The recent success of machine learning systems on various QA datasets could be interpreted as a significant improvement in models' language understanding abilities. However, using various perturbations, multiple recent works have shown that good performance on a dataset might not indicate performance that correlates well with human's expectations from models that "understand" language. In this work we consider a top performing model on several Multiple Choice Question Answering (MCQA) datasets, and evaluate it against a set of expectations one might have from such a model, using a series of zero-information perturbations of the model's inputs. Our results show that the model clearly falls short of our expectations, and motivates a modified training approach that forces the model to better attend to the inputs. We show that the new training paradigm leads to a model that performs on par with the original model while better satisfying our expectations.

We study the distribution of singular values of product of random matrices pertinent to the analysis of deep neural networks. The matrices resemble the product of the sample covariance matrices, however, an important difference is that the population covariance matrices assumed to be non-random or random but independent of the random data matrix in statistics and random matrix theory are now certain functions of random data matrices (synaptic weight matrices in the deep neural network terminology). The problem has been treated in recent work [25, 13] by using the techniques of free probability theory. Since, however, free probability theory deals with population covariance matrices which are independent of the data matrices, its applicability has to be justified. The justification has been given in [22] for Gaussian data matrices with independent entries, a standard analytical model of free probability, by using a version of the techniques of random matrix theory. In this paper we use another, more streamlined, version of the techniques of random matrix theory to generalize the results of [22] to the case where the entries of the synaptic weight matrices are just independent identically distributed random variables with zero mean and finite fourth moment. This, in particular, extends the property of the so-called macroscopic universality on the considered random matrices.

The predictive normalized maximum likelihood (pNML) approach has recently been proposed as the min-max optimal solution to the batch learning problem where both the training set and the test data feature are individuals, known sequences. This approach has yields a learnability measure that can also be interpreted as a stability measure. This measure has shown some potential in detecting out-of-distribution examples, yet it has considerable computational costs. In this project, we propose and analyze an approximation of the pNML, which is based on influence functions. Combining both theoretical analysis and experiments, we show that when applied to neural networks, this approximation can detect out-of-distribution examples effectively. We also compare its performance to that achieved by conducting a single gradient step for each possible label.

CATEGORY WINNER: MOLECULAR MEDICINE Dasha Nelidova Dasha Nelidova completed her undergraduate degrees at the University of Auckland, New Zealand. She completed her Ph.D. in neurobiology at the Friedrich Miescher Institute for Biomedical Research in Basel, Switzerland. Nelidova is currently a postdoctoral researcher at the Institute of Molecular and Clinical Ophthalmology Basel, where she is working to develop new translational technologies for treating retinal diseases that lead to blindness. [ www.sciencemag.org/content/370/6519/925.2 ][1] Photoreceptor degeneration, including age-related macular degeneration and retinitis pigmentosa, is a leading cause of blindness worldwide. Repair of retinal neurons by optogenetics—a technology that sensitizes neurons to light through the transfer of genes for light-sensitive proteins of microbial origin ([ 1 ][2], [ 2 ][3])—has entered clinical trials ([ 3 ][4], [ 4 ][5]). Trials began in 2018 in patients with advanced retinitis pigmentosa and minimal remaining vision ([ 4 ][5]). Optogenetic proteins are sensitive only to the brightest visible light, at intensities that overwhelm surviving functional photoreceptors. Yet, in a number of blinding diseases, light-sensitive and light-insensitive photoreceptor zones coexist within the same retina. In macular degeneration, for example, cone photoreceptors of the central retina lose their light sensitivity. Surrounding photoreceptors remain viable, and peripheral vision is largely unaffected. A key challenge for new translational technologies that aim to restore image-acquiring properties of the retina is the compatibility of such technologies with remaining vision. We reasoned that sensitizing the retina to wavelengths that functional photoreceptors are unable to detect (>900 nm) could supplement deteriorating natural vision, without interfering with the ability to