Several recent works demonstrate that transformers can implement algorithms like gradient descent. By a careful construction of weights, these works show that multiple layers of transformers are expressive enough to simulate iterations of gradient descent. Going beyond the question of expressivity, we ask: Can transformers learn to implement such algorithms by training over random problem instances? To our knowledge, we make the first theoretical progress on this question via an analysis of the loss landscape for linear transformers trained over random instances of linear regression. For a single attention layer, we prove the global minimum of the training objective implements a single iteration of preconditioned gradient descent. Notably, the preconditioning matrix not only adapts to the input distribution but also to the variance induced by data inadequacy. For a transformer with $L$ attention layers, we prove certain critical points of the training objective implement $L$ iterations of preconditioned gradient descent. Our results call for future theoretical studies on learning algorithms by training transformers.

Current state-of-the-art object-centric models use slots and attention-based routing for binding. However, this class of models has several conceptual limitations: the number of slots is hardwired; all slots have equal capacity; training has high computational cost; there are no object-level relational factors within slots. Synchrony-based models in principle can address these limitations by using complex-valued activations which store binding information in their phase components. However, working examples of such synchrony-based models have been developed only very recently, and are still limited to toy grayscale datasets and simultaneous storage of less than three objects in practice. Here we introduce architectural modifications and a novel contrastive learning method that greatly improve the state-of-the-art synchrony-based model. For the first time, we obtain a class of synchrony-based models capable of discovering objects in an unsupervised manner in multi-object color datasets and simultaneously representing more than three objects.

Feature selection (FS) plays an important role in machine learning, which extracts important features and accelerates the learning process. In this paper, we propose a deep FS method that simultaneously conducts feature selection and differentiable $ k $-NN graph learning based on the Dirichlet Energy. The Dirichlet Energy identifies important features by measuring their smoothness on the graph structure, and facilitates the learning of a new graph that reflects the inherent structure in new feature subspace. We employ Optimal Transport theory to address the non-differentiability issue of learning $ k $-NN graphs in neural networks, which theoretically makes our method applicable to other graph neural networks for dynamic graph learning. Furthermore, the proposed framework is interpretable, since all modules are designed algorithmically. We validate the effectiveness of our model with extensive experiments on both synthetic and real-world datasets.

We present a general methodology that learns to classify images without labels by leveraging pretrained feature extractors. Our approach involves self-distillation training of clustering heads based on the fact that nearest neighbours in the pretrained feature space are likely to share the same label. We propose a novel objective that learns associations between image features by introducing a variant of pointwise mutual information together with instance weighting. We demonstrate that the proposed objective is able to attenuate the effect of false positive pairs while efficiently exploiting the structure in the pretrained feature space. As a result, we improve the clustering accuracy over $k$-means on $17$ different pretrained models by $6.1$\% and $12.2$\% on ImageNet and CIFAR100, respectively. Finally, using self-supervised vision transformers, we achieve a clustering accuracy of $61.6$\% on ImageNet. The code is available at https://github.com/HHU-MMBS/TEMI-official-BMVC2023.

We study the problem of progressive ensemble distillation: Given a large, pretrained teacher model $g$, we seek to decompose the model into smaller, low-inference cost student models $f_i$, such that progressively evaluating additional models in this ensemble leads to improved predictions. The resulting ensemble allows for flexibly tuning accuracy vs. inference cost at runtime, which is useful for a number of applications in on-device inference. The method we propose, B-DISTIL , relies on an algorithmic procedure that uses function composition over intermediate activations to construct expressive ensembles with similar performance as $g$ , but with smaller student models. We demonstrate the effectiveness of B-DISTIL by decomposing pretrained models across standard image, speech, and sensor datasets. We also provide theoretical guarantees in terms of convergence and generalization.

