New approach detects out-of-distribution inputs without needing OOD samples.
problem Detecting incorrect classification of out-of-distribution inputs in deep neural networks.
method A one-class classifier trained on an early layer's output of the original classifier.
result Substantially better results compared to state-of-the-art approaches.
Sparse Transformers degrade semantic information first, with early layers encoding more.
problem Understanding how sparse Transformers affect learned representations and semantic information.
method Probed Transformers with progressively pruned weights to observe changes in semantic information and model behavior.
result Complex semantic information is first to degrade in sparse Transformers, with early layers encoding more.
SLED improves factuality in LLMs without external knowledge.
problem Unreliable or factually incorrect outputs from large language models.
method Contrasts final layer logits with early layers' logits, uses approximate gradient to refine outputs.
result Consistently improves factual accuracy over existing methods.
CNNs encode data augmentation transformations, especially in early layers.
problem Whether neural network features encode data augmentation transformations.
method Systematic approach using pre-trained vision models to predict augmentation transformations.
result Neural network features encode data augmentation transformations, especially in early layers.
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…
Measures difficulty of predictions to improve deep learning models.
problem Understanding and improving deep learning models.
method Introduces a measure called prediction depth to assess example difficulty.
result Surprising relationships between prediction depth and model performance.
CutMix training technique improves spatial locality in Vision Transformers.
problem Improving spatial locality in Vision Transformers trained from scratch.
method Comparison of Baseline and Modern training protocols on CIFAR-10, CIFAR-100, and Tiny-ImageNet.
result CutMix training component significantly reduces Mean Attention Distance (MAD) in early layers of Vision Transformers.
Deep networks preferentially learn shared features, avoiding memorization in early layers.
problem Understanding how deep neural networks generalize vs. memorize training data.
method Replica-based mean field geometric analysis of deep neural networks.
result Deep layers predominantly memorize, while early layers are minimally affected.
We characterize convolutional neural networks with respect to the relative amount of features per layer. Using a skew normal distribution as a parametrized framework, we investigate the common assumption of monotonously increasing feature-counts with higher layers of architecture designs. Our evaluation on models with …
The early layers of a deep neural net have the fewest parameters, but take up the most computation. In this extended abstract, we propose to only train the hidden layers for a set portion of the training run, freezing them out one-by-one and excluding them from the backward pass. Through experiments on CIFAR, we empiri…
Lower class selectivity makes networks more robust to natural perturbations but more vulnerable to adversarial attacks.
problem Understanding how class selectivity affects robustness to different types of perturbations in neural networks.
method Investigated the relationship between class selectivity and robustness to natural and adversarial perturbations in neural networks.
result Lower class selectivity increases robustness to natural perturbations but decreases robustness to adversarial attacks.
Exact solutions reveal how unbalanced initializations promote rapid feature learning in neural networks.
problem Understanding how neural networks efficiently extract features from data.
method Deriving exact solutions to a minimal model of neural networks transitioning between lazy and rich learning regimes.
result Unbalanced layer-specific initialization variances and learning rates determine the degree of feature learning.
New findings on representation changes in transfer learning.
problem Understanding representation changes in transfer learning.
method Analysis of pre-trained and fine-tuned representations in transfer learning.
result Representation changes mainly occur in the head, but early layers are also important for cross-domain adaptation.
Deep networks reveal colour opponent cells under retinal constraints.
problem Understanding colour vision in deep neural networks.
method Classifying cells in early layers of anatomically constrained convolutional neural networks.
result Emergence of single and double opponent cells in deep networks.
It is a known fact that training recurrent neural networks for tasks that have long term dependencies is challenging. One of the main reasons is the vanishing or exploding gradient problem, which prevents gradient information from propagating to early layers. In this paper we propose a simple recurrent architecture, th…
It is widely believed that the backpropagation algorithm is essential for learning good feature detectors in early layers of artificial neural networks, so that these detectors are useful for the task performed by the higher layers of that neural network. At the same time, the traditional form of backpropagation is bio…
Improved speech enhancement with larger neural networks using novel embeddings and biases.
problem Decreased robustness of speech enhancement models to real-world use cases.
method Frequency-positional embeddings, semi-supervised training, biased loss function.
result Better performance on real recordings with improved large neural network architecture.