see the visible spectrum. Inspired by infrared vision in snakes, we developed nanogenetic molecular tools that allowed blind mice and ex vivo human retinas to detect near-infrared (NIR) light ([ 5 ][6]). Snakes can see the world in two different ways. Like humans, they make use of their eyes to detect wavelengths of the visible spectrum (400 to 700 nm). In addition, several species can also generate thermal images ([ 6 ][7]). Snakes detect infrared light (1 to 30 μm) using temperature-sensitive transient receptor potential (TRP) cation channels expressed in a specialized “pit” organ ([ 6 ][7]). Infrared and visible spectrum images superimpose within the brain ([ 7 ][8]), presumably enabling the animals to react to the environment with greater precision than what is possible by using only a single image. Snakes can switch back and forth between the two imaging systems or use both simultaneously ([ 7 ][8], [ 8 ][9]). TRP channels could potentially be targeted to mammalian retinal cell types to make them sensitive to infrared radiation. However, infrared light would raise vibrational energies of water molecules throughout the eye. Shorter wavelength NIR light would be preferable because NIR has lower water absorption, although this same feature also makes direct NIR illumination an inefficient activator of TRP channels. To develop a more efficient NIR light detector for retinal cell types, we engineered a dual system that consists of a genetic and a nanomaterial component (see the figure). The genetic half of the sensor consists of TRP channels, engineered to incorporate an extracellular protein epitope tag recognizable by a specific antibody ([ 9 ][10]). The nanomaterial half of the sensor consists of gold nanorods conjugated to an antibody against the epitope ([ 10 ][11]). Gold nanorods serve as antennas for NIR light and convert light into local heat through surface plasmon resonance ([ 11 ][12]), driving photocurrents through antibody-bound TRP channels. Subretinal microinjection of virally packaged TRP and nanorods delivered the sensor components to cones. Our initial system was based on TRP vanilloid 1 (TRPV1) channels and gold nanorods with absorption maxima at 915 nm. We began by inserting a 6x-His epitope tag into the middle of the first TRPV1 extracellular loop, measuring sizes of evoked currents before and after the modification, and confirmed that channels remained functional. Next, we used adeno-associated virus (AAV)–mediated gene transfer to transduce cone photoreceptors of blind mice with the nanogenetic sensor. To measure neural activity, we performed two-photon calcium imaging of individual neurons within the retina and primary visual cortex. Expression of the nanogenetic sensor in cones rendered blind retinas to be sensitive to NIR light. Cone photoreceptors (retinal input) and retinal ganglion cells (retinal output) responded vigorously to 915-nm light, and NIR-evoked retinal activity propagated to the brain. This allowed treated mice to use their newly acquired NIR vision to perform behavioral tasks. In complementary experiments, we confirmed that NIR light was unable to activate wild-type cones and did not affect their visible light responses. Similarly, awake, wildtype mice failed to exploit NIR light cues during behavioral training. Nanorod properties depend on size and shape ([ 11 ][12]). By changing the length of the gold nanorods from ∼80 nm to ∼120 nm, we tuned NIR vision to a different NIR wavelength (980 nm). Wavelength tuning is important for several reasons. Certain NIR wavelengths might be better tolerated by patients than others. Also, maximum permissible light doses for the human eye depend on the wavelength. Additionally, NIR vision requires eye goggles that project images composed of specific NIR wavelengths onto the retina. Compatibility with current and future NIR projectors requires tunable NIR detectors. Across the animal kingdom, multiple variants of thermosensitive proteins can be found, and more can be created through mutagenesis. Channels, tags, and antibodies can be modified to gain additional desirable properties. We selected TRP ankyrin 1 (TRPA1) channels from the Texas rat snake because of their lower thermal thresholds and inserted the newer epitope tag OLLAS ( Escherichia coli OmpF Linker and mouse Langerin fusion sequence) ([ 12 ][13]) into the first extracellular loop. Mice transfected with engineered TRPA1 channels were better able to anticipate water rewards when lights were dimmed as compared with mice transfected with TRPV1, indicating an improvement in the sensitivity of the sensor. (Both TRPA1- and TRPV1-transduced animals performed behavioral tasks as well as wild-type animals that were trained by using visible light.) The next step was to validate findings in blind human retinas. To do this, we targeted TRPV1 and gold nanorods to light-insensitive photoreceptors of adult human ex vivo retinal explants. (We had previously developed a cocktail of molecules to keep human retinas alive for 8 weeks post mortem, giving gene expression time to take hold.) We