While the need for interpretable machine learning has been established, many common approaches are slow, lack fidelity, or hard to evaluate. Amortized explanation methods reduce the cost of providing interpretations by learning a global selector model that returns feature importances for a single instance of data. The selector model is trained to optimize the fidelity of the interpretations, as evaluated by a predictor model for the target. Popular methods learn the selector and predictor model in concert, which we show allows predictions to be encoded within interpretations. We introduce EVAL-X as a method to quantitatively evaluate interpretations and REAL-X as an amortized explanation method, which learn a predictor model that approximates the true data generating distribution given any subset of the input. We show EVAL-X can detect when predictions are encoded in interpretations and show the advantages of REAL-X through quantitative and radiologist evaluation.

Advances in deep learning have led to promising progress in inferring graphics programs by de-rendering computer-generated images. However, current methods do not explore which decoding methods lead to better inductive bias for inferring graphics programs. In our work, we first explore the effectiveness of LSTM-RNN versus Transformer networks as decoders for order-independent graphics programs. Since these are sequence models, we must choose an ordering of the objects in the graphics programs for likelihood training. We found that the LSTM performance was highly sensitive to the sequence ordering (random order vs. pattern-based order), while Transformer performance was roughly independent of the sequence ordering. Further, we present a policy gradient based reinforcement learning approach for better inductive bias in the decoder via multiple diverse rewards based both on the graphics program specification and the rendered image. We also explore the combination of these complementary rewards. We achieve state-of-the-art results on two graphics program generation datasets.

Early detection of skin cancers like melanoma is crucial to ensure high chances of survival for patients. Clinical application of Deep Learning (DL)-based Decision Support Systems (DSS) for skin cancer screening has the potential to improve the quality of patient care. The majority of work in the medical AI community focuses on a diagnosis setting that is mainly relevant for autonomous operation. Practical decision support should, however, go beyond plain diagnosis and provide explanations. This paper provides an overview of works towards explainable, DL-based decision support in medical applications with the example of skin cancer diagnosis from clinical, dermoscopic and histopathologic images. Analysis reveals that comparably little attention is payed to the explanation of histopathologic skin images and that current work is dominated by visual relevance maps as well as dermoscopic feature identification. We conclude that future work should focus on meeting the stakeholder's cognitive concepts, providing exhaustive explanations that combine global and local approaches and leverage diverse modalities. Moreover, the possibility to intervene and guide models in case of misbehaviour is identified as a major step towards successful deployment of AI as DL-based DSS and beyond.

Over the last decade, artificial intelligence has found many applications areas in the society. As AI solutions have become more sophistication and the use cases grew, they highlighted the need to address performance and energy efficiency challenges faced during the implementation process. To address these challenges, there has been growing interest in neuromorphic chips. Neuromorphic computing relies on non von Neumann architectures as well as novel devices, circuits and manufacturing technologies to mimic the human brain. Among such technologies, 3D integration is an important enabler for AI hardware and the continuation of the scaling laws. In this paper, we overview the unique opportunities 3D integration provides in neuromorphic chip design, discuss the emerging opportunities in next generation neuromorphic architectures and review the obstacles. Neuromorphic architectures, which relied on the brain for inspiration and emulation purposes, face grand challenges due to the limited understanding of the functionality and the architecture of the human brain. Yet, high-levels of investments are dedicated to develop neuromorphic chips. We argue that 3D integration not only provides strategic advantages to the cost-effective and flexible design of neuromorphic chips, it may provide design flexibility in incorporating advanced capabilities to further benefits the designs in the future.

Recent empirical results show that deep networks can approximate solutions to high dimensional PDEs, seemingly escaping the curse of dimensionality. However many open questions remain regarding the theoretical basis for such approximations, including the number of parameters required. In this paper, we investigate the representational power of neural networks for approximating solutions to linear elliptic PDEs with Dirichlet Boundary conditions. We prove that when a PDE's coefficients are representable by small neural networks, the parameters required to approximate its solution scale polynomially with the input dimension $d$ and are proportional to the parameter counts of the coefficient neural networks. Our proof is based on constructing a neural network which simulates gradient descent in an appropriate Hilbert space which converges to the solution of the PDE. Moreover, we bound the size of the neural network needed to represent each iterate in terms of the neural network representing the previous iterate, resulting in a final network whose parameters depend polynomially on $d$ and does not depend on the volume of the domain.

