Pattern Recognition -- Any Good?

[ph1l]

Any books I can buy/read to learn about multidimensional pattern recognition, tensor of data, convolution of Fourier transform and long termism?

Here is something about multi-dimensional pattern recogition:
"Multi-Dimensional Dynamic Time Warping for Gesture Recognition," G.A. ten Holt, M.J.T. Reinders, E.A. Hendriks, 2007

 

[TommyR]

when it comes to neural nets required essential reading is: functional proxies, second dual spaces, euler-lagrange geodesic characterisics, hamiltonians and the generator of the energy kernal, vector proxies for differential operators and familiarity with basic identities for div grad and curl on spaces with positive ricci curvature, tangent spaces for non-smooth manifolds.

do not read books on neural nets saying: nets are a childrens picture supposed to look like a brain capable of appromiating a very small number of functions. layers are pointless and the set of fucntions possible is the same as in one layer, consciousness is unlikely to appear at layer 10. it is unclear why adding parameters at o(data) is good.

they definitely are useful: excellent if you want to invert a deterministic function and you cannot analytically. so at bs vol/delta/greeks from price and all permuations for example they will be better than any series you can find and very fast so i use them there.

 

[ironchef]

i will answer any questions on the above. i would not recommend a book . a fourier transform is a long termist solution however if you are a beginner you can for your first project use it on the vix, it can do this very well, and there are good reasons why it should be quite good. you can use the simple one no need to decay these amplitudes because the function we are appoximating goes on forever and are extremely stable. word of advice: there are some very long term periods so use millions of terms. there are loads of programs you can buy to perform it. The idea is very simple: cos and sin are a so called orthogonal basis. as such we can write a function as a power series wrt to this basis and the coefficients for a given frequency are unique. Skill required:integrate a dummy on the unit disk, find the zeroes of the sin function, realise that integrating between -pi and pi is an inner product for this hoax.

when it comes to neural nets required essential reading is: functional proxies, second dual spaces, euler-lagrange geodesic characterisics, hamiltonians and the generator of the energy kernal, vector proxies for differential operators and familiarity with basic identities for div grad and curl on spaces with positive ricci curvature, tangent spaces for non-smooth manifolds.

do not read books on neural nets saying: nets are a childrens picture supposed to look like a brain capable of appromiating a very small number of functions. layers are pointless and the set of fucntions possible is the same as in one layer, consciousness is unlikely to appear at layer 10. it is unclear why adding parameters at o(data) is good.

they definitely are useful: excellent if you want to invert a deterministic function and you cannot analytically. so at bs vol/delta/greeks from price and all permuations for example they will be better than any series you can find and very fast so i use them there.

I am going to search Coursera and see if they offer a class on these topics.

 

[ironchef]

Here is something about multi-dimensional pattern recogition:
"Multi-Dimensional Dynamic Time Warping for Gesture Recognition," G.A. ten Holt, M.J.T. Reinders, E.A. Hendriks, 2007

Now you introduce another new term: Time warp? Like in General Relativity or Star Trek?

 

[TommyR]

i will answer any questions on the above. i would not recommend a book . a fourier transform is a long termist solution however if you are a beginner you can for your first project use it on the vix, it can do this very well, and there are good reasons why it should be quite good. you can use the simple one no need to decay these amplitudes because the function we are appoximating goes on forever and are extremely stable. word of advice: there are some very long term periods so use millions of terms. there are loads of programs you can buy to perform it. The idea is very simple: cos and sin are a so called orthogonal basis. as such we can write a function as a power series wrt to this basis and the coefficients for a given frequency are unique. Skill required:integrate a dummy on the unit disk, find the zeroes of the sin function, realise that integrating between -pi and pi is an inner product for this hoax.

After you can predict the vix perfectly do the vix futures by expiry given the level. You can use sums of f*exp-a(VIX-k)^2 to loaclise the behaviour of f = the termstructure near a value k of the vix and decaying rapidly away. there are so few values the vix takes this will be easy to do without transforms. Then you can see how much money selling the 1st and waiting a week will make.

 

[Kevin Schmit]

Any books I can buy/read to learn about multidimensional pattern recognition, tensor of data, convolution of Fourier transform and long termism?

Don't tell me, let me guess...

TommyR has somehow got access to an Internet-connected computer and is posting again. Usually the hospital staff tries to prevent this, but he somehow got around them.

Am I right? Checking in private browser mode....

Yes, of course. Tensors with 1e+17 individual elements processed routinely. That's plausible for sure!

Cubic spleens!

 

[Overnight]

when it comes to neural nets required essential reading is: functional proxies, second dual spaces, euler-lagrange geodesic characterisics, hamiltonians and the generator of the energy kernal, vector proxies for differential operators and familiarity with basic identities for div grad and curl on spaces with positive ricci curvature, tangent spaces for non-smooth manifolds.

do not read books on neural nets saying: nets are a childrens picture supposed to look like a brain capable of appromiating a very small number of functions. layers are pointless and the set of fucntions possible is the same as in one layer, consciousness is unlikely to appear at layer 10. it is unclear why adding parameters at o(data) is good.

they definitely are useful: excellent if you want to invert a deterministic function and you cannot analytically. so at bs vol/delta/greeks from price and all permuations for example they will be better than any series you can find and very fast so i use them there.

Something tells me you'd like this book...

 

[ph1l]

Now you introduce another new term: Time warp? Like in General Relativity or Star Trek?

Dynamic Time Warping is a fancy name for a way to calculate the distance between two time series.
See https://databricks.com/blog/2019/04/30/understanding-dynamic-time-warping.html

 

[ironchef]

Dynamic Time Warping is a fancy name for a way to calculate the distance between two time series.
See https://databricks.com/blog/2019/04/30/understanding-dynamic-time-warping.html

Thank you for the explanation and thank you for the link. Very helpful.

 

[themickey]

......realise that integrating between -pi and pi is an inner product for this hoax.

Soundsaboudright.