then recorded NIR light–evoked calcium activity and saw fast, strong activation of human photoreceptors and downstream retinal neurons, including ganglion cells. Taken together, these experiments provide proof of principle for the potential therapeutic translation of this technology. Light intensities required to drive genetically encoded NIR sensors met existing safety standards that specify exposure limits for the human eye, and we further demonstrated that components of the sensor may be exchanged, with predictable final outcomes. In the future, targeted central repair would allow an island of NIR sensitivity to be built in a sea of natural vision. Parallel developments in surgery ([ 13 ][14]) and NIR projectors with eye-tracking capabilities ([ 4 ][5]) make targeted central repair feasible. Ultimately, the user may be able to self-select the region of the electromagnetic spectrum most useful to view the external world, a decision guided by the state of their retina and ambient light conditions. 1. [↵][15]1. J. A. Sahel, 2. B. Roska , Annu. Rev. Neurosci. 36, 467 (2013). [OpenUrl][16][CrossRef][17][PubMed][18][Web of Science][19] 2. [↵][20]1. V. Busskamp et al ., Science 329, 413 (2010). [OpenUrl][21][Abstract/FREE Full Text][22] 3. [↵][23]1. S. Makin , Nat. Outlook 10.1038/d41586-019-01107-8 (2019). 4. [↵][24]Dose-escalation Study to Evaluate the Safety and Tolerability of GS030 in Subjects With Retinitis Pigmentosa (PIONEER). Clinical Trials ID: NCT03326336 (2018). 5. [↵][25]1. D. Nelidova et al ., Science 368, 1108 (2020). [OpenUrl][26][Abstract/FREE Full Text][27] 6. [↵][28]1. E. O. Gracheva et al ., Nature 464, 1006 (2010). [OpenUrl][29][CrossRef][30][PubMed][31][Web of Science][32] 7. [↵][33]1. E. A. Newman, 2. P. H. Hartline , Science 213, 789 (1981). [OpenUrl][34][Abstract/FREE Full Text][35] 8. [↵][36]1. E. A. Newman, 2. P. H. Hartline , Sci. Am. 246, 116 (1982). [OpenUrl][37] 9. [↵][38]1. S. A. Stanley et al ., Science 336, 604 (2012). [OpenUrl][39][Abstract/FREE Full Text][40] 10. [↵][41]1. P. P. Joshi, 2. S. J. Yoon, 3. W. G. Hardin, 4. S. Emelianov, 5. K. V. Sokolov , Bioconjug. Chem. 24, 878 (2013). [OpenUrl][42][CrossRef][43][PubMed][44] 11. [↵][45]1. Z. Qin, 2. J. C. Bischof , Chem. Soc. Rev. 41, 1191 (2012). [OpenUrl][46][CrossRef][47][PubMed][48] 12. [↵][49]1. S. H. Park et al ., J. Immunol. Methods 331, 27 (2008). [OpenUrl][50][CrossRef][51][PubMed][52][Web of Science][53] 13. [↵][54]1. A. M. Maguire et al ., N. Engl. J. Med. 358, 2240 (2008). [OpenUrl][55][CrossRef][56][PubMed][57][Web of Science][58] Acknowledgments: I thank my thesis adviser, B. Roska, and all molecular and clinical research colleagues at the Institute of Molecular and Clinical Ophthalmology Basel for their enthusiasm, advice, and help. I also thank our collaborators, especially A. Szabo, without whom human retinal experiments would not have been possible. [1]: http://www.sciencemag.org/content/370/6519/925.2 [2]: #ref-1 [3]: #ref-2 [4]: #ref-3 [5]: #ref-4 [6]: #ref-5 [7]: #ref-6 [8]: #ref-7 [9]: #ref-8 [10]: #ref-9 [11]: #ref-10 [12]: #ref-11 [13]: #ref-12 [14]: #ref-13 [15]: #xref-ref-1-1 "View reference 1 in text" [16]: {openurl}?query=rft.jtitle%253DAnnu.%2BRev.%2BNeurosci%26rft.volume%253D36%26rft.spage%253D467%26rft_id%253Dinfo%253Adoi%252F10.1146%252Fannurev-neuro-062012-170304%26rft_id%253Dinfo%253Apmid%252F23724995%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [17]: /lookup/external-ref?access_num=10.1146/annurev-neuro-062012-170304&link_type=DOI [18]: /lookup/external-ref?access_num=23724995&link_type=MED&atom=%2Fsci%2F370%2F6519%2F925.2.atom [19]: /lookup/external-ref?access_num=000323892300021&link_type=ISI [20]: #xref-ref-2-1 "View reference 2 in text" [21]: {openurl}?query=rft.jtitle%253DScience%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.1190897%26rft_id%253Dinfo%253Apmid%252F20576849%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [22]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEyOiIzMjkvNTk5MC80MTMiO3M6NDoiYXRvbSI7czoyNDoiL3NjaS8zNzAvNjUxOS85MjUuMi5hdG9tIjt9czo4OiJmcmFnbWVudCI7czowOiIiO30= [23]: #xref-ref-3-1 "View reference 3 in text" [24]: #xref-ref-4-1 "View reference 4 in text" [25]: #xref-ref-5-1 "View reference 5 in text" [26]: {openurl}?query=rft.jtitle%253DScience%26rft.stitle%253DScience%26rft.aulast%253DNelidova%26rft.auinit1%253DD.%26rft.volume%253D368%26rft.issue%253D6495%26rft.spage%253D1108%26rft.epage%253D1113%26rft.atitle%253DRestoring%2Blight%2Bsensitivity%2Busing%2Btunable%2Bnear-infrared%2Bsensors%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.aaz5887%26rft_id%253Dinfo%253Apmid%252F32499439%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [27]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEzOiIzNjgvNjQ5NS8xMTA4IjtzOjQ6ImF0b20iO3M6MjQ6Ii9zY2kvMzcwLzY1MTkvOTI1LjIuYXRvbSI7fXM6ODoiZnJhZ21lbnQiO3M6MDoiIjt9 [28]: #xref-ref-6-1 "View reference 6 in text" [29]: {openurl}?query=rft.jtitle%253DNature%26rft.stitle%253DNature%26rft.aulast%253DGracheva%26rft.auinit1%253DE.%2BO.