BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Women @Cl Talklet Event - Mariana Marasoiu\, Maria Perez-Ortiz\, H
 elena Andrés Terre
DTSTART:20180309T130000Z
DTEND:20180309T140000Z
UID:TALK102478@talks.cam.ac.uk
CONTACT:Ayat Fekry
DESCRIPTION:--------------------------------------------------------------
 \n\n*Speaker*: Mariana Marasoiu\n\n*Title:* Towards end-user tools for int
 eractive data visualisation\n\n*Abstract:* More data is being generated to
 day than ever before\, and the number of people wanting to analyse and und
 erstand this data is growing. Much of this data analysis is done using vis
 ualisations\, from simple bar charts and scatter plots to interactive dash
 boards and infographics. Unfortunately\, the existing tools for visual dat
 a exploration are not very accessible to non-experts. \nIn order to ground
  the design of new tools for non-expert end-users in the real-world practi
 ce of data analysts\, I have conducted an ethnographically informed study 
 of a team of expert visual analysts. In this talk\, I will summarise this 
 fieldwork and will discuss early designs for a new visualisation tool.\n\n
 --------------------------------------------------------------\n\n*Speaker
 *: Maria Perez-Ortiz\n\n*Title:* Learning from humans and the law of compa
 rative judgment\n\n*Abstract:*The problem of how to elicit judgments from 
 humans has its roots in the well-established fields of psychophysics and s
 ensory evaluation and is of crucial interest in applications involving sub
 jective judgments. Methods for elicitation of human judgments are usually 
 categorised under the term of scaling\, i.e. the generation of a scale of 
 the observer's response to a stimuli. Scaling methods attempt to represent
  preference judgments on a line or multidimensional space\, so as to effec
 tively retain distances between tested objects. This scale usually reveals
  the underlying structure or unique relationships among the objects\, allo
 wing to measure and compare them in a meaningful way. This talk will revie
 w some of these approaches. \n\n------------------------------------------
 --------------------\n\n*Speaker*: Helena Andrés Terre\n\n*Title:* Using 
 Auto-Encoders to interpret Single Cell Transcriptomic Data\n\n*Abstract:* 
 The introduction of single cell RNA-seq was a major breakthrough in the e
 ld of biology\, and particularly useful for research in areas like compara
 tive transcriptomics or disease studies. Stem Cell's dierentiation has al
 so beneted from this new technique\, being able now to characterise gene 
 expression levels for individual cells\, and analyse the different stages 
 of the differentiation process. Computational analysis of such data is ess
 ential to understand the experimental results\, therefore new techniques a
 re needed in order to process and interpret the data. Our goal is to ident
 ify the most relevant drivers of the underlying processes captured by the 
 gene\nexpression proles. Current methods are based on linear dimensionali
 ty reduction techniques\, combined with further analysis to classify the c
 ells and identify the dierentiation stages. But the nature of their assum
 ptions generate some restrictions when trying to characterise middle state
 s of differentiation.\nWe have developed an unsupervised Machine Learning 
 technique for dimensionality reduction of single cell data. We are using A
 uto-Encoders to extract a number of signicant components that characteris
 e individual cells based on their gene expression\, using a deep learning 
 "bottle-neck" approach. The encoded space provides a new representation of
  the individual cells with uncorrelated components\, which can then be use
 d for further analysis and classication. The reconstruction ability of Au
 toencoders can also give an insight on the noise level and relevance of sp
 ecic genes to the process. We evaluate the performance of these networks 
 in terms of reconstruction accuracy and the information transferred to the
  encoded dimensions. I will give an overview of the implementation and the
  rst results we have obtained\, together with some of the future challeng
 es and questions we will be tackling.
LOCATION:Computer Laboratory\, William Gates Building\, Room FW11
END:VEVENT
END:VCALENDAR