These are my recent Pin​board​.in links:

• [1107.0674] “Mem­ory foam” approach to unsu­per­vised learning

  “We pro­pose an alter­na­tive approach to con­struct an arti­fi­cial learn­ing sys­tem, which nat­u­rally learns in an unsu­per­vised man­ner. Its math­e­mat­i­cal pro­to­type is a dynam­i­cal sys­tem, which auto­mat­i­cally shapes its vec­tor field in response to the input sig­nal. The vec­tor field con­verges to a gra­di­ent of a multi-​​dimensional prob­a­bil­ity den­sity dis­tri­b­u­tion of the input process, taken with neg­a­tive sign. The most prob­a­ble pat­terns are rep­re­sented by the sta­ble fixed points, whose basins of attrac­tion are formed auto­mat­i­cally. The per­for­mance of this sys­tem is illus­trated with musi­cal signals.”
  
  machine-​​learning clas­si­fi­ca­tion learning-​​from-​​data algo­rithms nudge-​​targets

• [1107.0550] 3D Ter­res­trial LiDAR data clas­si­fi­ca­tion of com­plex nat­ural scenes using a multi-​​scale dimen­sion­al­ity cri­te­rion: appli­ca­tions in geomorphology

  “3D point clouds of nat­ural envi­ron­ments rel­e­vant to geo­mor­phol­ogy prob­lems (rivers, cliffs…) often require to clas­sify the data into ele­men­tary rel­e­vant classes. A typ­i­cal exam­ple is the sep­a­ra­tion of ripar­ian veg­e­ta­tion from soil in flu­vial envi­ron­ments, the dis­tinc­tion between fresh sur­faces and rock­fall in cliff envi­ron­ments, or more gen­er­ally the clas­si­fi­ca­tion of sur­faces accord­ing to their mor­phol­ogy (rip­ples, grain size…). Nat­ural sur­faces are very het­ero­ge­neous and their dis­tinc­tive prop­er­ties are sel­dom defined at a unique scale. We have thus defined a multi-​​scale mea­sure of the point cloud dimen­sion­al­ity around each point. The dimen­sion­al­ity char­ac­ter­izes the local 3D orga­ni­za­tion of the point cloud and varies from being 1D (points set along a line) to really tak­ing all 3D vol­ume, at each scale. We present the tech­nique and illus­trate its effi­ciency in sep­a­rat­ing ripar­ian veg­e­ta­tion from ground and clas­si­fy­ing a moun­tain stream in veg­e­ta­tion, rock, gravel and water sur­face. The supe­ri­or­ity of the multi-​​scale analy­sis in enhanc­ing class sep­a­ra­bil­ity and spa­tial res­o­lu­tion of the clas­si­fi­ca­tion is also demon­strated. Large scenes can be clas­si­fied on a com­mod­ity lap­top in a rea­son­able time. The tech­nique is robust to miss­ing data and espe­cially shadow zones. The clas­si­fi­ca­tion is fast and accu­rate and can account for some degree of intra-​​class mor­pho­log­i­cal vari­abil­ity such as dif­fer­ent veg­e­ta­tion types. A prob­a­bilis­tic con­fi­dence in the clas­si­fi­ca­tion result is given at each point allow­ing the user to remove the points for which the clas­si­fi­ca­tion is uncer­tain. The process can be both fully auto­mated but also fully cus­tomized by the user includ­ing a graph­i­cal def­i­n­i­tion of the clas­si­fiers if so desired. Although devel­oped for fully 3D data, the method can be read­ily applied to 2.5D air­borne LiDAR data.”
  
  image-​​analysis image-​​segmentation learning-​​from-​​data clas­si­fi­ca­tion nudge-​​targets

• [1105.6001] A Call to Arms: Revis­it­ing Data­base Design

  “Good data­base design is cru­cial to obtain a sound, con­sis­tent data­base, and — in turn — good data­base design method­olo­gies are the best way to achieve the right design. These method­olo­gies are taught to most Com­puter Sci­ence under­grad­u­ates, as part of any Intro­duc­tion to Data­base class. They can be con­sid­ered part of the “canon”, and indeed, the over­all approach to data­base design has been unchanged for years. More­over, none of the major data­base research assess­ments iden­tify data­base design as a strate­gic research direc­tion. Should we con­clude that data­base design is a solved prob­lem? Our the­sis is that data­base design remains a crit­i­cal unsolved prob­lem. Hence, it should be the sub­ject of more research. Our start­ing point is the obser­va­tion that tra­di­tional data­base design is not used in prac­tice — and if it were used it would result in designs that are not well adapted to cur­rent envi­ron­ments. In short, data­base design has failed to keep up with the times. In this paper, we put forth argu­ments to sup­port our view­point, ana­lyze the root causes of this sit­u­a­tion and sug­gest some avenues of research.”
  
  data­base ontol­ogy software-​​development computer-​​science design-​​patterns