Transfer learning aims to exploit pre-trained models for more efficient follow-up training on wide range of downstream tasks and datasets, enabling successful training also on small data. Recent line of work posits strong benefits for model generalization and transfer when model size, data size, and compute budget are increased for the pre-training. It remains however still largely unclear whether the observed transfer improvement due to increase in scale also holds when source and target data distributions are far apart from each other. In this work we conduct large-scale pre-training on large source datasets of either natural (ImageNet-21k/1k) or medical chest X-Ray images and compare full and few-shot transfer using different target datasets from both natural and medical imaging domains. Our observations provide evidence that while pre-training and transfer on closely related datasets do show clear benefit of increasing model and data size during pre-training, such benefits are not clearly visible when source and target datasets are further apart. These observations hold across both full and few-shot transfer and indicate that scaling laws pointing to improvement of generalization and transfer with increasing model and data size are incomplete and should be revised by taking into account the type and proximity of the source and target data, to correctly predict the effect of model and data scale during pre-training on transfer. Remarkably, in full shot transfer to a large X-Ray chest imaging target (PadChest), the largest model pre-trained on ImageNet-21k slightly outperforms best models pre-trained on large X-Ray chest imaging data. This indicates possibility to obtain high quality models for domain-specific transfer even without access to large domain-specific data, by pre-training instead on comparably very large, generic source data.

This paper explores the idea of utilising Long Short-Term Memory neural networks (LSTMNN) for the generation of musical sequences in ABC notation. The proposed approach takes ABC notations from the Nottingham dataset and encodes it to be fed as input for the neural networks. The primary objective is to input the neural networks with an arbitrary note, let the network process and augment a sequence based on the note until a good piece of music is produced. Multiple calibrations have been done to amend the parameters of the network for optimal generation. The output is assessed on the basis of rhythm, harmony, and grammar accuracy.

Despite its wide use, recent studies have revealed unexpected and undesirable properties of neural autoregressive sequence models trained with maximum likelihood, such as an unreasonably high affinity to short sequences after training and to infinitely long sequences at decoding time. We propose to study these phenomena by investigating how the modes, or local maxima, of a distribution are maintained throughout the full learning chain of the ground-truth, empirical, learned and decoding-induced distributions, via the newly proposed mode recovery cost. We design a tractable testbed where we build three types of ground-truth distributions: (1) an LSTM based structured distribution, (2) an unstructured distribution where probability of a sequence does not depend on its content, and (3) a product of these two which we call a semi-structured distribution. Our study reveals both expected and unexpected findings. First, starting with data collection, mode recovery cost strongly relies on the ground-truth distribution and is most costly with the semi-structured distribution. Second, after learning, mode recovery cost from the ground-truth distribution may increase or decrease compared to data collection, with the largest cost degradation occurring with the semi-structured ground-truth distribution. Finally, the ability of the decoding-induced distribution to recover modes from the learned distribution is highly impacted by the choices made earlier in the learning chain. We conclude that future research must consider the entire learning chain in order to fully understand the potentials and perils and to further improve neural autoregressive sequence models.

Normalizing flows are generative models that provide tractable density estimation by transforming a simple base distribution into a complex target distribution. However, this technique cannot directly model data supported on an unknown low-dimensional manifold, a common occurrence in real-world domains such as image data. Recent attempts to remedy this limitation have introduced geometric complications that defeat a central benefit of normalizing flows: exact density estimation. We recover this benefit with Conformal Embedding Flows, a framework for designing flows that learn manifolds with tractable densities. We argue that composing a standard flow with a trainable conformal embedding is the most natural way to model manifold-supported data. To this end, we present a series of conformal building blocks and apply them in experiments with real-world and synthetic data to demonstrate that flows can model manifold-supported distributions without sacrificing tractable likelihoods.

Work in deep clustering focuses on finding a single partition of data. However, high-dimensional data, such as images, typically feature multiple interesting characteristics one could cluster over. For example, images of objects against a background could be clustered over the shape of the object and separately by the colour of the background. In this paper, we introduce Multi-Facet Clustering Variational Autoencoders (MFCVAE), a novel class of variational autoencoders with a hierarchy of latent variables, each with a Mixture-of-Gaussians prior, that learns multiple clusterings simultaneously, and is trained fully unsupervised and end-to-end. MFCVAE uses a progressively-trained ladder architecture which leads to highly stable performance. We provide novel theoretical results for optimising the ELBO analytically with respect to the categorical variational posterior distribution, and corrects earlier influential theoretical work. On image benchmarks, we demonstrate that our approach separates out and clusters over different aspects of the data in a disentangled manner. We also show other advantages of our model: the compositionality of its latent space and that it provides controlled generation of samples.

