This paper introduces a novel approach to measuring privacy risks in deep computer vision models based on intermediate outputs.
problem The exposure of intermediate results in hidden layers of deep computer vision models poses significant privacy concerns.
method The approach leverages Degrees of Freedom (DoF) to evaluate the amount of information retained in each layer and combines this with the rank of the Jacobian matrix to assess sensitivity to input variations.
result The proposed framework provides deeper insights into privacy risks associated with intermediate representations without requiring adversarial attack simulations.
Self-supervised and supervised methods learn similar intermediate visual representations but diverge in final layers.
problem Comparing self-supervised and supervised methods for visual learning.
method Comparison of contrastive self-supervised and supervised methods on simple image data.
result Contrastive and supervised methods learn similar intermediate representations but diverge in final layers.
Paper analyzes why deeper layers of ViTs perform worse on out-of-distribution tasks.
problem Performance degradation of intermediate layers in ViTs under distribution shift.
method Extensive linear probing experiments across various benchmarks and fine-grained analysis of transformer modules.
result Probing feedforward network activations yields best performance under significant distribution shift.
Improves interpretability of neural network intermediate layers by making concept activations more robust and effective.
problem Challenges in interpreting neural network decisions and learning in intermediate layers due to opacity and shared interactions.
method Proposes A-CAV to increase effectiveness and employs Gram-Schmidt process to improve robustness.
result Significant improvement in recall rate of concept images from 18.35% to 76.83% for VGG16, with reduced variance in recall across different random seeds.
Freezing intermediate layers reduces DNN inference latency.
problem High latency in deep learning inference due to computationally expensive models.
method Introduce approximate caching at each intermediate layer to avoid running all layers for many inference requests.
result Reduces the number of effective layers by half for 91.58% of CIFAR-10 requests.
A robot learns to classify images with limited perception using a layered reinforcement learning approach.
problem Image classification for robots with partial perception.
method Three-layer architecture using deep reinforcement learning, including meta-layer, action-layer, and classification-layer.
result The method achieves high accuracy on the MNIST dataset and provides explainability of the agent's decision-making process.
The paper analyzes different neural network architectures for 3D point cloud processing.
problem Understanding the effects of specific intermediate-layer network architectures on 3D point cloud processing.
method Designing five metrics to diagnose various types of DNNs and conducting comparative studies.
result The hypotheses on the effects of specific intermediate-layer network architectures on the representation capacity of DNNs are verified.
We consider the problem of estimating a consensus community structure by combining information from multiple layers of a multi-layer network using methods based on the spectral clustering or a low-rank matrix factorization. As a general theme, these "intermediate fusion" methods involve obtaining a low column rank matr…
Neural network models have a reputation for being black boxes. We propose to monitor the features at every layer of a model and measure how suitable they are for classification. We use linear classifiers, which we refer to as "probes", trained entirely independently of the model itself. This helps us better understand …
Neural networks are vulnerable to adversarial examples, malicious inputs crafted to fool trained models. Adversarial examples often exhibit black-box transfer, meaning that adversarial examples for one model can fool another model. However, adversarial examples are typically overfit to exploit the particular architectu…
GANs can improve image reconstruction by using intermediate layers.
problem Improving the quality of image reconstruction using GANs.
method Exploiting the representation in intermediate layers of the generator.
result Intermediate layers in GANs can represent natural images with high visual fidelity.
LSTMs and other RNN variants have shown strong performance on character-level language modeling. These models are typically trained using truncated backpropagation through time, and it is common to assume that their success stems from their ability to remember long-term contexts. In this paper, we show that a deep (64-…
To reduce the large computation and storage cost of a deep convolutional neural network, the knowledge distillation based methods have pioneered to transfer the generalization ability of a large (teacher) deep network to a light-weight (student) network. However, these methods mostly focus on transferring the probabili…
The paper introduces REQNNs for robust 3D point cloud processing.
problem 3D point cloud processing robustness to rotations.
method Revised neural networks using quaternion features for rotation-equivariance.
result REQNNs exhibit higher rotation robustness compared to original networks.
Recently popularized graph neural networks achieve the state-of-the-art accuracy on a number of standard benchmark datasets for graph-based semi-supervised learning, improving significantly over existing approaches. These architectures alternate between a propagation layer that aggregates the hidden states of the local…
Proposes QNN to protect input privacy in neural networks.
problem Protecting input privacy in neural networks.
method Quaternion-valued neural network (QNN) to hide input information.
result QNN effectively protects input privacy without significant accuracy loss.
