FPGAs, hardware

Moore's law still applies for single-core FPGAs.
- Millions of floating point operations per second (MFLOPS): 10 years ago 1, 5 years ago 5, 2 years ago 50, now 5000.
Don't consider wiring diagrams any more - FPGAs change so fast that tailoring a design to one particular FGPA is a waste of time.
Don't code in VHDL at all, even behavioural. Too low level.
Compilers are now capable of compiling C down to FPGAs, but used to make sub-optimal decisions.
- FPGAs come with multiple floating point chips, multiple DSPs, and compiler not only needs to transform C but to schedule instructions on multiple chips, taking pipelines into account.
FPGAs just recently support 10 GbE.
Many tricks involved in making floating point operations faster
- Align mantissa and exponents
- Add small numbers first building up to large. Because adding large first means precision is impossible to maintain.

Latent Dirichlet Allocation (LDA)

Generative model of topics within a body of text.
Supervised.
Topics are random variables generating tokens.
Refer to NLP notes for refresher on generative models.

Clustering

Markov Cluster Algorithm, http://micans.org/mcl/
- Unsupervised, very fast, binary blob available.

Don't Use This Code

Lots of fun and insane hacks: http://seriously.dontusethiscode.com/

IPython 2 Widgets

Incredible! Coming out April 2014.

General notes

Many are PhDs, many are hard core into e.g. MATLAB
Want to emphasise how awesome Python is.
Almost everyone prefers to use IPython Notebook, it's common knowledge.
- For internal presentation, analyses, keeping notes and following presentations.
Cython has tremendous amount of mindshare, particular with NumPy. Almost standard.
Interesting that there are contrasting talks
- "Databases for Scientists" and "Scaling SQL" both contrast to Hive and Pig, wrt MapReduce.
- "Getting it out there" (JavaScript for webviz) contrasts with Bokeh (Python to auto-generate JavaScript) and IPython Notebook.



In [3]:

    
def f(x):
    return (x-1) ** 2

import scipy.optimize
scipy.optimize.minimize_scalar(f)









    Out[3]:





  fun: 0.0
 nfev: 5
  nit: 4
    x: 1.0



In [3]:

    
from IPython.core.display import HTML
def css_styling():
    styles = open("styles/custom.css", "r").read()
    return HTML(styles)
css_styling()









    Out[3]:



In [2]:

    
%autosave 10









    














    



Autosaving every 10 seconds