Subject: Re: Request for information on ICA-Maximum likelihood approach From: Brian Gygi <bgygi@xxxxxxxx> Date: Mon, 12 Apr 2010 20:55:43 +0000 List-Archive:<http://lists.mcgill.ca/scripts/wa.exe?LIST=AUDITORY>----=_vm_0011_W558620254_11012_1271105743 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Doesn't that kind of negate the whole purpose of ICA? It's supposed to fi= nd the dimensions that PCA might miss. But if you do a PCA first, all you= will get out are subsets of the dimensions the PCA finds in the first pl= ace. Brian Gygi, Ph.D. Speech and Hearing Research Veterans Affairs Northern California Health Care System 150 Muir Road Martinez, CA 94553 (925) 372-2000 x5653 -----Original Message----- From: Linda Seltzer [mailto:lseltzer@xxxxxxxx Sent: Friday, April 9, 2010 09:47 PM To: AUDITORY@xxxxxxxx Subject: Re: Request for information on ICA-Maximum likelihood approach I should add that when ICA is used in fMRI (e.g. the use of fast ICA inBr= ain Voyager), PCA is used first for dimensionality reduction.Linda Seltze= r> Hi Vijay,>> I'm not sure you could start a Matlab script from scratch = to do ICA> with Maximum likelihood approach. As far as I've gathered, the= re's> quite a lot of (heavy) maths involved to find equivalent ways to wr= ite> the initial ICA assumptions - like minimizing the maximum likelihood= > contrast function. Most of the different ICA implementations start> fro= m maximum likelihood (or similar criteria) and differ in the> strategies = they've used to actually allow calculation. (Does it make> sense?)...>> T= hat said, apart from the link given by Taylan and Linda, Tony Bell> has i= nteresting demos and papers (I'd recommend Bell and Sejnowski,> 1995) ava= ilable on his website:> http://cnl.salk.edu/~tony/ica.html> You might als= o find JADE (see Cardoso, 1999) or Fast-ICA (see> Hyvarinen & Oja, 2000) = Matlab implementations interesting "to play> with" :>> http://www.tsi.ens= t.fr/icacentral/algos.html> http://www.cs.helsinki.fi/u/ahyvarin/papers/f= astica.shtml>> Best regards,> Idrick>> Quoting Vijaykumar Peddinti :>>> D= ear All,>>>> This is might be a silly question. But I am trying to do a s= mall>> project on separating signals using Independent Component Analysis= >> (ICA) Maximum likelihood approach for my class. So far the articles>> = I found leave me with more questions than answers. I have never done>> an= ything with ICA before. I am lost in the process of finding the>> informa= tion.>>>> It would be of great help if anybody could point me in the corr= ect>> direction or send me an basic Introductory paper/tutorial (if>> pos= sible) on ICA (Maximum likelihood approach). I am not trying>> anything f= ancy, I am trying to get a basic simulation working in>> MATLAB.>>>> Than= k you very much.>>>> Best Regards,>> Vijay>>> ----=_vm_0011_W558620254_11012_1271105743 Content-Type: text/html; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable <html>Doesn't that kind of negate the whole purpose of ICA? It's su= pposed to find the dimensions that PCA might miss. But if you do a = PCA first, all you will get out are subsets of the dimensions the PCA fin= ds in the first place.<br><div><font face=3D"Verdana" size=3D"2"> </= font></div> Brian Gygi, Ph.D. <br> Speech and Hearing Research <br> Veterans Affairs Northern California Health Care System <br> 150 Muir Road <br> Martinez, CA 94553 <br> (925) 372-2000 x5653<div><font color=3D"#0000ff" face=3D"Verdana" size=3D= "2"></font> </div> <blockquote style=3D"border-left: 2px solid rgb(0, 0, 255); padding-left:= 5px; margin-left: 5px; margin-right: 0px;"><font face=3D"Tahoma" size=3D= "2">-----Original Message-----<br><b>From:</b> Linda Seltzer [mailto:lsel= tzer@xxxxxxxx<br><b>Sent:</b> Friday, April 9, 2010 09:47 PM<b= r><b>To:</b> AUDITORY@xxxxxxxx<br><b>Subject:</b> Re: Request for = information on ICA-Maximum likelihood approach<br><br></font>I should add= that when ICA is used in fMRI (e.g. the use of fast ICA in Brain Voyager), PCA is used first for dimensionality reduction. Linda Seltzer > Hi Vijay, > > I'm not sure you could start a Matlab script from scratch to do ICA > with Maximum likelihood approach. As far as I've gathered, there's > quite a lot of (heavy) maths involved to find equivalent ways to wri= te > the initial ICA assumptions - like minimizing the maximum likelihood= > contrast function. Most of the different ICA implementations start > from maximum likelihood (or similar criteria) and differ in the > strategies they've used to actually allow calculation. (Does it make= > sense?)... > > That said, apart from the link given by Taylan and Linda, Tony Bell > has interesting demos and papers (I'd recommend Bell and Sejnowski, > 1995) available on his website: > http://cnl.salk.edu/~tony/ica.html > You might also find JADE (see Cardoso, 1999) or Fast-ICA (see > Hyvarinen & Oja, 2000) Matlab implementations interesting "to pl= ay > with" : > > http://www.tsi.enst.fr/icacentral/algos.html > http://www.cs.helsinki.fi/u/ahyvarin/papers/fastica.shtml > > Best regards, > Idrick > > Quoting Vijaykumar Peddinti <vijay@xxxxxxxx>: > >> Dear All, >> >> This is might be a silly question. But I am trying to do a small= >> project on separating signals using Independent Component Analys= is >> (ICA) Maximum likelihood approach for my class. So far the artic= les >> I found leave me with more questions than answers. I have never = done >> anything with ICA before. I am lost in the process of finding th= e >> information. >> >> It would be of great help if anybody could point me in the corre= ct >> direction or send me an basic Introductory paper/tutorial (if >> possible) on ICA (Maximum likelihood approach). I am not trying >> anything fancy, I am trying to get a basic simulation working in= >> MATLAB. >> >> Thank you very much. >> >> Best Regards, >> Vijay >> > </vijay@xxxxxxxx></blockquote></html> ----=_vm_0011_W558620254_11012_1271105743--