Improved U-Nets with various intermediate blocks enhance singing voice separation.
problem Improving singing voice separation accuracy using U-Net architectures.
method Implemented and compared U-Nets with different intermediate spectrogram transformation blocks.
result A specific block type achieves state-of-the-art SDR by 0.9 dB.
A new KD layer lets student models learn and apply teacher knowledge explicitly.
problem Implicit action of traditional KD on student's feature transform limits its use in intermediate layers.
method Proposes a learnable KD layer that explicitly embeds teacher's knowledge in feature transform.
result Improves KD with two abilities: leveraging teacher's knowledge and feeding forward knowledge deeper.
Neural networks are vulnerable to adversarial examples, malicious inputs crafted to fool trained models. Adversarial examples often exhibit black-box transfer, meaning that adversarial examples for one model can fool another model. However, adversarial examples may be overfit to exploit the particular architecture and …
DFS dynamically decides bitwidths for layers to balance accuracy and efficiency.
problem Balancing model accuracy and inference speed for deep networks.
method Dynamic Fractional Skipping (DFS) framework that assigns bitwidths to layers for input-adaptive inference.
result DFS achieves superior tradeoff between computational cost and model accuracy.
The study proves Gaussian universality of deep random features learning.
problem Understanding the test error in deep random features learning.
method Proving Gaussian universality of test error in ridge regression and arbitrary convex losses.
result Sharp asymptotic formula for test error in ridge regression setting.
Neural networks benefit from intermediate representations, reducing sample complexity.
problem Understanding how neural networks leverage intermediate representations for hierarchical learning.
method Fixed, randomly initialized neural network as a representation function, compared with raw inputs and other trainable networks.
result Neural representations can achieve improved sample complexities compared to raw inputs, especially for low-rank polynomials.
Researchers analyze the geometric and statistical properties of transformer model representations.
problem Understanding the semantic structure of large transformer models across various data types.
method Characterization of geometric and statistical properties through analysis of intrinsic dimension and neighbor composition.
result The semantic information of the dataset is better expressed at the end of the first peak in transformer models.
A new method aggregates generative classifiers to resist adversarial attacks.
problem Adversarial attacks on deep neural networks.
method Rank-aggregating ensemble of generative classifiers trained on intermediate layer responses.
result The ensemble of generative classifiers shows robustness to adversarial attacks.
Detects adversarial examples using autoencoders at hidden layers.
problem Identifying adversarial examples in deep neural networks.
method Trains autoencoders on intermediate layers to detect deviations from true data.
result Outperforms state-of-the-art methods in both supervised and unsupervised settings.
Improves conditions for mode connectivity in deep neural networks.
problem Ensuring mode connectivity in deep neural networks under practical conditions.
method Exploiting feature quality and mild over-parameterization conditions.
result Connectivity of solutions found by stochastic gradient descent confirmed under new conditions.
Efficiently detects anomalies in videos with reduced computation.
problem Anomaly detection in videos for robust and efficient systems.
method Powers-of-two weights and denoising for CNN simplification and robustness.
result 10% faster detection with comparable accuracy and robustness.
Graph cross network improves graph classification accuracy.
problem Improving graph classification accuracy.
method Graph cross network (GXN) with vertex infomax pooling (VIPool) and feature-crossing layer.
result Improves graph classification accuracy by 2.12% and 1.15%.
Adversarial training has been shown to regularize deep neural networks in addition to increasing their robustness to adversarial examples. However, its impact on very deep state of the art networks has not been fully investigated. In this paper, we present an efficient approach to perform adversarial training by pertur…
Improves adversarial robustness of DEQ models by regulating neural dynamics.
problem Limited adversarial robustness of DEQ models.
method Interprets DEQs as neural dynamics, uses entropy reduction and random intermediate states.
result Significantly increases adversarial robustness of DEQ models.