We consider the straggler problem in decentralized learning over a logical ring while preserving user data privacy. Especially, we extend the recently proposed framework of differential privacy (DP) amplification by decentralization by Cyffers and Bellet to include overall training latency--comprising both computation and communication latency. Analytical results on both the convergence speed and the DP level are derived for both a skipping scheme (which ignores the stragglers after a timeout) and a baseline scheme that waits for each node to finish before the training continues. A trade-off between overall training latency, accuracy, and privacy, parameterized by the timeout of the skipping scheme, is identified and empirically validated for logistic regression on a real-world dataset and for image classification using the MNIST and CIFAR-10 datasets.

Imitation learning (IL), which learns control policies by imitating expert demonstrations, has demonstrated success across a variety of domains including self-driving cars (Codevilla et al., 2018) and robotics (Schaal, 1999). However, using IL to learn a robust behavior policy may require a large amount of training data (Ross et al., 2011), and expert demonstrations are often expensive to collect. One remedy for this problem is multi-task learning: using data from other tasks (source tasks) in addition to from the task of interest (target task) to jointly learn a policy. We study the application of multi-task learning to IL over linear systems, and demonstrate improved sample efficiency when learning a controller via representation learning. Our results expand on prior work that studies multi-task representation learning for supervised learning (Du et al., 2020; Tripuraneni et al., 2021), addressing the new challenges that arise in the imitation learning setting. First, the data for IL is temporally dependent, as it is generated from a dynamical system x [ t + 1] = f ( x[ t],u [t ],w [t ]). In contrast, the supervised learning setting assumes that both the train and test data are independent and identically distributed (i.i.d.) from the same underlying distribution. Furthermore, we are interested in the performance of the learned controller in closed-loop rather than its error on expert-controlled trajectories.

We aim to demonstrate in experiments that our cost sensitive PEGASOS SVM achieves good performance on imbalanced data sets with a Majority to Minority Ratio ranging from 8.6:1 to 130:1 and to ascertain whether the including intercept (bias), regularization and parameters affects performance on our selection of datasets. Although many resort to SMOTE methods, we aim for a less computationally intensive method. We evaluate the performance by examining the learning curves. These curves diagnose whether we overfit or underfit or whether the random sample of data chosen during the process was not random enough or diverse enough in dependent variable class for the algorithm to generalized to unseen examples. We will also see the background of the hyperparameters versus the test and train error in validation curves. We benchmark our PEGASOS Cost-Sensitive SVM's results of Ding's LINEAR SVM DECIDL method. He obtained an ROC-AUC of .5 in one dataset. Our work will extend the work of Ding by incorporating kernels into SVM. We will use Python rather than MATLAB as python has dictionaries for storing mixed data types during multi-parameter cross-validation.

Both humans and machines learn the meaning of unknown words through contextual information in a sentence, but not all contexts are equally helpful for learning. We introduce an effective method for capturing the level of contextual informativeness with respect to a given target word. Our study makes three main contributions. First, we develop models for estimating contextual informativeness, focusing on the instructional aspect of sentences. Our attention-based approach using pre-trained embeddings demonstrates state-of-the-art performance on our single-context dataset and an existing multi-sentence context dataset. Second, we show how our model identifies key contextual elements in a sentence that are likely to contribute most to a reader's understanding of the target word. Third, we examine how our contextual informativeness model, originally developed for vocabulary learning applications for students, can be used for developing better training curricula for word embedding models in batch learning and few-shot machine learning settings. We believe our results open new possibilities for applications that support language learning for both human and machine learners.

Federated Learning is expected to provide strong privacy guarantees, as only gradients or model parameters but no plain text training data is ever exchanged either between the clients or between the clients and the central server. In this paper, we challenge this claim by introducing a simple but still very effective membership inference attack algorithm, which relies only on a single training step. In contrast to the popular honest-but-curious model, we investigate a framework with a dishonest central server. Our strategy is applicable to models with ReLU activations and uses the properties of this activation function to achieve perfect accuracy. Empirical evaluation on visual classification tasks with MNIST, CIFAR10, CIFAR100 and CelebA datasets show that our method provides perfect accuracy in identifying one sample in a training set with thousands of samples. Occasional failures of our method lead us to discover duplicate images in the CIFAR100 and CelebA datasets.