%26rft.volume%253D464%26rft.issue%253D7291%26rft.spage%253D1006%26rft.epage%253D1011%26rft.atitle%253DMolecular%2Bbasis%2Bof%2Binfrared%2Bdetection%2Bby%2Bsnakes.%26rft_id%253Dinfo%253Adoi%252F10.1038%252Fnature08943%26rft_id%253Dinfo%253Apmid%252F20228791%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [30]: /lookup/external-ref?access_num=10.1038/nature08943&link_type=DOI [31]: /lookup/external-ref?access_num=20228791&link_type=MED&atom=%2Fsci%2F370%2F6519%2F925.2.atom [32]: /lookup/external-ref?access_num=000276635000031&link_type=ISI [33]: #xref-ref-7-1 "View reference 7 in text" [34]: {openurl}?query=rft.jtitle%253DScience%26rft.stitle%253DScience%26rft.aulast%253DNewman%26rft.auinit1%253DE.%26rft.volume%253D213%26rft.issue%253D4509%26rft.spage%253D789%26rft.epage%253D791%26rft.atitle%253DIntegration%2Bof%2Bvisual%2Band%2Binfrared%2Binformation%2Bin%2Bbimodal%2Bneurons%2Bin%2Bthe%2Brattlesnake%2Boptic%2Btectum%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.7256281%26rft_id%253Dinfo%253Apmid%252F7256281%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [35]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEyOiIyMTMvNDUwOS83ODkiO3M6NDoiYXRvbSI7czoyNDoiL3NjaS8zNzAvNjUxOS85MjUuMi5hdG9tIjt9czo4OiJmcmFnbWVudCI7czowOiIiO30= [36]: #xref-ref-8-1 "View reference 8 in text" [37]: {openurl}?query=rft.jtitle%253DSci.%2BAm%26rft.volume%253D246%26rft.spage%253D116%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [38]: #xref-ref-9-1 "View reference 9 in text" [39]: {openurl}?query=rft.jtitle%253DScience%26rft.stitle%253DScience%26rft.aulast%253DStanley%26rft.auinit1%253DS.%2BA.%26rft.volume%253D336%26rft.issue%253D6081%26rft.spage%253D604%26rft.epage%253D608%26rft.atitle%253DRadio-Wave%2BHeating%2Bof%2BIron%2BOxide%2BNanoparticles%2BCan%2BRegulate%2BPlasma%2BGlucose%2Bin%2BMice%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.1216753%26rft_id%253Dinfo%253Apmid%252F22556257%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [40]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEyOiIzMzYvNjA4MS82MDQiO3M6NDoiYXRvbSI7czoyNDoiL3NjaS8zNzAvNjUxOS85MjUuMi5hdG9tIjt9czo4OiJmcmFnbWVudCI7czowOiIiO30= [41]: #xref-ref-10-1 "View reference 10 in text" [42]: {openurl}?query=rft.jtitle%253DBioconjug.%2BChem%26rft.volume%253D24%26rft.spage%253D878%26rft_id%253Dinfo%253Adoi%252F10.1021%252Fbc3004815%26rft_id%253Dinfo%253Apmid%252F23631707%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [43]: /lookup/external-ref?access_num=10.1021/bc3004815&link_type=DOI [44]: /lookup/external-ref?access_num=23631707&link_type=MED&atom=%2Fsci%2F370%2F6519%2F925.2.atom [45]: #xref-ref-11-1 "View reference 11 in text" [46]: {openurl}?query=rft.stitle%253DChem%2BSoc%2BRev%26rft.aulast%253DQin%26rft.auinit1%253DZ.%26rft.volume%253D41%26rft.issue%253D3%26rft.spage%253D1191%26rft.epage%253D1217%26rft.atitle%253DThermophysical%2Band%2Bbiological%2Bresponses%2Bof%2Bgold%2Bnanoparticle%2Blaser%2Bheating.%26rft_id%253Dinfo%253Adoi%252F10.1039%252FC1CS15184C%26rft_id%253Dinfo%253Apmid%252F21947414%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [47]: /lookup/external-ref?access_num=10.1039/C1CS15184C&link_type=DOI [48]: /lookup/external-ref?access_num=21947414&link_type=MED&atom=%2Fsci%2F370%2F6519%2F925.2.atom [49]: #xref-ref-12-1 "View reference 12 in text" [50]: {openurl}?query=rft.jtitle%253DJournal%2Bof%2Bimmunological%2Bmethods%26rft.stitle%253DJ%2BImmunol%2BMethods%26rft.aulast%253DPark%26rft.auinit1%253DS.%2BH.%26rft.volume%253D331%26rft.issue%253D1-2%26rft.spage%253D27%26rft.epage%253D38%26rft.atitle%253DGeneration%2Band%2Bapplication%2Bof%2Bnew%2Brat%2Bmonoclonal%2Bantibodies%2Bagainst%2Bsynthetic%2BFLAG%2Band%2BOLLAS%2Btags%2Bfor%2Bimproved%2Bimmunodetection.%26rft_id%253Dinfo%253Adoi%252F10.1016%252Fj.jim.2007.10.012%26rft_id%253Dinfo%253Apmid%252F18054954%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [51]: /lookup/external-ref?access_num=10.1016/j.jim.2007.10.012&link_type=DOI [52]: /lookup/external-ref?access_num=18054954&link_type=MED&atom=%2Fsci%2F370%2F6519%2F925.2.atom [53]: /lookup/external-ref?access_num=000254059400003&link_type=ISI [54]: #xref-ref-13-1 "View reference 13 in text" [55]: {openurl}?query=rft.jtitle%253DNew%2BEngland%2BJournal%2Bof%2BMedicine%26rft.stitle%253DNEJM%26rft.volume%253D358%26rft.issue%253D21%26rft.spage%253D2240%26rft.epage%253D2248%26rft.atitle%253DSafety%2Band%2Befficacy%2Bof%2Bgene%2Btransfer%2Bfor%2BLeber%2527s%2Bcongenital%2Bamaurosis.%26rft_id%253Dinfo%253Adoi%252F10.1056%252FNEJMoa0802315%26rft_id%253Dinfo%253Apmid%252F18441370%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [56]: /lookup/external-ref?access_num=10.1056/NEJMoa0802315&link_type=DOI [57]: /lookup/external-ref?access_num=18441370&link_type=MED&atom=%2Fsci%2F370%2F6519%2F925.2.atom [58]: /lookup/external-ref?access_num=000256023600006&link_type=ISI