Investigates optimal parameter allocation in Transformers for efficiency and expressivity.
problem Balancing expressivity and efficiency in Transformer model parameters.
method Mathematical analysis and theoretical characterization of attention heads and head dimensions.
result Later layers can operate more efficiently with reduced parameters due to saturation of softmax activations.
TCT improves federated learning by convexifying neural networks.
problem Federated learning performance disparity due to nonconvexity.
method Train-Convexify-Train procedure using neural tangent kernels.
result Up to +36% accuracy improvement on FMNIST and +37% on CIFAR10.
Adversarially-trained models transfer better in limited data scenarios.
problem Improving transfer learning performance with limited data.
method Adversarially-train deep nets, freeze early layers, fine-tune last layers, observe shape vs texture bias.
result Adversarially-trained models transfer better than non-adversarially-trained models, especially with limited data.
Transformers learn to generalize unseen tasks by composing self-attention layers.
problem How Transformers generalize to unseen, out-of-distribution tasks.
method Examined synthetic examples and pretrained LLMs, focusing on induction heads and latent subspace.
result Transformers can learn hidden rules for unseen tasks by composing self-attention layers, achieving OOD generalization.
We present an approach to adaptively utilize deep neural networks in order to reduce the evaluation time on new examples without loss of accuracy. Rather than attempting to redesign or approximate existing networks, we propose two schemes that adaptively utilize networks. We first pose an adaptive network evaluation sc…
Detects corruption in agentic models during execution.
problem Inconsistent context, retrieval errors, or adversarial inputs corrupt intermediate steps of reasoning chains.
method Analyzes token graphs induced by attention and computes spectral statistics to emit accept/reject signals.
result A single threshold on the high frequency energy ratio optimally detects context inconsistency in agentic models.
This study investigates how much knowledge from natural images can be transferred to pathology images.
problem Quantifying how much knowledge from natural images can be transferred to pathology images.
method Proposes a framework to quantify knowledge gain by a particular layer, conducts empirical investigation in pathology image centered transfer learning.
result Early layers of deep models can transfer knowledge to pathology image classification tasks.
Scattering networks are a class of designed Convolutional Neural Networks (CNNs) with fixed weights. We argue they can serve as generic representations for modelling images. In particular, by working in scattering space, we achieve competitive results both for supervised and unsupervised learning tasks, while making pr…
IGSD separates task-specific content channels in transformer components by comparing activation replacement with zero ablation.
problem Mechanistic interpretability of transformer components
method IGSD: paired-intervention framework for comparing activation replacement with zero ablation
result IGSD identifies an early-layer content channel in transformer components that standard importance methods underestimate.
Transformer models show distinct spectral fingerprints under voice changes.
problem Detecting architectural biases in transformer models.
method Spectral analysis of attention-induced token graphs.
result Clear architectural signatures in model fingerprints correlate with language-specific behavior.
The paper presents a novel approach to multi-output regression using probabilistic circuits.
problem Capturing correlations between multiple output dimensions in large-scale regression problems.
method Employing a mixture of single-output Gaussian process experts encoded via a probabilistic circuit.
result The method can capture correlations between output dimensions and often outperforms other approaches.
A scalable MOGP model with stochastic variational inference for many outputs.
problem Efficiently modeling data from multiple sources with many outputs.
method Stochastic variational inference for Latent Variable MOGP (LV-MOGP).
result Computational complexity per iteration is independent of the number of outputs.
Multi-output learning aims to simultaneously predict multiple outputs given an input. It is an important learning problem due to the pressing need for sophisticated decision making in real-world applications. Inspired by big data, the 4Vs characteristics of multi-output imposes a set of challenges to multi-output learn…
This paper proposes an alternative to E2E training for deep networks, reducing memory footprint.
problem High GPUs memory footprint in end-to-end training of deep networks.
method Locally supervised learning with information propagation loss to avoid information collapse.
result The proposed method achieves competitive performance with less than 40% memory footprint compared to E2E training.
Unified framework for output analysis using Monte Carlo sampling.
problem Accurately assess the quality of estimated values in predictive models.
method Unified output analysis framework through Monte Carlo sampling, leveraging fast iterative bootstrap sampling and higher-order influence functions.
result Clear advantage in building more robust confidence intervals with higher coverage probability.
