skip to main content


Title: Rich Feature Construction for the Optimization-Generalization Dilemma
There often is a dilemma between ease of optimization and robust out-of-distribution (OoD) generalization. For instance, many OoD methods rely on penalty terms whose optimization is challenging. They are either too strong to optimize reliably or too weak to achieve their goals. We propose to initialize the networks with a rich representation containing a palette of potentially useful features, ready to be used by even simple models. On the one hand, a rich representation provides a good initialization for the optimizer. On the other hand, it also provides an inductive bias that helps OoD generalization. Such a representation is constructed with the Rich Feature Construction (RFC) algorithm, also called the Bonsai algorithm, which consists of a succession of training episodes. During discovery episodes, we craft a multi-objective optimization criterion and its associated datasets in a manner that prevents the network from using the features constructed in the previous iterations. During synthesis episodes, we use knowledge distillation to force the network to simultaneously represent all the previously discovered features. Initializing the networks with Bonsai representations consistently helps six OoD methods achieve top performance on ColoredMNIST benchmark. The same technique substantially outperforms comparable results on the Wilds Camelyon17 task, eliminates the high result variance that plagues other methods, and makes hyperparameter tuning and model selection more reliable.  more » « less
Award ID(s):
1922658
NSF-PAR ID:
10343272
Author(s) / Creator(s):
; ;
Date Published:
Journal Name:
International Conference on Machine Learning
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Does the dominant approach to learn representations (as a side effect of optimizing an expected cost for a single training distribution) remain a good approach when we are dealing with multiple distributions? Our thesis is that such scenarios are better served by representations that are richer than those obtained with a single optimization episode. We support this thesis with simple theoretical arguments and with experiments utilizing an apparently na\"ıve ensembling technique: concatenating the representations obtained from multiple training episodes using the same data, model, algorithm, and hyper-parameters, but different random seeds. These independently trained networks perform similarly. Yet, in a number of scenarios involving new distributions, the concatenated representation performs substantially better than an equivalently sized network trained with a single training run. This proves that the representations constructed by multiple training episodes are in fact different. Although their concatenation carries little additional information about the training task under the training distribution, it becomes substantially more informative when tasks or distributions change. Meanwhile, a single training episode is unlikely to yield such a redundant representation because the optimization process has no reason to accumulate features that do not incrementally improve the training performance. 
    more » « less
  2. Abstract

    Climate models are essential to understand and project climate change, yet long‐standing biases and uncertainties in their projections remain. This is largely associated with the representation of subgrid‐scale processes, particularly clouds and convection. Deep learning can learn these subgrid‐scale processes from computationally expensive storm‐resolving models while retaining many features at a fraction of computational cost. Yet, climate simulations with embedded neural network parameterizations are still challenging and highly depend on the deep learning solution. This is likely associated with spurious non‐physical correlations learned by the neural networks due to the complexity of the physical dynamical system. Here, we show that the combination of causality with deep learning helps removing spurious correlations and optimizing the neural network algorithm. To resolve this, we apply a causal discovery method to unveil causal drivers in the set of input predictors of atmospheric subgrid‐scale processes of a superparameterized climate model in which deep convection is explicitly resolved. The resulting causally‐informed neural networks are coupled to the climate model, hence, replacing the superparameterization and radiation scheme. We show that the climate simulations with causally‐informed neural network parameterizations retain many convection‐related properties and accurately generate the climate of the original high‐resolution climate model, while retaining similar generalization capabilities to unseen climates compared to the non‐causal approach. The combination of causal discovery and deep learning is a new and promising approach that leads to stable and more trustworthy climate simulations and paves the way toward more physically‐based causal deep learning approaches also in other scientific disciplines.