We each developed from a single cell—a fertilized egg—that divided and divided and eventually gave rise to the trillions of cells, of hundreds of types, that constitute the tissues and organs of our adult bodies. Advancing our understanding of the molecular programs underlying the emergence and differentiation of these diverse cell types is of fundamental interest and will affect almost every aspect of biology and medicine. Recently, technological advances have made it possible to directly measure the gene expression patterns of individual cells ([ 1 ][1]). Such methods can be used to clarify cell types and to determine the developmental stage of individual cells ([ 2 ][2]). Single-cell transcriptional profiling of successive developmental stages has the potential to be particularly informative, as the data can be used to reconstruct developmental processes, as well as characterize the underlying genetic programs ([ 3 ][3], [ 4 ][4]). ![Figure][5] A genomic technique for tracking cellular development High-throughput single-cell genomic methods enable a global view of cell type diversifcation by transcriptome and epigenome CREDIT: N. DESAI/ SCIENCE FROM CAO ET AL. ([7][6]) AND BIORENDER When I began my doctoral studies in Jay Shendure's lab at the University of Washington, available single-cell sequencing techniques relied on the isolation of individual cells within physical compartments and thus were limited in terms of both throughput and cost. As a graduate student, I developed four high-throughput single-cell genomic techniques to overcome these limitations ([ 5 ][7]–[ 8 ][8]). Leveraging these methods, I profiled millions of single-cell transcriptomes from organisms, in species that included worms, mice, and humans. By quantifying the dynamics of embryonic development at single-cell resolution, I was able to map out the global genetic programs that control cell proliferation and differentiation at the whole-organism scale. By the 1980s, biologists had documented every developmental step in the nematode Caenorhabditis elegans , from a single-cell embryo to the adult worm, and mapped the connections of all of the worm's neurons ([ 9 ][9]). However, although the nematode worm has a relatively small cell number (558 cells at hatching), a comprehensive understanding of the molecular basis for the specification of these cell types remains difficult. To resolve cellular heterogeneity, I first developed a method to specifically label the transcriptomes of large numbers of single cells, which we called sci-RNA-seq (single-cell combinatorial indexing RNA sequencing) ([ 5 ][7]). This method is based on combinatorial indexing, a strategy using split-pool barcoding of nucleic acids to label vast numbers of single cells within a single experiment ([ 9 ][9]). In this study, I profiled nearly 50,000 cells from C. elegans at the L2 stage, which is more than 50-fold “shotgun cellular coverage” of its somatic cell composition. We further defined consensus expression profiles for 27 cell types and identified rare neuronal cell types corresponding to as few as one or two cells in the L2 worm. This was the first study to show that single-cell transcriptional profiling is sufficient to separate all major cell types from an entire animal. C. elegans development follows a tightly controlled genetic program. Other multicellular organisms, such as mice and humans, have much more developmental flexibility. However, conventional approaches for mammalian single-cell profiling lack the throughput and resolution to obtain a global view of the molecular states and trajectories of the rapidly diversifying and expanding cell types. To investigate cell state dynamics in mammalian development, I developed an even more scalable single-cell profiling technique, sci-RNA-seq3 ([ 7 ][6]), and used it to trace the development path of 2 million mouse cells as they traversed diverse paths in a 4-day window of development corresponding to organogenesis (embryonic day 9.5 to embryonic day 13.5). From these data, we characterized the dynamics of cell proliferation and key regulators for each cell lineage, a potentially foundational resource for understanding how the hundreds of cell types forming a mammalian body are generated in development. This was, and remains, the largest publicly available single-cell transcriptional dataset. The sci-RNA-seq3 method enabled this dataset to be generated rapidly, within a few weeks, by a single individual. A major challenge regarding current single-cell assays is that nearly all such methods capture just one aspect of cellular biology (typically mRNA expression), limiting the ability to relate different components to one another and to infer causal relationships. Another technique that I developed, sci-CAR (single-cell combinatorial indexing chromatin accessibility and mRNA) ([ 6 ][10]), was created with the goal of overcoming this limitation, allowing the user to jointly profile the epigenome (chromatin accessibility) and transcriptome (mRNA). I applied sci-CAR to the mouse whole kidneys, recovering all major cell types and linking cis-regulatory sites to their target genes on the basis of the covariance of chromatin accessibility and transcription across large numbers of single cells. To further explore the gene regulatory mechanisms, I invented sci-fate ([ 8 ][8]), a new method that identifies the temporal dynamics