Time dependent data is a main source of information in today's data driven world. Generating this type of data though has shown its challenges and made it an interesting research area in the field of generative machine learning. One such approach was that by Smith et al. who developed Time Series Generative Adversarial Network (TSGAN) which showed promising performance in generating time dependent data and the ability of few shot generation though being flawed in certain aspects of training and learning. This paper looks to improve on the results from TSGAN and address those flaws by unifying the training of the independent networks in TSGAN and creating a dependency both in training and learning. This improvement, called unified TSGAN (uTSGAN) was tested and comapred both quantitatively and qualitatively to its predecessor on 70 benchmark time series data sets used in the community. uTSGAN showed to outperform TSGAN in 80\% of the data sets by the same number of training epochs and 60\% of the data sets in 3/4th the amount of training time or less while maintaining the few shot generation ability with better FID scores across those data sets.

We consider the problem of finding a two-layer neural network with sigmoid, rectified linear unit (ReLU), or binary step activation functions that "fits" a training data set as accurately as possible as quantified by the training error; and study the following question: \emph{does a low training error guarantee that the norm of the output layer (outer norm) itself is small?} We answer affirmatively this question for the case of non-negative output weights. Using a simple covering number argument, we establish that under quite mild distributional assumptions on the input/label pairs; any such network achieving a small training error on polynomially many data necessarily has a well-controlled outer norm. Notably, our results (a) have a polynomial (in $d$) sample complexity, (b) are independent of the number of hidden units (which can potentially be very high), (c) are oblivious to the training algorithm; and (d) require quite mild assumptions on the data (in particular the input vector $X\in\mathbb{R}^d$ need not have independent coordinates). We then leverage our bounds to establish generalization guarantees for such networks through \emph{fat-shattering dimension}, a scale-sensitive measure of the complexity class that the network architectures we investigate belong to. Notably, our generalization bounds also have good sample complexity (polynomials in $d$ with a low degree), and are in fact near-linear for some important cases of interest.

In this work we investigate different machine learning based strategies for denoising raw simulation data from ProtoDUNE experiment. Proto-DUNE detector is hosted by CERN and it aims to test and calibrate the technologies for DUNE, a forthcoming experiment in neutrino physics. Our models leverage deep learning algorithms to make the first step in the reconstruction workchain, which consists in converting digital detector signals into physical high level quantities. We benchmark this approach against traditional algorithms implemented by the DUNE collaboration. We test the capabilities of graph neural networks, while exploiting multi-GPU setups to accelerate training and inference processes.

Batch Normalization (BN) is a commonly used technique to accelerate and stabilize training of deep neural networks. Despite its empirical success, a full theoretical understanding of BN is yet to be developed. In this work, we analyze BN through the lens of convex optimization. We introduce an analytic framework based on convex duality to obtain exact convex representations of weight-decay regularized ReLU networks with BN, which can be trained in polynomial-time. Our analyses also show that optimal layer weights can be obtained as simple closed-form formulas in the high-dimensional and/or overparameterized regimes. Furthermore, we find that Gradient Descent provides an algorithmic bias effect on the standard non-convex BN network, and we design an approach to explicitly encode this implicit regularization into the convex objective. Experiments with CIFAR image classification highlight the effectiveness of this explicit regularization for mimicking and substantially improving the performance of standard BN networks.

Deep learning techniques have achieved remarkable performance in wide-ranging tasks. However, when trained on privacy-sensitive datasets, the model parameters may expose private information in training data. Prior attempts for differentially private training, although offering rigorous privacy guarantees, lead to much lower model performance than the non-private ones. Besides, different runs of the same training algorithm produce models with large performance variance. To address these issues, we propose DPlis--Differentially Private Learning wIth Smoothing. The core idea of DPlis is to construct a smooth loss function that favors noise-resilient models lying in large flat regions of the loss landscape. We provide theoretical justification for the utility improvements of DPlis. Extensive experiments also demonstrate that DPlis can effectively boost model quality and training stability under a given privacy budget.