Does machine learning really involve data?

Many definitions of machine learning start by proclaiming it uses data, to learn. I want to challenge this, or remind us where the term originally came from and consider why the meaning has shifted.

For a long time machine learning seemed to be a new technology, but I notice we're starting to say AI and machine learning interchangeably. Job postings often sneak the word scientist in there too. What is a data scientist? What do any of these words mean?

Current trends often come with an air of mystery. I suspect a lot of data science roles involve data entry, in order to clean input data. Not as appealing as the headline role suggests. Several day to day techniques being described as machine learning  could also be described as statistics. In fact, look at the table of contents of a statistics book, such as An Introduction to Statistical Learning. Look at a small selection of the topics:

• accuracy
• k-means clustering
• making predictions
• cross-validation
• support vector machines, SVM
• principal component analysis, PCA

Most, if not all, of these topics are covered in an average machine learning course and included in ML software packages. Yet statistics doesn't sound as exciting as machine learning, to many people.

Wikipedia defines statistics as "a branch of mathematics dealing with data collection, organization, analysis, interpretation and presentation." No mention of learning, though each of these activities form an essential part of data science. The article goes on to discuss descriptive and inferential statistics. Inference involves making predictions: many people use the term machine learning to mean the very same. Can you spot patterns in purchases automatically and suggest other items a customer might be interested in? Can you detect unusual or anomalous behaviour, indicating fraud or similar? Again, these are now labelled as AI or machine learning, but usually rely on well established statistical techniques. Admittedly, today's faster machines mean number crunching can happen quickly. This has contributed to the resurgence of machine learning.

Many problem solving algorithms are not about numbers. Some techniques, such as evolutionary computing, including genetic algorithms, don't fit comfortably into a data-driven view of learning. Do these methods count as machine learning? I'll leave that for you to think about. My book explores genetic algorithms and several other areas that do not need numbers to learn.

Arthur Samuel came out with the phrase "machine learning", by which he meant something along the lines of a "field of study that gives computers the ability to learn without being explicitly programmed." The abstract of his 1959 paper, "Some studies in machine learning using the game of checkers" states,

Two machine-learning procedures have been investigated in some detail using the game of checkers. Enough work has been done to verify the fact that a computer can be programmed so that it will learn to play a better game of checkers than can be played by the person who wrote the program. Furthermore, it can learn to do this in a remarkably short period of time (8 or 10 hours of machine-playing time) when given only the rules of the game, a sense of direction, and a redundant and incomplete list of parameters which are thought to have something to do with the game, but whose correct signs and relative weights are unknown and unspecified. The principles of machine learning verified by these experiments are, of course, applicable to many other situations.

AI and machine learning are both very old terms. I think they encompass a much broader field than data analysis. As a final thought, Turing designed an algorithm to play chess. In effect, he was trying to make an artificial brain, before the term AI was invented or computers, in their modern sense, existed.

I think machine learning is much broader than investigating data. Its history involves attempting to get computers to learn, and specifically to learn to play games. Let the games continue.

Read my book and see what you think.

I wrote a book about Genetic algorithms and Machine learning

I've written a book pulling together some of my previous talks showing how to code your way out of a paper bag using a variety of machine learning techniques and models, including genetic algorithms.
It available at Amazon and you can download free excerpts from the publishers website.

The sales figures show I've sold over 1,000 copies already. I'm going through the copy edits at the moment. I can't wait to see the actual paper book.

I've sold over 3,000 copies, which is phenomenal. 

Thank you to everyone at ACCU who helped and encouraged me while I wrote this.

I will be giving some talks at conferences and hopefully some meetups based on ideas in some of the chapters in 2019.

Watch this space.

Update: here are some of my talks on YouTube including diffusion, simulated annealing, cellular automata and genetic algorithms. 

Gitlab certificates

On Ubuntu, cloning a repo from a machine you don't have a certificate for will give the error:

fatal: unable to access 'https://servername': server certificate verification failed. CAFuile /etc/ssl/certs/your_filename CRLfile: None

You can work around this by tell git clone not to use the certificate e.g.

git config --system http.sslverify false

which is asking for trouble. However you can install the certificate, so you don't need to keep doing this. 

Using an answer here: https://stackoverflow.com/questions/21181231/server-certificate-verification-failed-cafile-etc-ssl-certs-ca-certificates-c  looks to have worked, by trying things one step at a time:

hostname=gitlab.city.ac.uk

port=443

trust_cert_file_location=`curl-config --ca`

sudo bash -c "echo -n | openssl s_client -showcerts -connect $hostname:$port \

    2>/dev/null  | sed -ne '/-BEGIN CERTIFICATE-/,/-END CERTIFICATE-/p'  \

    >> $trust_cert_file_location"

I did try this first – so errors don’t end up in dev null:

openssl s_client -showcerts -connect $hostname:$port

Also, I first got the error sed: unrecognised option '--ca'

It took a moment to realise the --ca came from curl-config, which I needed to install.