of transcription by distinguishing newly synthesized mRNA transcripts from “older” mRNA transcripts in thousands of individual cells. Applying the strategy to cancer cell state dynamics in response to glucocorticoids, we were able to link transcription factors (TFs) with their target genes on the basis of the covariance between TF expression and the amount of newly synthesized RNA across thousands of cells. In summary, my dissertation involved developing the technical framework for quantifying gene expression and chromatin dynamics across thousands to millions of single cells and applying these technologies to profile complex, developing organisms. The methods that I developed enable such projects to be achievable by a single individual, rather than requiring large consortia. Looking ahead, I anticipate that the integration of single-cell views of the transcriptome, epigenome, proteome, and spatial-temporal information throughout development will enable an increasingly complete view of how life is formed. GRAND PRIZE WINNER Junyue Cao Junyue Cao received his undergraduate degree from Peking University and a Ph.D. from the University of Washington. After completing his postdoctoral fellowship at the University of Washington, Junyue Cao started his lab as an assistant professor and lab head of single-cell genomics and population dynamics at the Rockefeller University in 2020. His current research focuses on studying how a cell population in our body maintains homeostasis by developing genomic techniques to profile and perturb cell dynamics at single-cell resolution. CATEGORY WINNER: ECOLOGY AND EVOLUTION Orsi Decker Orsi Decker completed her undergraduate degree at Eötvös Loránd University in Budapest, Hungary. She went on to receive her master's degree in Ecology and Evolution at the University of Amsterdam. Decker completed her doctoral research at La Trobe University in Melbourne, Australia, where she investigated the extinctions of native digging mammals and their context-dependent impacts on soil processes. She is currently a postdoctoral researcher at La Trobe University, where she is examining how land restoration efforts could be improved to regain soil functions through the introduction of soil fauna to degraded areas. [www.sciencemag.org/content/370/6519/925.1][11] CATEGORY WINNER: MOLECULAR MEDICINE Dasha Nelidova Dasha Nelidova completed her undergraduate degrees at the University of Auckland, New Zealand. She completed her Ph.D. in neurobiology at the Friedrich Miescher Institute for Biomedical Research in Basel, Switzerland. Nelidova is currently a postdoctoral researcher at the Institute of Molecular and Clinical Ophthalmology Basel, where she is working to develop new translational technologies for treating retinal diseases that lead to blindness. [www.sciencemag.org/content/370/6519/925.2][12] CATEGORY WINNER: CELL AND MOLECULAR BIOLOGY William E. Allen William E. Allen received his undergraduate degree from Brown University in 2012, M.Phil. in Computational Biology from the University of Cambridge in 2013, and Ph.D. in Neurosciences from Stanford University in 2019. At Stanford, he worked to develop new tools for the large-scale characterization of neural circuit structure and function, which he applied to understand the neural basis of thirst. After completing his Ph.D., William started as an independent Junior Fellow in the Society of Fellows at Harvard University, where he is developing and applying new approaches to map mammalian brain function and dysfunction over an animal's life span. [www.sciencemag.org/content/370/6519/925.3][13] 1. [↵][14]1. D. Ramsköld et al ., Nat. Biotechnol. 30, 777 (2012). [OpenUrl][15][CrossRef][16][PubMed][17] 2. [↵][18]1. C. Trapnell , Genome Res. 25, 1491 (2015). [OpenUrl][19][Abstract/FREE Full Text][20] 3. [↵][21]1. D. E. Wagner et al ., Science 360, 981 (2018). [OpenUrl][22][Abstract/FREE Full Text][23] 4. [↵][24]1. K. Davie et al ., Cell 174, 982 (2018). [OpenUrl][25][CrossRef][26][PubMed][27] 5. [↵][28]1. J. Cao et al ., Science 357, 661 (2017). [OpenUrl][29][Abstract/FREE Full Text][30] 6. [↵][31]1. J. Cao et al ., Science 361, 1380 (2018). [OpenUrl][32][Abstract/FREE Full Text][33] 7. [↵][34]1. J. Cao et al ., Nature 566, 496 (2019). [OpenUrl][35][CrossRef][36][PubMed][37] 8. [↵][38]1. J. Cao, 2. W. Zhou, 3. F. Steemers, 4. C. Trapnell, 5. J. Shendure , Nat. Biotechnol. 38, 980 (2020). [OpenUrl][39][CrossRef][40][PubMed][41] 9. [↵][42]1. D. A. Cusanovich et al ., Science 348, 910 (2015). [OpenUrl][43][Abstract/FREE Full Text][44] [1]: #ref-1 [2]: #ref-2 [3]: #ref-3 [4]: #ref-4 [5]: pending:yes [6]: #ref-7 [7]: #ref-5 [8]: #ref-8 [9]: #ref-9 [10]: #ref-6 [11]: http://www.sciencemag.org/content/370/6519/925.1 [12]: http://www.sciencemag.org/content/370/6519/925.2 [13]: http://www.sciencemag.org/content/370/6519/925.3 [14]: #xref-ref-1-1 "View reference 1 in text" [15]: {openurl}?query=rft.jtitle%253DNature%2Bbiotechnology%26rft.stitle%253DNat%2BBiotechnol%26rft.aulast%253DRamskold%26rft.auinit1%253DD.