70 research outputs found
Continual learning of conjugated visual representations through higher-order motion flows
No full textLearning with neural networks from a continuous stream of visual information presents several challenges due to the non-i.i.d. nature of the data. However, it also offers novel opportunities to develop representations that are consistent with the information flow. In this paper we investigate the case of unsupervised continual learning of pixel-wise features subject to multiple motion-induced constraints, therefore named motion-conjugated feature representations. Differently from existing approaches, motion is not a given signal (either ground-truth or estimated by external modules), but is the outcome of a progressive and autonomous learning process, occurring at various levels of the feature hierarchy. Multiple motion flows are estimated with neural networks and characterized by different levels of abstractions, spanning from traditional optical flow to other latent signals originating from higher-level features, hence called higher-order motions. Continuously learning to develop consistent multi-order flows and representations is prone to trivial solutions, which we counteract by introducing a self-supervised contrastive loss, spatially-aware and based on flow-induced similarity. We assess our model on photorealistic synthetic streams and real-world videos, comparing to pre-trained state-of-the art feature extractors (also based on Transformers) and to recent unsupervised learning models, significantly outperforming these alternatives
Being Friends Instead of Adversaries: Deep Networks Learn from Data Simplified by Other Networks
No full textAmongst a variety of approaches aimed at making the learning procedure of neural networks more effective, the scientifc community developed strategies to order the examples
according to their estimated complexity, to distil knowledge
from larger networks, or to exploit the principles behind adversarial machine learning. A different idea has been recently
proposed, named Friendly Training, which consists in altering the input data by adding an automatically estimated perturbation, with the goal of facilitating the learning process
of a neural classifer. The transformation progressively fadesout as long as training proceeds, until it completely vanishes.
In this work we revisit and extend this idea, introducing a
radically different and novel approach inspired by the effectiveness of neural generators in the context of Adversarial
Machine Learning. We propose an auxiliary multi-layer network that is responsible of altering the input data to make
them easier to be handled by the classifer at the current stage
of the training procedure. The auxiliary network is trained
jointly with the neural classifer, thus intrinsically increasing
the “depth” of the classifer, and it is expected to spot general regularities in the data alteration process. The effect of
the auxiliary network is progressively reduced up to the end
of training, when it is fully dropped and the classifer is deployed for applications. We refer to this approach as Neural Friendly Training. An extended experimental procedure
involving several datasets and different neural architectures
shows that Neural Friendly Training overcomes the originally
proposed Friendly Training technique, improving the generalization of the classifer, especially in the case of noisy data
Summertime Primary and Secondary Contributions to Southern Ocean Cloud Condensation Nuclei
No full textAtmospheric aerosols in clean remote oceanic regions contribute significantly to the global albedo through the formation of haze and cloud layers; however, the relative importance of ‘primary’ wind-produced sea-spray over secondary (gas-to-particle conversion) sulphate in forming marine clouds remains unclear. Here we report on marine aerosols (PM1) over the Southern Ocean around Antarctica, in terms of their physical, chemical, and cloud droplet activation properties. Two predominant pristine air masses and aerosol populations were encountered: modified continental Antarctic (cAA) comprising predominantly sulphate with minimal sea-salt contribution and maritime Polar (mP) comprising sulphate plus sea-salt. We estimate that in cAA air, 75% of the CCN are activated into cloud droplets while in mP air, 37% are activated into droplets, for corresponding peak supersaturation ranges of 0.37–0.45% and 0.19–0.31%, respectively. When realistic marine boundary layer cloud supersaturations are considered (e.g. ~0.2–0.3%), sea-salt CCN contributed 2–13% of the activated nuclei in the cAA air and 8–51% for the marine air for surface-level wind speed < 16 m s−1. At higher wind speeds, primary marine aerosol can even contribute up to 100% of the activated CCN, for corresponding peak supersaturations as high as 0.32%. © 2018, The Author(s)
Continual Learning with Pretrained Backbones by Tuning in the Input Space
No full textThe intrinsic difficulty in adapting deep learning models to non-stationary environments limits the applicability of neural networks to real-world tasks. This issue is critical in practical supervised learning settings, such as the ones in which a pre-trained model computes projections toward a latent space where different task predictors are sequentially learned over time. As a matter of fact, incrementally fine-tuning the whole model to better adapt to new tasks usually results in catastrophic forgetting, with decreasing performance over the past experiences and losing valuable knowledge from the pretraining stage. In this paper, we propose a novel strategy to make the fine-tuning procedure more effective, by avoiding to update the pre-trained part of the network and learning not only the usual classification head, but also a set of newly-introduced learnable parameters that are responsible for transforming the input data. This process allows the network to effectively leverage the pre-training knowledge and find a good trade-off between plasticity and stability with modest computational efforts, thus especially suitable for on-the-edge settings. Our experiments on four image classification problems in a continual learning setting confirm the quality of the proposed approach when compared to several fine-tuning procedures and to popular continual learning methods
Linking mixing processes and climate variability to the heat content distribution of the Eastern Mediterranean abyss
No full textThe heat contained in the ocean (OHC) dominates the Earth’s energy budget and hence represents a fundamental parameter for understanding climate changes. However, paucity of observational data hampers our knowledge on OHC variability, particularly in abyssal areas. Here, we analyze water characteristics, observed during the last three decades in the abyssal Ionian Sea (Eastern Mediterranean), where two competing convective sources of bottom water exist. We find a heat storage of ~1.6 W/m2 – twice that assessed globally in the same period – exceptionally well-spread throughout the local abyssal layers. Such an OHC accumulation stems from progressive warming and salinification of the Eastern Mediterranean, producing warmer near-bottom waters. We analyze a new process that involves convectively-generated waters reaching the abyss as well as the triggering of a diapycnal mixing due to rough bathymetry, which brings to a warming and thickening of the bottom layer, also influencing water-column potential vorticity. This may affect the prevailing circulation, altering the local cyclonic/anticyclonic long-term variability and hence precondition future water-masses formation and the redistribution of heat along the entire water-column. © 2018, The Author(s)
