Deep neural networks often have millions of parameters. This can hinder their deployment to low-end devices, not only due to high memory requirements but also because of increased latency at inference. We propose a novel model compression method that generates a sparse trained model without additional overhead: by allowing (i) dynamic allocation of the sparsity pattern and (ii) incorporating feedback signal to reactivate prematurely pruned weights we obtain a performant sparse model in one single training pass (retraining is not needed, but can further improve the performance). We evaluate our method on CIFAR-10 and ImageNet, and show that the obtained sparse models can reach the state-of-the-art performance of dense models. Moreover, their performance surpasses that of models generated by all previously proposed pruning schemes. Highly overparametrized deep neural networks show impressive results on machine learning tasks. However, with the increase in model size comes also the demand for memory and computer power at inference stage--two resources that are scarcely available on low-end devices. Pruning techniques have been successfully applied to remove a significant fraction of the network weights while preserving test accuracy attained by dense models. In some cases, the generalization of compressed networks has even been found to be better than with full models (Han et al., 2015; 2017; Mocanu et al., 2018). The sparsity of a network is the number of weights that are identically zero, and can be obtained by applying a sparsity mask on the weights.

Spiking neural networks are nature's solution for parallel information processing with high temporal precision at a low metabolic energy cost. To that end, biological neurons integrate inputs as an analog sum and communicate their outputs digitally as spikes, i.e., sparse binary events in time. These architectural principles can be mirrored effectively in analog neuromorphic hardware. Nevertheless, training spiking neural networks with sparse activity on hardware devices remains a major challenge. Primarily this is due to the lack of suitable training methods that take into account device-specific imperfections and operate at the level of individual spikes instead of firing rates. To tackle this issue, we developed a hardware-in-the-loop strategy to train multi-layer spiking networks using surrogate gradients on the analog BrainScales-2 chip. Specifically, we used the hardware to compute the forward pass of the network, while the backward pass was computed in software. We evaluated our approach on downscaled 16x16 versions of the MNIST and the fashion MNIST datasets in which spike latencies encoded pixel intensities. The analog neuromorphic substrate closely matched the performance of equivalently sized networks implemented in software. It is capable of processing 70 k patterns per second with a power consumption of less than 300 mW. Added activity regularization resulted in sparse network activity with about 20 spikes per input, at little to no reduction in classification performance. Thus, overall, our work demonstrates low-energy spiking network processing on an analog neuromorphic substrate and sets several new benchmarks for hardware systems in terms of classification accuracy, processing speed, and efficiency. Importantly, our work emphasizes the value of hardware-in-the-loop training and paves the way toward energy-efficient information processing on non-von-Neumann architectures.

The power machine learning models consume when making predictions can be affected by a model's architecture. This paper presents various estimates of power consumption for a range of different activation functions, a core factor in neural network model architecture design. Substantial differences in hardware performance exist between activation functions. This difference informs how power consumption in machine learning models can be reduced. The field of deep neural networks has reported strong progress in many problem areas, including natural language processing (NLP), image recognition, and game playing.

This paper introduces a new approach for multi-agent communication learning called multi-agent counterfactual communication (MACC) learning. Many real-world problems are currently tackled using multi-agent techniques. However, in many of these tasks the agents do not observe the full state of the environment but only a limited observation. This absence of knowledge about the full state makes completing the objectives significantly more complex or even impossible. The key to this problem lies in sharing observation information between agents or learning how to communicate the essential data. In this paper we present a novel multi-agent communication learning approach called MACC. It addresses the partial observability problem of the agents. MACC lets the agent learn the action policy and the communication policy simultaneously. We focus on decentralized Markov Decision Processes (Dec-MDP), where the agents have joint observability. This means that the full state of the environment can be determined using the observations of all agents. MACC uses counterfactual reasoning to train both the action and the communication policy. This allows the agents to anticipate on how other agents will react to certain messages and on how the environment will react to certain actions, allowing them to learn more effective policies. MACC uses actor-critic with a centralized critic and decentralized actors. The critic is used to calculate an advantage for both the action and communication policy. We demonstrate our method by applying it on the Simple Reference Particle environment of OpenAI and a MNIST game. Our results are compared with a communication and non-communication baseline. These experiments demonstrate that MACC is able to train agents for each of these problems with effective communication policies.

Neural Architecture Search (NAS) is a promising and rapidly evolving research area. Training a large number of neural networks requires an exceptional amount of computational power, which makes NAS unreachable for those researchers who have limited or no access to high-performance clusters and supercomputers. A few benchmarks with precomputed neural architectures performances have been recently introduced to overcome this problem and ensure more reproducible experiments. However, these benchmarks are only for the computer vision domain and, thus, are built from the image datasets and convolution-derived architectures. In this work, we step outside the computer vision domain by leveraging the language modeling task, which is the core of natural language processing (NLP). Our main contribution is as follows: we have provided search space of recurrent neural networks on the text datasets and trained 14k architectures within it; we have conducted both intrinsic and extrinsic evaluation of the trained models using datasets for semantic relatedness and language understanding evaluation; finally, we have tested several NAS algorithms to demonstrate how the precomputed results can be utilized. We believe that our results have high potential of usage for both NAS and NLP communities.