Enhances robustness of MOGP regression for multiple correlated outputs.
problem Model misspecification and outliers in MOGP regression.
method Extends RCGP framework to multi-output setting.
result Provable robust MOGP with joint correlation capture.
DeepICMGP surrogate models multiple outputs efficiently.
problem Challenges in modeling dependencies between multiple outputs using traditional multi-output GPs.
method Introduces hierarchical coregionalization structures across layers in DGPs.
result Demonstrates competitive performance and active learning strategies.
Improved code translation by preserving structure with composed fine-tuning.
problem Improving code translation accuracy with unlabeled code outputs.
method Pre-trained denoiser to capture output structure, composed fine-tuning to fine-tune predictor.
result Composed fine-tuning significantly improves generalization over standard fine-tuning.
Multi-Output Dependence (MOD) learning is a generalization of standard classification problems that allows for multiple outputs that are dependent on each other. A primary issue that arises in the context of MOD learning is that for any given input pattern there can be multiple correct output patterns. This changes the…
New method learns output embeddings for structured prediction.
problem Structured prediction with output embeddings.
method Jointly learns output embedding and regression function.
result Structured predictor is a consistent estimator with smaller complexity.
Proposes GPLFR for predicting high-dimensional outputs with few data.
problem Predicting high-dimensional outputs from limited data.
method GPLFR combines Gaussian process and linear-Gaussian decoding for high-dimensional prediction.
result GPLFR outperforms existing methods in predicting high-dimensional outputs.
New model predicts multiple outputs with missing labels.
problem Missing group labels in multi-output regression.
method Weakly-supervised multi-output model using correlated Gaussian processes.
result Model excels in multi-output settings with missing labels.
RePULSe improves language model alignment by reducing undesired outputs without sacrificing overall performance.
problem Aligning language models with human preferences while minimizing undesired outputs.
method Integrates probabilistic inference into RL training to reduce undesired outputs.
result RePULSe achieves a better balance between expected reward and undesired output probability.
Sig-PCA integrates model outputs and observations to correct model biases.
problem Improving model accuracy and reliability by correcting biases and numerical approximations.
method Sig-PCA framework that combines summary statistics from model outputs with localized observations via a neural network.
result Corrects model outputs to align closely with observational data, preserving essential statistical information.
Models such as Sequence-to-Sequence and Image-to-Sequence are widely used in real world applications. While the ability of these neural architectures to produce variable-length outputs makes them extremely effective for problems like Machine Translation and Image Captioning, it also leaves them vulnerable to failures o…
Gaussian processes (GPs), or distributions over arbitrary functions in a continuous domain, can be generalized to the multi-output case: a linear model of coregionalization (LMC) is one approach. LMCs estimate and exploit correlations across the multiple outputs. While model estimation can be performed efficiently for …
Automatically identifies geometric flat outputs for robotic systems.
problem Lack of systematic and practical means to identify flat outputs for arbitrary robotic systems.
method Casts the search for a globally valid, equivariant flat output as an optimization problem using Riemannian geometry, Lie group theory, and differential forms.
result Approximate transcription of continuum formulation to a quadratic program achieves precise agreement with known closed-form flat outputs.
Wide Boosting improves GB's performance on multivariate output tasks.
problem Lack of flexibility in fitting probabilistic multi-dimensional outputs.
method Inserts matrix multiplication between GB output and loss function.
result Wide Boosting outperforms Gradient Boosting on multivariate output tasks.
Proposes SHORE model for efficient MOR with sparsity and scalability.
problem Challenges of interpretability and scalability in MOR with high-dimensional outputs.
method Incorporates sparsity requirements and a two-stage optimization framework for efficient compression.
result Theoretical and empirical validation of the proposed framework's efficiency and accuracy.
Paper proposes active learning for structured output design, improving Gaussian process model predictions.
problem Finding optimal input parameters for achieving desired structured outputs.
method Developed new acquisition functions to minimize prediction error of Gaussian process model, incorporating output correlations.
result Effectiveness demonstrated in synthetic and real data experiments, including materials informatics.
In many applications of supervised learning, multiple classification or regression outputs have to be predicted jointly. We consider several extensions of gradient boosting to address such problems. We first propose a straightforward adaptation of gradient boosting exploiting multiple output regression trees as base le…