%26rft.volume%253D30%26rft.issue%253D8%26rft.spage%253D777%26rft.epage%253D782%26rft.atitle%253DFull-length%2BmRNA-Seq%2Bfrom%2Bsingle-cell%2Blevels%2Bof%2BRNA%2Band%2Bindividual%2Bcirculating%2Btumor%2Bcells.%26rft_id%253Dinfo%253Adoi%252F10.1038%252Fnbt.2282%26rft_id%253Dinfo%253Apmid%252F22820318%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [16]: /lookup/external-ref?access_num=10.1038/nbt.2282&link_type=DOI [17]: /lookup/external-ref?access_num=22820318&link_type=MED&atom=%2Fsci%2F370%2F6519%2F924.atom [18]: #xref-ref-2-1 "View reference 2 in text" [19]: {openurl}?query=rft.jtitle%253DGenome%2BRes.%26rft_id%253Dinfo%253Adoi%252F10.1101%252Fgr.190595.115%26rft_id%253Dinfo%253Apmid%252F26430159%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [20]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6NjoiZ2Vub21lIjtzOjU6InJlc2lkIjtzOjEwOiIyNS8xMC8xNDkxIjtzOjQ6ImF0b20iO3M6MjI6Ii9zY2kvMzcwLzY1MTkvOTI0LmF0b20iO31zOjg6ImZyYWdtZW50IjtzOjA6IiI7fQ== [21]: #xref-ref-3-1 "View reference 3 in text" [22]: {openurl}?query=rft.jtitle%253DScience%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.aar4362%26rft_id%253Dinfo%253Apmid%252F29700229%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [23]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEyOiIzNjAvNjM5Mi85ODEiO3M6NDoiYXRvbSI7czoyMjoiL3NjaS8zNzAvNjUxOS85MjQuYXRvbSI7fXM6ODoiZnJhZ21lbnQiO3M6MDoiIjt9 [24]: #xref-ref-4-1 "View reference 4 in text" [25]: {openurl}?query=rft.jtitle%253DCell%26rft.volume%253D174%26rft.spage%253D982%26rft_id%253Dinfo%253Adoi%252F10.1016%252Fj.cell.2018.05.057%26rft_id%253Dinfo%253Apmid%252F29909982%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [26]: /lookup/external-ref?access_num=10.1016/j.cell.2018.05.057&link_type=DOI [27]: /lookup/external-ref?access_num=29909982&link_type=MED&atom=%2Fsci%2F370%2F6519%2F924.atom [28]: #xref-ref-5-1 "View reference 5 in text" [29]: {openurl}?query=rft.jtitle%253DScience%26rft.stitle%253DScience%26rft.aulast%253DCao%26rft.auinit1%253DJ.%26rft.volume%253D357%26rft.issue%253D6352%26rft.spage%253D661%26rft.epage%253D667%26rft.atitle%253DComprehensive%2Bsingle-cell%2Btranscriptional%2Bprofiling%2Bof%2Ba%2Bmulticellular%2Borganism%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.aam8940%26rft_id%253Dinfo%253Apmid%252F28818938%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [30]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEyOiIzNTcvNjM1Mi82NjEiO3M6NDoiYXRvbSI7czoyMjoiL3NjaS8zNzAvNjUxOS85MjQuYXRvbSI7fXM6ODoiZnJhZ21lbnQiO3M6MDoiIjt9 [31]: #xref-ref-6-1 "View reference 6 in text" [32]: {openurl}?query=rft.jtitle%253DScience%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.aau0730%26rft_id%253Dinfo%253Apmid%252F30166440%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [33]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEzOiIzNjEvNjQwOS8xMzgwIjtzOjQ6ImF0b20iO3M6MjI6Ii9zY2kvMzcwLzY1MTkvOTI0LmF0b20iO31zOjg6ImZyYWdtZW50IjtzOjA6IiI7fQ== [34]: #xref-ref-7-1 "View reference 7 in text" [35]: {openurl}?query=rft.jtitle%253DNature%26rft.volume%253D566%26rft.spage%253D496%26rft_id%253Dinfo%253Adoi%252F10.1038%252Fs41586-019-0969-x%26rft_id%253Dinfo%253Apmid%252F30787437%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [36]: /lookup/external-ref?access_num=10.1038/s41586-019-0969-x&link_type=DOI [37]: /lookup/external-ref?access_num=30787437&link_type=MED&atom=%2Fsci%2F370%2F6519%2F924.atom [38]: #xref-ref-8-1 "View reference 8 in text" [39]: {openurl}?query=rft.jtitle%253DNat.%2BBiotechnol.%26rft.volume%253D38%26rft.spage%253D980%26rft_id%253Dinfo%253Adoi%252F10.1038%252Fs41587-020-0480-9%26rft_id%253Dinfo%253Apmid%252F32284584%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [40]: /lookup/external-ref?access_num=10.1038/s41587-020-0480-9&link_type=DOI [41]: /lookup/external-ref?access_num=32284584&link_type=MED&atom=%2Fsci%2F370%2F6519%2F924.atom [42]: #xref-ref-9-1 "View reference 9 in text" [43]: {openurl}?query=rft.jtitle%253DScience%26rft_id%253Dinfo%253Adoi%252F10.1126%252Fscience.aab1601%26rft_id%253Dinfo%253Apmid%252F25953818%26rft.genre%253Darticle%26rft_val_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Ajournal%26ctx_ver%253DZ39.88-2004%26url_ver%253DZ39.88-2004%26url_ctx_fmt%253Dinfo%253Aofi%252Ffmt%253Akev%253Amtx%253Actx [44]: /lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6Mzoic2NpIjtzOjU6InJlc2lkIjtzOjEyOiIzNDgvNjIzNy85MTAiO3M6NDoiYXRvbSI7czoyMjoiL3NjaS8zNzAvNjUxOS85MjQuYXRvbSI7fXM6ODoiZnJhZ21lbnQiO3M6MDoiIjt9

An artificial intelligence that can diagnose tinnitus based on the results of brain imaging, rather than subjective tests, may help improve treatments for the condition. Mehrnaz Shoushtarian at the Bionics Institute in Melbourne, Australia, and her colleagues have developed an algorithm that can detect whether a person has tinnitus, and also how severe it is. The AI can spot the presence of tinnitus with 78 per cent accuracy, and distinguish between mild and severe forms with 87 per cent accuracy. Chronic tinnitus affects around 15 per cent of adults. The condition is usually diagnosed by a hearing test, by self-reporting or based on a subjective questionnaire.