Windows batch files

I've been writing a batch file to run some mathematical models over a set of inputs.
The models are software reliability growth models, described here.

We are using

• du: Duane
• go: Goel and Okumto
• jm: Jelinski and Moranda
• kl: Keiller and Littlewood
• lm: Littlewood model
• lnhpp: Littlewood non-homogeneous Poisson process
• lv: Littlewood and Verrall
• mo: Musa and Okumoto

Littlewood appears many times: he founded the group where I currently work. 

So, far too much background. I have one executable for each model, after making a make file; yet another story. And a folder of input files, named as f3[some dataset]du.dat, f3[some dataset]go.dat,... f3[some dataset]mo.dat. I also have some corresponding output files someone else produced a while ago, so in theory I can check I get the same numbers.I don't but that's going to be yet another story.

You can also use the original file and generated file to recalibrate, giving yet another file. Which I have previously generated results from. Which also don't match. 

I wanted to be able to run this on Ubuntu and Windows, and managed to make a bash script easily enough. Then I tried to make a Windows batch file to do the same thing. I'll just put my final result here, and point out the things I tripped up on several times.


ECHO OFF
setlocal EnableDelayedExpansion
setlocal 

for %%m in (du go jm kl lm lnhpp lv mo) do (
  echo %%m
  for %%f in (*_%%m.dat) do (
    echo %%~nf
    set var=%%~nf
    echo var: !var!
    set var=!var:~2!
    echo var now: !var!

    swrelpred\%%m.exe %%~nf.dat "f4!var!"
    swrelpred\%%mcal.exe %%~nf.dat "f4!var!" "f9!var!"
  )
)

1. First, turn the echo off because there's way too nosie otherwise.
2. Next, enable delayed expansion, otherwise things in blocks get expanded on sight and therefore don't change in the loop: "Delayed expansion causes variables delimited by exclamation marks (!) to be evaluated on execution"  from stack exchanges' Superuser site
3. Corollary: Use ! in the variables in the block not % for delayed expansion.
4.  But we're getting ahead of ourselves. The setlocal at the top means I don't set the variables back at my prompt. Without this, as I changed my script to fix mistakes it did something different between two runs, since a variable I had previously set might end up being empty when I broke stuff.
5. "Echo is off" spewed to the prompt means I was trying to echo empty variables, so the var: etc tells me which line something is coming from.
6. !var:~2! gives me everything from the second character, so I could drop the f3 at the start of the filename and make f4 and f9 files to try a diff on afterwards. Again pling for delayed expansion.




I suspect I could improve this, but it's six importnat things to remember another time.

Writing this in Python might have been easier. Or perhaps I should learn Powrshell one day.

Call a dll from Python

Letting VS2015 make a dll called SomeDLL for me with these implementations

// SomeDll.cpp : 
// Defines the exported functions for the DLL application.
//

#include "stdafx.h"
#include "SomeDll.h"


// This is an example of an exported variable
SOMEDLL_API int nSomeDll=0;

// This is an example of an exported function.
SOMEDLL_API int fnSomeDll(void)
{
    return 42;
}

// This is the constructor of a class that has been exported.
// see SomeDll.h for the class definition
CSomeDll::CSomeDll()
{
    return;
}
 
I then make a python script, using ctypes and loading the dll, using os to find it:

import os
import ctypes

os.chdir("C:\\Users\\sbkg525\\src\\SomeDll\\Debug")
SomeDll = ctypes.WinDLL("SomeDll.dll")

I can either use attributes of the library or use protoypes. I tried protoypes first.The function returns an int and takes no parameters:

proto = ctypes.WINFUNCTYPE(ctypes.c_int)
params = ()

answer = proto(("fnSomeDll", SomeDll), params)

Unfortunately this says

AttributeError: function 'fnSomeDll' not found

because C++ is name mangled. extern "C" FTW another time; for now

link.exe /dump /exports Debug\SomeDll.dll

          1    0 0001114F ??0CSomeDll@@QAE@XZ =
                  @ILT+330(??0CSomeDll@@QAE@XZ)
          2    1 00011244 ??4CSomeDll@@QAEAAV0@$$QAV0@@Z =
                  @ILT+575(??4CSomeDll@@QAEAAV0@$$QAV0@@Z)
          3    2 0001100A ??4CSomeDll@@QAEAAV0@ABV0@@Z =
                  @ILT+5(??4CSomeDll@@QAEAAV0@ABV0@@Z)
          4    3 0001111D ?fnSomeDll@@YAHXZ =
                  @ILT+280(?fnSomeDll@@YAHXZ)
          5    4 00018138 ?nSomeDll@@3HA =
                  ?nSomeDll@@3HA (int nSomeDll)

Looks like we want 4;

answer = proto(("?fnSomeDll@@YAHXZ", SomeDll), params)

print answer
>>> <WinFunctionType object at 0x0248A7B0>


Of course. It's a function. Let's call the function

print answer()
>>> 42

Done. I'll try attributes next. And functions which take parameters. Later, I'll look at other calling conventions.

