3D imaging enables accurate diagnosis by providing spatial information about organ anatomy. However, using 3D images to train AI models is computationally challenging because they consist of 10x or 100x more pixels than their 2D counterparts. To be trained with high-resolution 3D images, convolutional neural networks resort to downsampling them or projecting them to 2D. We propose an effective alternative, a neural network that enables efficient classification of full-resolution 3D medical images. Compared to off-the-shelf convolutional neural networks, our network, 3D Globally-Aware Multiple Instance Classifier (3D-GMIC), uses 77.98%-90.05% less GPU memory and 91.23%-96.02% less computation. While it is trained only with image-level labels, without segmentation labels, it explains its predictions by providing pixel-level saliency maps. On a dataset collected at NYU Langone Health, including 85,526 patients with full-field 2D mammography (FFDM), synthetic 2D mammography, and 3D mammography, 3D-GMIC achieves an AUC of 0.831 (95% CI: 0.769-0.887) in classifying breasts with malignant findings using 3D mammography. This is comparable to the performance of GMIC on FFDM (0.816, 95% CI: 0.737-0.878) and synthetic 2D (0.826, 95% CI: 0.754-0.884), which demonstrates that 3D-GMIC successfully classified large 3D images despite focusing computation on a smaller percentage of its input compared to GMIC. Therefore, 3D-GMIC identifies and utilizes extremely small regions of interest from 3D images consisting of hundreds of millions of pixels, dramatically reducing associated computational challenges. 3D-GMIC generalizes well to BCS-DBT, an external dataset from Duke University Hospital, achieving an AUC of 0.848 (95% CI: 0.798-0.896).

Large-scale transformer-based methods are gaining popularity as a tool for predicting the properties of chemical compounds, which is of central importance to the drug discovery process. To accelerate their development and dissemination among the community, we are releasing HuggingMolecules -- an open-source library, with a simple and unified API, that provides the implementation of several state-of-the-art transformers for molecular property prediction. In addition, we add a comparison of these methods on several regression and classification datasets. HuggingMolecules package is available at: github.com/gmum/huggingmolecules.

We hypothesize that due to the greedy nature of learning in multi-modal deep neural networks, these models tend to rely on just one modality while under-fitting the other modalities. Such behavior is counter-intuitive and hurts the models’ generalization, as we observe empirically. To estimate the model’s dependence on each modality, we compute the gain on the accuracy when the model has access to it in addition to another modality. We refer to this gain as the conditional utilization rate. In the experiments, we consistently observe an imbalance in conditional utilization rates between modalities, across multiple tasks and architectures. Since conditional utilization rate cannot be computed efficiently during training, we introduce a proxy for it based on the pace at which the model learns from each modality, which we refer to as the conditional learning speed. We propose an algorithm to balance the conditional learning speeds between modalities during training and demonstrate that it indeed addresses the issue of greedy learning. The proposed algorithm improves the model’s generalization on three datasets: Colored MNIST, ModelNet40, and NVIDIA Dynamic Hand Gesture.

The early phase of training a deep neural network has a dramatic effect on the local curvature of the loss function. For instance, using a small learning rate does not guarantee stable optimization because the optimization trajectory has a tendency to steer towards regions of the loss surface with increasing local curvature. We ask whether this tendency is connected to the widely observed phenomenon that the choice of the learning rate strongly influences generalization. We first show that stochastic gradient descent (SGD) implicitly penalizes the trace of the Fisher Information Matrix (FIM), a measure of the local curvature, from the start of training. We argue it is an implicit regularizer in SGD by showing that explicitly penalizing the trace of the FIM can significantly improve generalization. We highlight that poor final generalization coincides with the trace of the FIM attaining a large value early in training, to which we refer as catastrophic Fisher explosion. Finally, to gain insight into the regularization effect of penalizing the trace of the FIM, we show that it limits memorization by reducing the learning speed of examples with noisy labels more than that of the examples with clean labels.

The computation of the distance to the true distribution is a key component of most state-of-the-art generative models. Inspired by prior works on the Sliced-Wasserstein Auto-Encoders (SWAE) and the Wasserstein Auto-Encoders with MMD-based penalty (WAE-MMD), we propose a new generative model - a Cramer-Wold Auto-Encoder (CWAE). A fundamental component of CWAE is the characteristic kernel, the construction of which is one of the goals of this paper, from here on referred to as the Cramer-Wold kernel. Its main distinguishing feature is that it has a closed-form of the kernel product of radial Gaussians. Consequently, CWAE model has a~closed-form for the distance between the posterior and the normal prior, which simplifies the optimization procedure by removing the need to sample in order to compute the loss function. At the same time, CWAE performance often improves upon WAE-MMD and SWAE on standard benchmarks.