Multi-objective reinforcement learning (MORL) is an extension of ordinary, single-objective reinforcement learning (RL) that is applicable to many real world tasks where multiple objectives exist without known relative costs. We study the problem of single policy MORL, which learns an optimal policy given the preference of objectives. Existing methods require strong assumptions such as exact knowledge of the multi-objective Markov decision process, and are analyzed in the limit of infinite data and time. We propose a new algorithm called model-based envelop value iteration (EVI), which generalizes the enveloped multi-objective $Q$-learning algorithm in Yang, 2019. Our method can learn a near-optimal value function with polynomial sample complexity and linear convergence speed. To the best of our knowledge, this is the first finite-sample analysis of MORL algorithms.

We advance the modeling capability of electron particle precipitation from the magnetosphere to the ionosphere through a new database and use of machine learning tools to gain utility from those data. We have compiled, curated, analyzed, and made available a new and more capable database of particle precipitation data that includes 51 satellite years of Defense Meteorological Satellite Program (DMSP) observations temporally aligned with solar wind and geomagnetic activity data. The new total electron energy flux particle precipitation nowcast model, a neural network called PrecipNet, takes advantage of increased expressive power afforded by machine learning approaches to appropriately utilize diverse information from the solar wind and geomagnetic activity and, importantly, their time histories. With a more capable representation of the organizing parameters and the target electron energy flux observations, PrecipNet achieves a >50% reduction in errors from a current state-of-the-art model (OVATION Prime), better captures the dynamic changes of the auroral flux, and provides evidence that it can capably reconstruct mesoscale phenomena. We create and apply a new framework for space weather model evaluation that culminates previous guidance from across the solar-terrestrial research community. The research approach and results are representative of the `new frontier' of space weather research at the intersection of traditional and data science-driven discovery and provides a foundation for future efforts.

Traditionally, robust statistics has focused on designing estimators tolerant to a minority of contaminated data. Robust list-decodable learning focuses on the more challenging regime where only a minority $\frac 1 k$ fraction of the dataset is drawn from the distribution of interest, and no assumptions are made on the remaining data. We study the fundamental task of list-decodable mean estimation in high dimensions. Our main result is a new list-decodable mean estimation algorithm for bounded covariance distributions with optimal sample complexity and error rate, running in nearly-PCA time. Assuming the ground truth distribution on $\mathbb{R}^d$ has bounded covariance, our algorithm outputs a list of $O(k)$ candidate means, one of which is within distance $O(\sqrt{k})$ from the truth. Our algorithm runs in time $\widetilde{O}(ndk)$ for all $k = O(\sqrt{d}) \cup \Omega(d)$, where $n$ is the size of the dataset. We also show that a variant of our algorithm has runtime $\widetilde{O}(ndk)$ for all $k$, at the expense of an $O(\sqrt{\log k})$ factor in the recovery guarantee. This runtime matches up to logarithmic factors the cost of performing a single $k$-PCA on the data, which is a natural bottleneck of known algorithms for (very) special cases of our problem, such as clustering well-separated mixtures. Prior to our work, the fastest list-decodable mean estimation algorithms had runtimes $\widetilde{O}(n^2 d k^2)$ and $\widetilde{O}(nd k^{\ge 6})$. Our approach builds on a novel soft downweighting method, $\mathsf{SIFT}$, which is arguably the simplest known polynomial-time mean estimation technique in the list-decodable learning setting. To develop our fast algorithms, we boost the computational cost of $\mathsf{SIFT}$ via a careful "win-win-win" analysis of an approximate Ky Fan matrix multiplicative weights procedure we develop, which we believe may be of independent interest.

This study introduces a framework for quality control of measured weather data, including anomaly detection, and infilling missing values. Weather data is a fundamental input to building performance simulations, in which anomalous values defect the results while missing data lead to an unexpected termination of the simulation process. Traditionally, infilling missing values in weather data is performed through periodic or linear interpolations. However, when missing values exceed many consecutive hours, the accuracy of traditional methods is subject to debate. This study demonstrates how Neural Networks can increase the accuracy of data imputation when compared to other supervised learning methods. The framework is validated by predicting missing temperature and relative humidity data for an observation site, through a network of nearby weather stations in Milan, Italy. Results show that the proposed method can facilitate real-time building simulations with accurate and rapid quality control.