With the rising number of machine learning competitions, the world has witnessed an exciting race for the best algorithms. However, the involved data selection process may fundamentally suffer from evidence ambiguity and concept drift issues, thereby possibly leading to deleterious effects on the performance of various models. This paper proposes a new Reinforced Data Sampling (RDS) method to learn how to sample data adequately on the search for useful models and insights. We formulate the optimisation problem of model diversification $\delta{-div}$ in data sampling to maximise learning potentials and optimum allocation by injecting model diversity. This work advocates the employment of diverse base learners as value functions such as neural networks, decision trees, or logistic regressions to reinforce the selection process of data subsets with multi-modal belief. We introduce different ensemble reward mechanisms, including soft voting and stochastic choice to approximate optimal sampling policy. The evaluation conducted on four datasets evidently highlights the benefits of using RDS method over traditional sampling approaches. Our experimental results suggest that the trainable sampling for model diversification is useful for competition organisers, researchers, or even starters to pursue full potentials of various machine learning tasks such as classification and regression. The source code is available at https://github.com/probeu/RDS.

End-to-end training of neural network solvers for combinatorial problems such as the Travelling Salesman Problem is intractable and inefficient beyond a few hundreds of nodes. While state-of-the-art Machine Learning approaches perform closely to classical solvers for trivially small sizes, they are unable to generalize the learnt policy to larger instances of practical scales. Towards leveraging transfer learning to solve large-scale TSPs, this paper identifies inductive biases, model architectures and learning algorithms that promote generalization to instances larger than those seen in training. Our controlled experiments provide the first principled investigation into such zero-shot generalization, revealing that extrapolating beyond training data requires rethinking the entire neural combinatorial optimization pipeline, from network layers and learning paradigms to evaluation protocols.

Disentanglement is an enjoyable property in representation learning which increases the interpretability of generative models such as Variational Auto-Encoders (VAE), Generative Adversarial Models and their many variants. In the context of latent space models, this work presents a representation learning framework that explicitly promotes disentanglement thanks to the combination of an auto-encoder with Principal Component Analysis (PCA) in latent space. The proposed objective is the sum of an auto-encoder error term along with a PCA reconstruction error in the feature space. This has an interpretation of a Restricted Kernel Machine with an interconnection matrix on the Stiefel manifold. The construction encourages a matching between the principal directions in latent space and the directions of orthogonal variation in data space. The training algorithm involves a stochastic optimization method on the Stiefel manifold, which increases only marginally the computing time compared to an analogous VAE. Our theoretical discussion and various experiments show that the proposed model improves over many VAE variants along with special emphasis on disentanglement learning.

Retrosynthesis prediction is a fundamental problem in organic synthesis, where the task is to identify precursor molecules that can be used to synthesize a target molecule. Despite recent advancements in neural retrosynthesis algorithms, they are unable to fully recapitulate the strategies employed by chemists and do not generalize well to infrequent reaction types. In this paper, we propose a graph-based approach that capitalizes on the idea that the graph topology of precursor molecules is largely unaltered during the reaction. The model first predicts the set of graph edits transforming the target into incomplete molecules called synthons. Next, the model learns to expand synthons into complete molecules by attaching relevant leaving groups. Since the model operates at the level of molecular fragments, it avoids full generation, greatly simplifying the underlying architecture and improving its ability to generalize. The model yields $11.7\%$ absolute improvement over state-of-the-art approaches on the USPTO-50k dataset, and a $4\%$ absolute improvement on a rare reaction subset of the same dataset.

Federated learning (FL) is an emerging distributed machine learning paradigm that stands out with its inherent privacy-preserving advantages. Heterogeneity is one of the core challenges in FL, which resides in the diverse user behaviors and hardware capacity across devices who participate in the training. Heterogeneity inherently exerts a huge influence on the FL training process, e.g., causing device unavailability. However, existing FL literature usually ignores the impacts of heterogeneity. To fill in the knowledge gap, we build FLASH, the first heterogeneity-aware FL platform. Based on FLASH and a large-scale user trace from 136k real-world users, we demonstrate the usefulness of FLASH in anatomizing the impacts of heterogeneity in FL by exploring three previously unaddressed research questions: whether and how can heterogeneity affect FL performance; how to configure a heterogeneity-aware FL system; and what are heterogeneity's impacts on existing FL optimizations. It shows that heterogeneity causes nontrivial performance degradation in FL from various aspects, and even invalidates some typical FL optimizations.