Search

Mahalanobis distance (MD) is a simple and popular post-processing method for detecting out-of-distribution (OOD) inputs in neural networks. We analyze its failure modes for near-OOD detection and propose a simple fix called relative Mahalanobis distance (RMD) which improves performance and is more robust to hyperparameter choice. On a wide selection of challenging vision, language, and biology OOD benchmarks (CIFAR-100 vs CIFAR-10, CLINC OOD intent detection, Genomics OOD), we show that RMD meaningfully improves upon MD performance (by up to 15% AUROC on genomics OOD).

Posté le 5 décembre 2020 sur BioRxiv; Spatiotemporal patterns of microtubule modifications such as acetylation underlie diverse cellular functions. While the molecular identity of the acetylating agent, α-tubulin N-acetyltransferase 1 (α-TAT1), as well as the functional consequences of microtubule acetylation have been revealed, the molecular mechanisms that regulate multi-tasking α-TAT1 action for dynamic acetylation remain obscure. Here we identified a signal motif in the intrinsically disordered C-terminus of α-TAT1, which comprises three functional elements - nuclear export, nuclear import and cytosolic retention. Their balance is tuned via phosphorylation by serine-threonine kinases to determine subcellular localization of α-TAT1. While the phosphorylated form binds to 14-3-3 adapters and accumulates in the cytosol for maximal substrate access, the non-phosphorylated form is sequestered inside the nucleus, thus keeping microtubule acetylation minimal. As cancer mutations have been reported to this motif, the unique ensemble regulation of α-TAT1 localization may hint at a role of microtubule acetylation in aberrant physiological conditions.

A generalized theory of spatial coherence for superposition of two speckle patterns with polarization diversity is presented. The presented theory deals with superposition in different scenarios i.e. superposition of two fully correlated, partially correlated and completely uncorrelated speckle patterns, and describes the effect on the spatial coherence-polarization (CP) property of the superposed random field from the study of the spatial degree of coherence and degree of polarization. The change in the spatial CP property with the polarization diversity is studied for different correlation factors of the constituent random fields, and a modulation of the spatial CP property is observed, when the constituent speckle patterns are not fully correlated. The effect of the average intensity ratio of the two random fields on the spatial CP property is also studied. Thus, this theoretical study completes the case of superposition of two speckle patterns with polarization diversity and its effect on the spatial CP property.

The random intensity distribution observed due to the propagation of a coherent beam of light through a scattering medium is known as a speckle pattern. The interaction of the speckles with a birefringent crystal, here, YVO4 results in the e-ray and o-ray speckles. It is observed that the e-ray and o-ray speckles experience an angular shift with the rotation of the crystal and the shift depends on the rotation angle of the crystal. It is also found that the rate of the shift is different for e-ray and o-ray speckles, and is independent of the topological charge of the beam incident on the scattering medium. The observed rate of shift of the speckle pattern is found to be same as that of the e-ray and o-ray beams. It is experimentally demonstrated that the rotation of a birefringent crystal can be tracked very accurately from the speckle shift measurement.

High-speed shadowgraphy showed droplets generated during flap cut with microkeratome. However, aerosolization of these droplets was unlikely because these droplets would settle on surface before aerosolization., Purpose: To quantify the atomization of liquid over the cornea during flap creation using microkeratome using high-speed shadowgraphy. Setting: Laboratory investigational study. Design: Laboratory study. Method: In an experimental setup, flap creation was performed on enucleated goat's eyes (n = 8) mounted on a stand using One Use-Plus SBK Moria microkeratome (Moria SA) to assess the spread of aerosols and droplets using high-speed shadowgraphy. Two conditions were computed. A constant airflow assumed uniform air velocity throughout the room. A decaying jet assumed that local air velocity at the site of measurements was smaller than the exit velocity from the air duct. Results: With the advancement of the microkeratome across the wet corneal surface, the atomization of a balanced salt solution was recorded on shadowgraphy. The minimum droplet size was ∼90 μm. The maximum distance traversed was ∼1.8 m and ∼1.3 m assuming a constant airflow (setting of refractive surgery theater) and decaying jet condition (setting of an operating theater with air-handling unit), respectively. Conclusions: The microkeratome-assisted LASIK flap creation does seem to cause spread of droplets. The droplet diameters and velocities did not permit the formation of aerosols. Therefore, the risk of transmission of the virus to the surgeon and surgical personnel due to the microkeratome procedure seems to be low.

Reliable detection of out-of-distribution (OOD) inputs is increasingly understood to be a precondition for deployment of machine learning systems. This paper proposes and investigates the use of contrastive training to boost OOD detection performance. Unlike leading methods for OOD detection, our approach does not require access to examples labeled explicitly as OOD, which can be difficult to collect in practice. We show in extensive experiments that contrastive training significantly helps OOD detection performance on a number of common benchmarks. By introducing and employing the Confusion Log Probability (CLP) score, which quantifies the difficulty of the OOD detection task by capturing the similarity of inlier and outlier datasets, we show that our method especially improves performance in the `near OOD' classes -- a particularly challenging setting for previous methods.

Non-invasive medical diagnostics demonstrate a propensity for droplet generation and should be studied to devise risk mitigation strategies against the spread of the SARS-CoV-2 virus. We investigate the air-puff tonometry, which uses a short-timed air-puff to applanate the human eye in a bid to detect the early onset of glaucoma by measuring the intraocular pressure. The air-puff consists of a vortex trailed by a high-speed jet. High-speed imaging of the eye during a typical tonometry measurement reveals a sequence of events starting with the interaction between the tear layer and the air puff leading to an initial sheet ejection. It is immediately followed by the trailing jet applanating the central corneal section, causing capillary waves to form and interact with the highly 3D transient expanding sheet. Such interaction with the capillary waves and the surrounding airfield due to the trailing jet causes the expanding sheet to undergo bag breakup, finger formation by Rayleigh Taylor instability and further break up into subsequent droplets by Rayleigh Plateau instability, which eventually splashes onto nearby objects, potentially forming fomites or aerosols which can lead to infections. The complex spatiotemporal phenomenon is carefully documented by rigorous experiments and corroborated using comprehensive theoretical analyses.

The ability of neural networks to continuously learn and adapt to new tasks while retaining prior knowledge is crucial for many applications. However, current neural networks tend to forget previously learned tasks when trained on new ones, i.e., they suffer from Catastrophic Forgetting (CF). The objective of Continual Learning (CL) is to alleviate this problem, which is particularly relevant for medical applications, where it may not be feasible to store and access previously used sensitive patient data. In this work, we propose a Continual Learning approach for brain segmentation, where a single network is consecutively trained on samples from different domains. We build upon an importance driven approach and adapt it for medical image segmentation. Particularly, we introduce learning rate regularization to prevent the loss of the network's knowledge. Our results demonstrate that directly restricting the adaptation of important network parameters clearly reduces Catastrophic Forgetting for segmentation across domains.