Convolutional neural networks (CNNs) have demonstrated remarkable success in vision-related tasks. However, their susceptibility to failing when inputs deviate from the training distribution is well-documented. Recent studies suggest that CNNs exhibit a bias toward texture instead of object shape in image classification tasks, and that background information may affect predictions. This paper investigates the ability of CNNs to adapt to different color distributions in an image while maintaining context and background. The results of our experiments on modified MNIST and FashionMNIST data demonstrate that changes in color can substantially affect classification accuracy. The paper explores the effects of various regularization techniques on generalization error across datasets and proposes a minor architectural modification utilizing the dropout regularization in a novel way that enhances model reliance on color-invariant intensity-based features for improved classification accuracy. Overall, this work contributes to ongoing efforts to understand the limitations and challenges of CNNs in image classification tasks and offers potential solutions to enhance their performance.
相關內容
One of the most surprising puzzles in neural network generalisation is grokking: a network with perfect training accuracy but poor generalisation will, upon further training, transition to perfect generalisation. We propose that grokking occurs when the task admits a generalising solution and a memorising solution, where the generalising solution is slower to learn but more efficient, producing larger logits with the same parameter norm. We hypothesise that memorising circuits become more inefficient with larger training datasets while generalising circuits do not, suggesting there is a critical dataset size at which memorisation and generalisation are equally efficient. We make and confirm four novel predictions about grokking, providing significant evidence in favour of our explanation. Most strikingly, we demonstrate two novel and surprising behaviours: ungrokking, in which a network regresses from perfect to low test accuracy, and semi-grokking, in which a network shows delayed generalisation to partial rather than perfect test accuracy.
We propose a novel algorithm for solving the composite Federated Learning (FL) problem. This algorithm manages non-smooth regularization by strategically decoupling the proximal operator and communication, and addresses client drift without any assumptions about data similarity. Moreover, each worker uses local updates to reduce the communication frequency with the server and transmits only a $d$-dimensional vector per communication round. We prove that our algorithm converges linearly to a neighborhood of the optimal solution and demonstrate the superiority of our algorithm over state-of-the-art methods in numerical experiments.
Recent architectural developments have enabled recurrent neural networks (RNNs) to reach and even surpass the performance of Transformers on certain sequence modeling tasks. These modern RNNs feature a prominent design pattern: linear recurrent layers interconnected by feedforward paths with multiplicative gating. Here, we show how RNNs equipped with these two design elements can exactly implement (linear) self-attention, the main building block of Transformers. By reverse-engineering a set of trained RNNs, we find that gradient descent in practice discovers our construction. In particular, we examine RNNs trained to solve simple in-context learning tasks on which Transformers are known to excel and find that gradient descent instills in our RNNs the same attention-based in-context learning algorithm used by Transformers. Our findings highlight the importance of multiplicative interactions in neural networks and suggest that certain RNNs might be unexpectedly implementing attention under the hood.
Benefiting from the development of deep learning, text-to-speech (TTS) techniques using clean speech have achieved significant performance improvements. The data collected from real scenes often contains noise and generally needs to be denoised by speech enhancement models. Noise-robust TTS models are often trained using the enhanced speech, which thus suffer from speech distortion and background noise that affect the quality of the synthesized speech. Meanwhile, it was shown that self-supervised pre-trained models exhibit excellent noise robustness on many speech tasks, implying that the learned representation has a better tolerance for noise perturbations. In this work, we therefore explore pre-trained models to improve the noise robustness of TTS models. Based on HiFi-GAN, we first propose a representation-to-waveform vocoder, which aims to learn to map the representation of pre-trained models to the waveform. We then propose a text-to-representation FastSpeech2 model, which aims to learn to map text to pre-trained model representations. Experimental results on the LJSpeech and LibriTTS datasets show that our method outperforms those using speech enhancement methods in both subjective and objective metrics. Audio samples are available at: //zqs01.github.io/rep2wav.
Trojans are one of the most threatening network attacks currently. HTTP-based Trojan, in particular, accounts for a considerable proportion of them. Moreover, as the network environment becomes more complex, HTTP-based Trojan is more concealed than others. At present, many intrusion detection systems (IDSs) are increasingly difficult to effectively detect such Trojan traffic due to the inherent shortcomings of the methods used and the backwardness of training data. Classical anomaly detection and traditional machine learning-based (TML-based) anomaly detection are highly dependent on expert knowledge to extract features artificially, which is difficult to implement in HTTP-based Trojan traffic detection. Deep learning-based (DL-based) anomaly detection has been locally applied to IDSs, but it cannot be transplanted to HTTP-based Trojan traffic detection directly. To solve this problem, in this paper, we propose a neural network detection model (HSTF-Model) based on hierarchical spatiotemporal features of traffic. Meanwhile, we combine deep learning algorithms with expert knowledge through feature encoders and statistical characteristics to improve the self-learning ability of the model. Experiments indicate that F1 of HSTF-Model can reach 99.4% in real traffic. In addition, we present a dataset BTHT consisting of HTTP-based benign and Trojan traffic to facilitate related research in the field.