First, the docs says, "foreign functions can be accessed as attributes of loaded shared libraries.
Given

extern "C"
{
    SOMEDLL_API void hello_world()
    {
        //now way of telling this has worked
    }
}

we call it like this

lib = ctypes.cdll.LoadLibrary('SomeDll.dll')
lib.hello_world()

It seems to assume a return type of int. For example,

extern "C"
{
    SOMEDLL_API double some_number()
    {
        return 1.01;
    }
}

called as follows

lib = ctypes.cdll.LoadLibrary('SomeDll.dll')
lib.hello_world()
val = lib.some_number()
print val, type(val)
Gives

-858993460, <type 'int'>


We need to specify the return type, since it's not an int:



lib.some_number.restype = ctypes.c_double
val = lib.some_number()
print val, type(val)

then we get what we want

1.01 <type 'float'>

We need to do likewise for parameters

extern "C"
{
    SOMEDLL_API double add_numbers(double x, double y)
    {
        return x + y;
    }
}

If we just try to call it with some floats we get an error

    total = lib.add_numbers(10.5, 25.7)
ctypes.ArgumentError: argument 1: <type 'exceptions.TypeError'>: Don't know how to convert parameter 1

Stating the parameter type fixes this

lib.add_numbers.restype = ctypes.c_double
lib.add_numbers.argtypes = [ctypes.c_double, ctypes.c_double]
total = lib.add_numbers(10.5, 25.7)
print total, type(total)

36.2 <type 'float'>

Just one starter thought on strings. Returning a const char * seems to give a str in python

SOMEDLL_API const char * speak()
{
    return "Hello";
} 

called like this



lib.speak.restype = ctypes.c_char_p
print lib.speak()

says "Hello"; using parameters as strings and particularly as refrences that can be changed needs some investigation.

Does a similar approach using attributes of the loaded library work for non-extern "C" functions? Can I use the proto approach on the extern "C" functions?

Let's see...

Codes of Conduct

There's been yet another thread about code of conducts at conferences on twitter, and I wanted to compare them to conditions of sale for gig or festival tickets. Here's an example:

Let's consider a sample:
1. I can't turn up before it starts? Why even say that? Oh, and 2. I have to leave when it's over. THIS IS PC GONE MAD. OK, not PC. But why do you need to spell this out? It's like putting a precondition of a string has to terminate with a null for you to call strlen.
3. Yeah, rulez. Whatever. Surprised you didn't say it's illegal to break the law. What is this?
Blahblahblah.
9. Yeah, fair enough. But what about flame throwers?
No, no, no...
16. Does that mean I can bring *Real* weapons?
Blahblahblah.

25. There might be swear words?! Why tell me that?

Or maybe some people will bring kids with them and that might need a sane conversation about a time and a place for certain behaviours. I know the bands, I know what kind of music to expect. I know what I'm walking in to.

I also know there's a clearly marked welfare tent on site just in case.

And wardens on the campsite wearing obvious vests.

And hundreds of programmers listening to their favourite rock stars.
(OK, not all the punters are programmers, but many are).

Why did I ever go to a field full of metallers? (Knowing I'd probably be one of the few women there). I went with friends, which made me feel safer. I wanted to listen to bands I knew, and discover new music. If I'd gone by myself, knowing in advance about wardens and the welfare tent would have made me feel OK. Hey, the conditions of sale show the organisers have thought about things that might go wrong or concern people, and that makes me feel safer.

Do they make people think, "This is PC gone mad? If you need to tell me not to set fire to myself and drink myself to death then you are insulting me and I want nothing to do with this?"

Not by the looks of the number of people who turn up. And there have been more women and kids recently which is great.

What are codes of conduct for? Quite frankly me. But not just me. They are not there to tell you how to behave because the organisers think white guys don't know how to conduct themselves, and that all such guys are potential rapists or murders. In some ways, the logic of stating a CoC is pointless because we know how to be nice taken to extreme could mean laws are pointless. Why say "You aren't allowed to commit murder?" Surely that doesn't need pointing out? And yet most countries have such a law.

Conditions of sale and code of conducts aren't laws, but they give a shared statement of expectation and make me feel OK about going to conferences/talks/festivals/gigs alone. And then I meet loads of new people and discover new music. And I can't wait til Bloodstock. Or the next tech talk/conference I go to.

Elastic stack - RTFM

I tried to setup ELK (well, just elasticsearch and kibana initially), with a view to monitoring a network.