Domain shifts in the training data are common in practical applications of machine learning, they occur for instance when the data is coming from different sources. Ideally, a ML model should work well independently of these shifts, for example, by learning a domain-invariant representation. Moreover, privacy concerns regarding the source also require a domain-invariant representation. In this work, we provide theoretical results that link domain invariant representations -- measured by the Wasserstein distance on the joint distributions -- to a practical semi-supervised learning objective based on a cross-entropy classifier and a novel domain critic. Quantitative experiments demonstrate that the proposed approach is indeed able to practically learn such an invariant representation (between two domains), and the latter also supports models with higher predictive accuracy on both domains, comparing favorably to existing techniques.

Model discovery aims at autonomously discovering differential equations underlying a dataset. Approaches based on Physics Informed Neural Networks (PINNs) have shown great promise, but a fully-differentiable model which explicitly learns the equation has remained elusive. In this paper we propose such an approach by combining neural network based surrogates with Sparse Bayesian Learning (SBL). We start by reinterpreting PINNs as multitask models, applying multitask learning using uncertainty, and show that this leads to a natural framework for including Bayesian regression techniques. We then construct a robust model discovery algorithm by using SBL, which we showcase on various datasets. Concurrently, the multitask approach allows the use of probabilistic approximators, and we show a proof of concept using normalizing flows to directly learn a density model from single particle data. Our work expands PINNs to various types of neural network architectures, and connects neural network-based surrogates to the rich field of Bayesian parameter inference.

The Mixture-of-experts (MoE) architecture is showing promising results in multi-task learning (MTL) and in scaling high-capacity neural networks. State-of-the-art MoE models use a trainable sparse gate to select a subset of the experts for each input example. While conceptually appealing, existing sparse gates, such as Top-k, are not smooth. The lack of smoothness can lead to convergence and statistical performance issues when training with gradient-based methods. In this paper, we develop DSelect-k: the first, continuously differentiable and sparse gate for MoE, based on a novel binary encoding formulation. Our gate can be trained using first-order methods, such as stochastic gradient descent, and offers explicit control over the number of experts to select. We demonstrate the effectiveness of DSelect-k in the context of MTL, on both synthetic and real datasets with up to 128 tasks. Our experiments indicate that MoE models based on DSelect-k can achieve statistically significant improvements in predictive and expert selection performance. Notably, on a real-world large-scale recommender system, DSelect-k achieves over 22% average improvement in predictive performance compared to the Top-k gate. We provide an open-source TensorFlow implementation of our gate.

The specialists of the Academy of Artificial Intelligence in Beijing (China) this week presented the most sophisticated natural language processing model in the world, which uses 1.75 trillion parameters to simulate conversational speech, write poems, understand images and even generate recipes, pick up the South China Morning Post newspaper. El WuDao 2.0, which in Chinese means'understanding of natural laws', is a previously trained artificial intelligence model that was developed with the help of more than 100 scientists. It is more powerful than the models of its main competitors: the GPT-3 from the company OpenAI (co-founded by Elon Musk), which was launched with 175,000 million parameters, and the Switch Transformer from Google, which uses 1.6 trillion parameters. The model develops both in Chinese and English acquired skills as you have'studied' 4.9 terabytes of images and texts, including 1.2 terabytes of text in those two languages. WuDao 2.0 already has 22 partners, such as smartphone maker Xiaomi or short video giant Kuaishou.

Deep learning is a booming field at the current time, most of the projects and problem statement uses deep learning in any sort of work. If you have to pick a deep learning technique for solving any computer vision problem statement then many of you including myself will go with a conventional neural network. In this article, we will build our first image processing project using CNN and understand its power and why it has become so popular. In this article, we will walk through every step of developing our own convolutional model and build our first amazing project. Image classification is a task where the system takes an input image and classifies it with an appropriate label.