We present a deep convolutional neural network for breast cancer screening exam classification, trained, and evaluated on over 200000 exams (over 1000000 images). Our network achieves an AUC of 0.895 in predicting the presence of cancer in the breast, when tested on the screening population. We attribute the high accuracy to a few technical advances. 1) Our network's novel two-stage architecture and training procedure, which allows us to use a high-capacity patch-level network to learn from pixel-level labels alongside a network learning from macroscopic breast-level labels. 2) A custom ResNet-based network used as a building block of our model, whose balance of depth and width is optimized for high-resolution medical images. 3) Pretraining the network on screening BI-RADS classification, a related task with more noisy labels. 4) Combining multiple input views in an optimal way among a number of possible choices. To validate our model, we conducted a reader study with 14 readers, each reading 720 screening mammogram exams, and show that our model is as accurate as experienced radiologists when presented with the same data. We also show that a hybrid model, averaging the probability of malignancy predicted by a radiologist with a prediction of our neural network, is more accurate than either of the two separately. To further understand our results, we conduct a thorough analysis of our network's performance on different subpopulations of the screening population, the model's design, training procedure, errors, and properties of its internal representations. Our best models are publicly available at https://github.com/nyukat/breast_cancer_classifier.

We demonstrate that in residual neural networks (ResNets) dynamical isometry is achievable irrespective of the activation function used. We do that by deriving, with the help of Free Probability and Random Matrix Theories, a universal formula for the spectral density of the input-output Jacobian at initialization, in the large network width and depth limit. The resulting singular value spectrum depends on a single parameter, which we calculate for a variety of popular activation functions, by analyzing the signal propagation in the artificial neural network. We corroborate our results with numerical simulations of both random matrices and ResNets applied to the CIFAR-10 classification problem. Moreover, we study consequences of this universal behavior for the initial and late phases of the learning processes. We conclude by drawing attention to the simple fact, that initialization acts as a confounding factor between the choice of activation function and the rate of learning. We propose that in ResNets this can be resolved based on our results by ensuring the same level of dynamical isometry at initialization.

Fine-tuning large pretrained models is an effective transfer mechanism in NLP. However, in the presence of many downstream tasks, fine-tuning is parameter inefficient: an entire new model is required for every task. As an alternative, we propose transfer with adapter modules. Adapter modules yield a compact and extensible model; they add only a few trainable parameters per task, and new tasks can be added without revisiting previous ones. The parameters of the original network remain fixed, yielding a high degree of parameter sharing. To demonstrate adapter’s effectiveness, we transfer the recently proposed BERT Transformer model to $26$ diverse text classification tasks, including the GLUE benchmark. Adapters attain near state-of-the-art performance, whilst adding only a few parameters per task. On GLUE, we attain within $0.8%$ of the performance of full fine-tuning, adding only $3.6%$ parameters per task. By contrast, fine-tuning trains $100%$ of the parameters per task.

There are many surprising and perhaps counter-intuitive properties of optimization of deep neural networks. We propose and experimentally verify a unified phenomenological model of the loss landscape that incorporates many of them. High dimensionality plays a key role in our model. Our core idea is to model the loss landscape as a set of high dimensional \emph{wedges} that together form a large-scale, inter-connected structure and towards which optimization is drawn. We first show that hyperparameter choices such as learning rate, network width and $L_2$ regularization, affect the path optimizer takes through the landscape in similar ways, influencing the large scale curvature of the regions the optimizer explores. Finally, we predict and demonstrate new counter-intuitive properties of the loss-landscape. We show an existence of low loss subspaces connecting a set (not only a pair) of solutions, and verify it experimentally. Finally, we analyze recently popular ensembling techniques for deep networks in the light of our model.

We examine the role of memorization in deep learning, drawing connections to capacity, generalization, and adversarial robustness. While deep networks are capable of memorizing noise data, our results suggest that they tend to prioritize learning simple patterns first. In our experiments, we expose qualitative differences in gradient-based optimization of deep neural networks (DNNs) on noise vs.~real data. We also demonstrate that for appropriately tuned explicit regularization (e.g.,~dropout) we can degrade DNN training performance on noise datasets without compromising generalization on real data. Our analysis suggests that the notions of effective capacity which are dataset independent are unlikely to explain the generalization performance of deep networks when trained with gradient based methods because training data itself plays an important role in determining the degree of memorization.