Neural networks have been able to generate high-quality single-sentence speech with substantial expressiveness. However, it remains a challenge concerning paragraph-level speech synthesis due to the need for coherent acoustic features while delivering fluctuating speech styles. Meanwhile, training these models directly on over-length speech leads to a deterioration in the quality of synthesis speech. To address these problems, we propose a high-quality and expressive paragraph speech synthesis system with a multi-step variational autoencoder. Specifically, we employ multi-step latent variables to capture speech information at different grammatical levels before utilizing these features in parallel to generate speech waveform. We also propose a three-step training method to improve the decoupling ability. Our model was trained on a single-speaker French audiobook corpus released at Blizzard Challenge 2023. Experimental results underscore the significant superiority of our system over baseline models.
Deep neural networks have shown remarkable performance when trained on independent and identically distributed data from a fixed set of classes. However, in real-world scenarios, it can be desirable to train models on a continuous stream of data where multiple classification tasks are presented sequentially. This scenario, known as Continual Learning (CL) poses challenges to standard learning algorithms which struggle to maintain knowledge of old tasks while learning new ones. This stability-plasticity dilemma remains central to CL and multiple metrics have been proposed to adequately measure stability and plasticity separately. However, none considers the increasing difficulty of the classification task, which inherently results in performance loss for any model. In that sense, we analyze some limitations of current metrics and identify the presence of setup-induced forgetting. Therefore, we propose new metrics that account for the task's increasing difficulty. Through experiments on benchmark datasets, we demonstrate that our proposed metrics can provide new insights into the stability-plasticity trade-off achieved by models in the continual learning environment.
In this paper, we propose a human trajectory prediction model that combines a Long Short-Term Memory (LSTM) network with an attention mechanism. To do that, we use attention scores to determine which parts of the input data the model should focus on when making predictions. Attention scores are calculated for each input feature, with a higher score indicating the greater significance of that feature in predicting the output. Initially, these scores are determined for the target human position, velocity, and their neighboring individual's positions and velocities. By using attention scores, our model can prioritize the most relevant information in the input data and make more accurate predictions. We extract attention scores from our attention mechanism and integrate them into the trajectory prediction module to predict human future trajectories. To achieve this, we introduce a new neural layer that processes attention scores after extracting them and concatenates them with positional information. We evaluate our approach on the publicly available ETH and UCY datasets and measure its performance using the final displacement error (FDE) and average displacement error (ADE) metrics. We show that our modified algorithm performs better than the Social LSTM in predicting the future trajectory of pedestrians in crowded spaces. Specifically, our model achieves an improvement of 6.2% in ADE and 6.3% in FDE compared to the Social LSTM results in the literature.
To create effective data visualizations, it helps to represent data using visual features in intuitive ways. When visualization designs match observer expectations, visualizations are easier to interpret. Prior work suggests that several factors influence such expectations. For example, the dark-is-more bias leads observers to infer that darker colors map to larger quantities, and the opaque-is-more bias leads them to infer that regions appearing more opaque (given the background color) map to larger quantities. Previous work suggested that the background color only plays a role if visualizations appear to vary in opacity. The present study challenges this claim. We hypothesized that the background color modulate inferred mappings for colormaps that should not appear to vary in opacity (by previous measures) if the visualization appeared to have a "hole" that revealed the background behind the map (hole hypothesis). We found that spatial aspects of the map contributed to inferred mappings, though the effects were inconsistent with the hole hypothesis. Our work raises new questions about how spatial distributions of data influence color semantics in colormap data visualizations.
Recent advances in 3D fully convolutional networks (FCN) have made it feasible to produce dense voxel-wise predictions of volumetric images. In this work, we show that a multi-class 3D FCN trained on manually labeled CT scans of several anatomical structures (ranging from the large organs to thin vessels) can achieve competitive segmentation results, while avoiding the need for handcrafting features or training class-specific models. To this end, we propose a two-stage, coarse-to-fine approach that will first use a 3D FCN to roughly define a candidate region, which will then be used as input to a second 3D FCN. This reduces the number of voxels the second FCN has to classify to ~10% and allows it to focus on more detailed segmentation of the organs and vessels. We utilize training and validation sets consisting of 331 clinical CT images and test our models on a completely unseen data collection acquired at a different hospital that includes 150 CT scans, targeting three anatomical organs (liver, spleen, and pancreas). In challenging organs such as the pancreas, our cascaded approach improves the mean Dice score from 68.5 to 82.2%, achieving the highest reported average score on this dataset. We compare with a 2D FCN method on a separate dataset of 240 CT scans with 18 classes and achieve a significantly higher performance in small organs and vessels. Furthermore, we explore fine-tuning our models to different datasets. Our experiments illustrate the promise and robustness of current 3D FCN based semantic segmentation of medical images, achieving state-of-the-art results. Our code and trained models are available for download: //github.com/holgerroth/3Dunet_abdomen_cascade.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191