Having tried to read the documentation for an older version than I'd downloaded and furthermore one for *Nix when I'm using Windows, I eventually restarted at the "Learn" pages on https://www.elastic.co/

There are a lot of links in there, and it's easy to get lost, but it is very well written.

This is my executive summary of what I think I did.

First, download the zip of kibana and elasticsearch.

From the bin directory for elasticsearch, run elasticsearch.bat file, or run service install then service run. If you run the batch file it will spew logs to the console, as well as a log file (in the logs folder). You can tail the file if you choose to run it as a service. Either works.

If you then open http://localhost:9200/ in a suitable browser you should see something like this:

{
  "name" : "Barbarus",
  "cluster_name" : "elasticsearch",
  "cluster_uuid" : "bE-p5dLXQ_69o0FWQqsObw",
  "version" : {
    "number" : "2.4.1",
    "build_hash" : "c67dc32e24162035d18d6fe1e952c4cbcbe79d16",
    "build_timestamp" : "2016-09-27T18:57:55Z",
    "build_snapshot" : false,
    "lucene_version" : "5.5.2"
  },
  "tagline" : "You Know, for Search"
}

The name is a randomly assigned Marvel character. You can configure all of this, but don't need to just to get something up and running to explore. kibana will expect elasticsearch to be on port 9200, but again that is configurable. I am getting ahead of myself though.

Second, unzip kibana, and run the batch file kibana.bat in the bin directory. This will witter to itself. This starts a webserver, on port 5601 (again configurable, but this by default): so open http://localhost:5601 in your browser.

kibana wants an "index" (way to find data), so we need to get some into elasticsearch: the first page will say "Configure an index pattern". This blog has a good walk through of kibana (so do the official docs).

All of the official docs tell you to use curl to add (or CRUD) data in elasticsearch, for example

curl -XPUT 'localhost:9200/customer/external/1?pretty' -d '
{
  "name": "John Doe"
}'

NEVER try that from a Windows prompt, even if you have a curl library installed. You need to escape out the quote, and even then I had trouble. You can put the data (-d part) in a file instead and use @, but it's not worth it.

Python to the rescue. And Requests:HTTP for Humans

pip install requests

to the rescue.

Now I can run the instructions in Python instead of shouting at a cmd prompt.

import requests
r = requests.get('http://localhost:9200/_cat/health?v')
r.text


Simple. The text shows me the response. There is a status code property too. And other gooides. The the manual. For this simple get command you could just point your browser at localhost:9200/_cat/health?v


Don't worry if the status is yellow - this just means you only have omne node so it can't replicate in cause of disaster.

Notice the transport, http:// at the start. If you forget this, you'll get an error like
>>> r = requests.put('localhost:9200/customer/external/1?pretty', json={"name": "John Doe"})
...
    raise InvalidSchema("No connection adapters were found for '%s'" % url) requests.exceptions.InvalidSchema: No connection adapters were found for 'localhost:9200/customer/external/1?pretty'


Now we can put in some data.

First make an index (elastic might add this if you try to put data under a non-existent index). We will then be able to point kibana at that index - I mentioned kibana wanted an index earlier.
r = requests.put('http://localhost:9200/customer?pretty')


Right, now we want some data.
>>> payload = {'name': 'John Doe'}
>>> r = requests.post('http://localhost:9200/customer/external/1?pretty', json=payload)

If you point your browser at localhost:9200/customer/external/1?pretty you (should) then see the data you created. We gave it an id of 1, but it will be automatically assigned a unique id if we left that off.

We can use requests.delete to delete, and requests.post to update:
 >>> r = requests.post('http://localhost:9200/customer/external/1/_update', \
 json={ "doc" : {"name" : "Jane Doe"}})

Now, this small record set won't be much use to us. The docs have a link to some json data. I downloaded some ficticious account data. SO to the rescue for uploading the file:


>>> with open('accounts.json', 'rb') as payload:
...   headers = {'content-type': 'application/x-www-form-urlencoded'}
...   r = requests.post('http://localhost:9200/bank/account/_bulk?pretty', \ 
              data=payload,  verify=False, headers=headers)
...

>>> r = requests.get('http://localhost:9200/bank/_search?q=*&pretty')
>>> r.json()This is equivalent to using

>>> r = requests.post('http://localhost:9200/bank/_search?pretty', \
      json={"query" : {"match_all": {}}})
 i.e. instead of q=* in the uri we have put it in the rest body.



Either way, you now have some data which you can point kibana at. In kibana, the discover tab allows you to view the data by clicking through fields. The visualise tab allows you to set up graphs. What wasn't immeditely apparent was once you have selected your buckets, fields and so forth, you need to press the green "play" button by the "options" to make it render your visualisation. And finally, I got a pie chart of the data.  I now need to point it at some real data.