on our results, we plan to enhance our SCNN by integrating lateral inhibition-based redundancy reduction prior to rank-ordering, which will further improve the classification accuracy by the network.

Finding parameters in a deep neural network (NN) that fit training data is a nonconvex optimization problem, but a basic first-order optimization method (gradient descent) finds a global optimizer with perfect fit (zero-loss) in many practical situations. We examine this phenomenon for the case of Residual Neural Networks (ResNet) with smooth activation functions in a limiting regime in which both the number of layers (depth) and the number of neurons in each layer (width) go to infinity. First, we use a mean-field-limit argument to prove that the gradient descent for parameter training becomes a gradient flow for a probability distribution that is characterized by a partial differential equation (PDE) in the large-NN limit. Next, we show that the solution to the PDE converges in the training time to a zero-loss solution. Together, these results imply that the training of the ResNet gives a near-zero loss if the ResNet is large enough. We give estimates of the depth and width needed to reduce the loss below a given threshold, with high probability.

In the past decade the mathematical theory of machine learning has lagged far behind the triumphs of deep neural networks on practical challenges. However, the gap between theory and practice is gradually starting to close. In this paper I will attempt to assemble some pieces of the remarkable and still incomplete mathematical mosaic emerging from the efforts to understand the foundations of deep learning. The two key themes will be interpolation, and its sibling, over-parameterization. Interpolation corresponds to fitting data, even noisy data, exactly. Over-parameterization enables interpolation and provides flexibility to select a right interpolating model. As we will see, just as a physical prism separates colors mixed within a ray of light, the figurative prism of interpolation helps to disentangle generalization and optimization properties within the complex picture of modern Machine Learning. This article is written with belief and hope that clearer understanding of these issues brings us a step closer toward a general theory of deep learning and machine learning.

Instance-based interpretation methods have been widely studied for supervised learning methods as they help explain how black box neural networks predict. However, instance-based interpretations remain ill-understood in the context of unsupervised learning. In this paper, we investigate influence functions [20], a popular instance-based interpretation method, for a class of deep generative models called variational auto-encoders (VAE). We formally frame the counter-factual question answered by influence functions in this setting, and through theoretical analysis, examine what they reveal about the impact of training samples on classical unsupervised learning methods. We then introduce VAE-TracIn, a computationally efficient and theoretically sound solution based on Pruthi et al. [28], for VAEs. Finally, we evaluate VAE-TracIn on several real world datasets with extensive quantitative and qualitative analysis.

Self-supervised learning and pre-training strategies have developed over the last few years especially for Convolutional Neural Networks (CNNs). Recently application of such methods can also be noticed for Graph Neural Networks (GNNs). In this paper, we have used a graph based self-supervised learning strategy with different loss functions (Barlow Twins[ 7], HSIC[ 4], VICReg[ 1]) which have shown promising results when applied with CNNs previously. We have also proposed a hybrid loss function combining the advantages of VICReg and HSIC and called it as VICRegHSIC. The performance of these aforementioned methods have been compared when applied to two different datasets namely MUTAG and PROTEINS. Moreover, the impact of different batch sizes, projector dimensions and data augmentation strategies have also been explored. The results are preliminary and we will be continuing to explore with other datasets.