Physical forcing and physical/biochemical variability of the Mediterranean Sea: A review of unresolved issues and directions for future research
No full textThis paper is the outcome of a workshop held in Rome in November 2011 on the occasion of the 25th anniversary of the POEM (Physical Oceanography of the Eastern Mediterranean) program. In the workshop discussions, a number of unresolved issues were identified for the physical and biogeochemical properties of the Mediterranean Sea as a whole, i.e., comprising the Western and Eastern sub-basins. Over the successive two years, the related ideas were discussed among the group of scientists who participated in the workshop and who have contributed to the writing of this paper. Three major topics were identified, each of them being the object of a section divided into a number of different sub-sections, each addressing a specific physical, chemical or biological issue: 1. Assessment of basin-wide physical/biochemical properties, of their variability and interactions. 2. Relative importance of external forcing functions (wind stress, heat/moisture fluxes, forcing through straits) vs. internal variability. 3. Shelf/deep sea interactions and exchanges of physical/biogeochemical properties and how they affect the sub-basin circulation and property distribution. Furthermore, a number of unresolved scientific/methodological issues were also identified and are reported in each sub-section after a short discussion of the present knowledge. They represent the collegial consensus of the scientists contributing to the paper. Naturally, the unresolved issues presented here constitute the choice of the authors and therefore they may not be exhaustive and/or complete. The overall goal is to stimulate a broader interdisciplinary discussion among the scientists of the Mediterranean oceanographic community, leading to enhanced collaborative efforts and exciting future discoveries. © Author(s) 2014
Continual Neural Computation
No full textContinuously processing a stream of not-i.i.d. data by neural models with the goal of progressively learning new skills is largely known to introduce significant challenges, frequently leading to catastrophic forgetting. In this paper we tackle this problem focusing on the low-level aspects of the neural computation model, differently from the most common existing approaches. We propose a novel neuron model, referred to as Continual Neural Unit (CNU), which does not only compute a response to an input pattern, but also diversifies computations to preserve what was previously learned, while being plastic enough to adapt to new knowledge. The values attached to weights are the outcome of a computational process which depends on the neuron input, implemented by a key-value map to select and blend multiple sets of learnable memory units. This computational mechanism implements a natural, learnable form of soft parameter isolation, virtually defining multiple computational paths within each neural unit. We show that such a computational scheme is related to the ones of popular models which perform computations relying on a set of samples stored in a memory buffer, including Kernel Machines and Transformers. Experiments in class-and-domain incremental streams processed in online and single-pass manner show how CNUs can mitigate forgetting without any replays or more informed learning criteria, while keeping competitive or better performance with respect to continual learning methods that explicitly store and replay data
The Mag-Gripper: A Soft-Rigid Gripper Augmented With an Electromagnet to Precisely Handle Clothes
No full textThis letter introduces Mag-Gripper, a novel robotic gripper specifically designed for autonomous clothing manipulation. It is capable of improving grasp repeatability, and precision, compensating uncertainties in the target grasping locations. We propose to approach the autonomous clothing manipulation challenge by involving a suitable magnetic force. For this reason, Mag-Gripperis equipped with an electromagnet capable of interacting with small metal parts properly placed on the garment to be grasped. Electromagnet exploitation is not a novelty in literature, but our design innovation consists in embedding the electromagnet in the structure of a jaw gripper. In so doing, we revisit a classic end-effector type, corresponding to the simplest representation of a hand capable of opposability, allowing easily controllable devices to perform grasps similar to the human pinch grasp. Mag-Grippercan find applications either in Research labs investigating Machine Learning-based clothing manipulation techniques either in companies having to manage a large amount of returns, either in home setting scenarios
Acute zonal occult outer retinopathy complex disease: Lessons learned about choroid, photoreceptors, and retinal function
No full textPurpose: Retinal photoreceptors layer integrity is considered essential to visual function. We report a case of acute zonal occult outer retinopathy (AZOOR) complex disease (namely AIBSE: acute idiopathic blind spot enlargement) in which apparently a full anatomic regeneration is not needed for a complete functional recovery. Methods: Case report with multimodal imaging. Reports: Visual field recovery in the presence of photoreceptors layer disruption studied by means of Optical Coherence Tomography. Choroid and photoreceptors layer thickness thinned progressively during recovery. Conclusion: This case suggests that anatomical retinal integrity as shown by OCT does not always correspond to visual function. Our case highlights that a complete visual recovery can occur even when structural abnormalities are still observable
Developing constrained neural units over time
No full textIn this paper we present a foundational study on a constrained method that defines learning problems with Neural Networks in the context of the Least Cognitive Action, which very much resembles the Least Action Principle in mechanics. Starting from a general approach to enforce constraints into the dynamical laws of learning, this work focuses on an alternative way of defining Neural Networks, that is different from the majority of existing approaches. In particular, the structure of the neural architecture is defined by means of a special class of constraints that are extended also to the interaction with data, leading to “architectural” and “input-related” constraints, respectively. The proposed theory is cast into the time domain, in which data are presented to the network in an ordered manner, that makes this study an important step toward alternative ways of processing continuous streams of data with Neural Networks. The connection with the classic Backpropagation-based update rule of the weights of networks is discussed, showing that there are conditions under which our approach degenerates to Backpropagation. Moreover, the theory is experimentally evaluated on a simple problem that allows us to deeply study several aspects of the theory itself and to show the soundness of the model
- …