     
    more » « less
  3. null (Ed.)
    Abstract: Radio access network (RAN) in 5G is expected to satisfy the stringent delay requirements of a variety of applications. The packet scheduler plays an important role by allocating spectrum resources to user equipments (UEs) at each transmit time interval (TTI). In this paper, we show that optimal scheduling is a challenging combinatorial optimization problem, which is hard to solve within the channel coherence time with conventional optimization methods. Rule-based scheduling methods, on the other hand, are hard to adapt to the time-varying wireless channel conditions and various data request patterns of UEs. Recently, integrating artificial intelligence (AI) into wireless networks has drawn great interest from both academia and industry. In this paper, we incorporate deep reinforcement learning (DRL) into the design of cellular packet scheduling. A delay-aware cell traffic scheduling algorithm is developed to map the observed system state to scheduling decision. Due to the huge state space, a recurrent neural network (RNN) is utilized to approximate the optimal action-policy function. Different from conventional rule-based scheduling methods, the proposed scheme can learn from the interactions with the environment and adaptively choosing the best scheduling decision at each TTI. Simulation results show that the DRL-based packet scheduling can achieve the lowest average delay compared with several conventional approaches. Meanwhile, the UEs' average queue lengths can also be significantly reduced. The developed method also exhibits great potential in real-time scheduling in delay-sensitive scenarios. 
    more » « less
  4. Methods which utilize the outputs or feature representations of predictive models have emerged as promising approaches for out-of-distribution (OOD) detection of image inputs. However, these methods struggle to detect OOD inputs that share nuisance values (e.g. background) with in-distribution inputs. The detection of shared-nuisance out-of-distribution (SN-OOD) inputs is particularly relevant in real-world applications, as anomalies and in-distribution inputs tend to be captured in the same settings during deployment. In this work, we provide a possible explanation for SN-OOD detection failures and propose nuisance-aware OOD detection to address them. Nuisance-aware OOD detection substitutes a classifier trained via empirical risk minimization and cross-entropy loss with one that 1. is trained under a distribution where the nuisance-label relationship is broken and 2. yields representations that are independent of the nuisance under this distribution, both marginally and conditioned on the label. We can train a classifier to achieve these objectives using Nuisance-Randomized Distillation (NuRD), an algorithm developed for OOD generalization under spurious correlations. Output- and feature-based nuisance-aware OOD detection perform substantially better than their original counterparts, succeeding even when detection based on domain generalization algorithms fails to improve performance. 
    more » « less
  5. BACKGROUND Optical sensing devices measure the rich physical properties of an incident light beam, such as its power, polarization state, spectrum, and intensity distribution. Most conventional sensors, such as power meters, polarimeters, spectrometers, and cameras, are monofunctional and bulky. For example, classical Fourier-transform infrared spectrometers and polarimeters, which characterize the optical spectrum in the infrared and the polarization state of light, respectively, can occupy a considerable portion of an optical table. Over the past decade, the development of integrated sensing solutions by using miniaturized devices together with advanced machine-learning algorithms has accelerated rapidly, and optical sensing research has evolved into a highly interdisciplinary field that encompasses devices and materials engineering, condensed matter physics, and machine learning. To this end, future optical sensing technologies will benefit from innovations in device architecture, discoveries of new quantum materials, demonstrations of previously uncharacterized optical and optoelectronic phenomena, and rapid advances in the development of tailored machine-learning algorithms. ADVANCES Recently, a number of sensing and imaging demonstrations have emerged that differ substantially from conventional sensing schemes in the way that optical information is detected. A typical example is computational spectroscopy. In this new paradigm, a compact spectrometer first collectively captures the comprehensive spectral information of an incident light beam using multiple elements or a single element under different operational states and generates a high-dimensional photoresponse vector. An advanced algorithm then interprets the vector to achieve reconstruction of the spectrum. This scheme shifts the physical complexity of conventional grating- or interference-based spectrometers to computation. Moreover, many of the recent developments go well beyond optical spectroscopy, and we discuss them within a common framework, dubbed “geometric deep optical sensing.” The term “geometric” is intended to emphasize that in this sensing scheme, the physical properties of an unknown light beam and the corresponding photoresponses can be regarded as points in two respective high-dimensional vector spaces and that the sensing process can be considered to be a mapping from one vector space to the other. The mapping can be linear, nonlinear, or highly entangled; for the latter two cases, deep artificial neural networks represent a natural choice for the encoding and/or decoding processes, from which the term “deep” is derived. In addition to this classical geometric view, the quantum geometry of Bloch electrons in Hilbert space, such as Berry curvature and quantum metrics, is essential for the determination of the polarization-dependent photoresponses in some optical sensors. In this Review, we first present a general perspective of this sensing scheme from the viewpoint of information theory, in which the photoresponse measurement and the extraction of light properties are deemed as information-encoding and -decoding processes, respectively. We then discuss demonstrations in which a reconfigurable sensor (or an array thereof), enabled by device reconfigurability and the implementation of neural networks, can detect the power, polarization state, wavelength, and spatial features of an incident light beam. OUTLOOK As increasingly more computing resources become available, optical sensing is becoming more computational, with device reconfigurability playing a key role. On the one hand, advanced algorithms, including deep neural networks, will enable effective decoding of high-dimensional photoresponse vectors, which reduces the physical complexity of sensors. Therefore, it will be important to integrate memory cells near or within sensors to enable efficient processing and interpretation of a large amount of photoresponse data. On the other hand, analog computation based on neural networks can be performed with an array of reconfigurable devices, which enables direct multiplexing of sensing and computing functions. We anticipate that these two directions will become the engineering frontier of future deep sensing research. On the scientific frontier, exploring quantum geometric and topological properties of new quantum materials in both linear and nonlinear light-matter interactions will enrich the information-encoding pathways for deep optical sensing. In addition, deep sensing schemes will continue to benefit from the latest developments in machine learning. Future highly compact, multifunctional, reconfigurable, and intelligent sensors and imagers will find applications in medical imaging, environmental monitoring, infrared astronomy, and many other areas of our daily lives, especially in the mobile domain and the internet of things. Schematic of deep optical sensing. The n -dimensional unknown information ( w ) is encoded into an m -dimensional photoresponse vector ( x ) by a reconfigurable sensor (or an array thereof), from which w′ is reconstructed by a trained neural network ( n ′ = n and w′   ≈   w ). Alternatively, x may be directly deciphered to capture certain properties of w . Here, w , x , and w′ can be regarded as points in their respective high-dimensional vector spaces ℛ n , ℛ m , and ℛ n ′ . 
    more » « less