We present a method for feature interpretation that makes use of recent advances in autoregressive density estimation models to invert model representations. We train generative inversion models to express a distribution over input features conditioned on intermediate model representations. Insights into the invariance…
Our work presents extensive empirical evidence that layer rotation, i.e. the evolution across training of the cosine distance between each layer's weight vector and its initialization, constitutes an impressively consistent indicator of generalization performance. In particular, larger cosine distances between final an…
Formalizes explanations as blending input and model output.
problem Creating clear and consistent explanations for model predictions.
method Defines properties of explanation functions and links them to model layers.
result Consistency of activations across layers implies consistency of explanations.
Deep neural network algorithms are difficult to analyze because they lack structure allowing to understand the properties of underlying transforms and invariants. Multiscale hierarchical convolutional networks are structured deep convolutional networks where layers are indexed by progressively higher dimensional attrib…
RENNs protect input privacy by rotating d-ary features.
problem Protecting input privacy from intermediate-layer features.
method Rotation-equivariant neural networks using d-ary vectors/tensors.
result RENNs effectively hide input information without degrading output accuracy.
Our proposed deeply-supervised nets (DSN) method simultaneously minimizes classification error while making the learning process of hidden layers direct and transparent. We make an attempt to boost the classification performance by studying a new formulation in deep networks. Three aspects in convolutional neural netwo…
Transformers solve parity problems efficiently with step-by-step reasoning.
problem Training transformers to solve complex, recursive problems like parity.
method Training a one-layer transformer to solve k-parity, incorporating intermediate parities into the loss function, and using teacher forcing or augmented data. result Transformers can learn parity in one gradient update with intermediate supervision or self-consistency checks.
Deep networks improve by progressively refining approximations at each layer.
problem Standard approximation theory doesn't explain the role of intermediate layers in deep neural networks.
method Developed a mixed-activation architecture with a geometric scale interpretation of depth.
result Each intermediate layer approximates the target function with a geometric rate.
Recently, there is increasing interest and research on the interpretability of machine learning models, for example how they transform and internally represent EEG signals in Brain-Computer Interface (BCI) applications. This can help to understand the limits of the model and how it may be improved, in addition to possi…
Backpropagation-free RL method trains layers using local signals.
problem Vanishing or exploding gradients in backpropagation-based RL.
method Local pairwise distance matching for layer-wise training without backpropagation.
result Backpropagation-free method achieves competitive performance and stability.
Shallow supervised 1-hidden layer neural networks have a number of favorable properties that make them easier to interpret, analyze, and optimize than their deep counterparts, but lack their representational power. Here we use 1-hidden layer learning problems to sequentially build deep networks layer by layer, which ca…
New adversarial attack method based on deep feature distributions.
problem Adversarial attacks on CNN classifiers using output layer information.
method Modeling and exploiting class-wise and layer-wise deep feature distributions.
result Achieves state-of-the-art transfer-based attack results for undefended ImageNet models.
Missing data imputation (MDI) is a fundamental problem in many scientific disciplines. Popular methods for MDI use global statistics computed from the entire data set (e.g., the feature-wise medians), or build predictive models operating independently on every instance. In this paper we propose a more general framework…
Law explains how deep networks separate data for classification.
problem Black-box nature of deep learning limits architecture design and interpretation.
method Studied how deep neural networks process data in intermediate layers.
result Law of geometric data separation emerges in various architectures and datasets.
Proposes PSCs for UQ in deep nets without retraining.
problem Estimating uncertainty in deep nets with a single pass.
method Identifies sensitive, smooth intermediate layer, fits probabilistic model.
result PSCs achieve UQ and OOD detection performance matching existing methods.
Modified RV-coefficient reveals how training affects neural network representations.
problem Understanding how training affects intermediate representations in convolutional neural networks.
method Experimented with modified RV-coefficient (RV2) to compare activation patterns in deep networks trained on varying amounts of data and layers.
result RV2 successfully recovered expected similarity patterns and provided interpretable similarity matrices.
Since the 2007-2009 financial crisis, substantial academic effort has been dedicated to improving our understanding of interbank lending networks (ILNs). Because of data limitations or by choice, the literature largely lacks multiple loan maturities. We employ a complete interbank loan contract dataset to investigate w…
Deep learning can learn compositional functions more efficiently by breaking them into stages.
problem Understanding why deep learning performs better than shallow models in learning compositional functions.
method Analyzed learnability of compositional target functions using a three-layer fitting model trained with layer-wise spectral estimators.
result Learning compositional functions can be simplified by breaking them into stages, reducing the complexity of the learning problem.