Machine Learning

[avatar user=”malm” size=”small” align=”left” link=”file” /]

I recently completed the Stanford University Machine Learning online course available via Coursera and taught by Professor Andrew Ng. A companion post I’ve written surveys the course contents and provides some personal perspectives on the experience.  It includes a worked example of a data compression technique called Principal Component Analysis. The course exercises are conducted using the MATLAB environment whose origins go all the way back to Alan Turing, creator of the eponymous test for determining a machine’s ability to exhibit intelligent behaviour.

Machine Learning can be considered as a key subset within the broader field of Artificial Intelligence which this week marked a major milestone with the news that Deep Mind’s AlphaGo system had beaten Europe’s best Go player at the legendarily difficult game 5 times in a row.  Perhaps the most important aspect of this amazing feat lies in the cunning combination of AI and Machine Learning techniques being used to achieve it.  They are more sophisticated versions of the same concepts covered in online courses – notably neural nets and reinforcement learning:

After training on 30 million human moves, a DeepMind neural net could predict the next human move about 57 percent of the time—an impressive number (the previous record was 44 percent). Then Hassabis and team matched this neural net against slightly different versions of itself through what’s called reinforcement learning. Essentially, as the neural nets play each other, the system tracks which move brings the most reward

It’s an achievement with profound implications that have resonated around the tech world this week:

it’s conceivable that AlphaGo could play the game at a level that no human could ever attain. Hassabis said that games like Go represent perfect stepping stones for researchers to hit on the pathway to potentially creating something that could be considered a true artificial intelligence system. … Google has cracked what has long been held as one of the “grand challenges” in AI research, giving credence to the idea that true, general-purpose, artificial intelligence may be possible.

The Kernel’s Cyborg issue recently reminded readers that Roy Batty the leader of the runaway replicants in Bladerunner, was “incepted” in Philip K. Dick’s storyline on Friday January 8th 2016.

Go notwithstanding, we’re a long way from the fictional Nexus 6 but the date is an appropriate signifier that we are already living in a future where AI is big and going to get bigger.   Data from the US job market suggests demand is already beginning to eclipse other skillsets and that “AI skills are commanding the premium that big data know how used to“.

Artificial Intelligence on the loose

• As many as 40 countries are working on killer robots and there is as yet no law outlining how their development should be policed. It seems inevitable that military application of ‘warbots’ will be a primary focus.  It’s not a stretch to image deep neural nets with billions of hours of Call of Duty training under their belt being placed inside of the case of a mechanised robot a few years from now.  Not quite Roy Batty but it’s enough to make you think again about those cute videos of Atlas doing the housework. And wonder what developers might create with the CNTK deep learning toolkit Microsoft just released. In time perhaps BSOD  will have a very different meaning.

• Researchers in the US can now “track perception at the speed of thought” using an EEG rig and lots of real-time data analysis:

• Meanwhile in Europe other researchers are using memristors to create a physical neural network.

• No wonder Werner Herzog has seen fit to choose the tech world as the focus for his next film Lo and Behold:

“for those looking for a ride through our modern technological world, or indeed a preview of what is to come, this is it.”

• Here’s something utterly inconceivable 35 years ago when the craze was at its peak.  The world’s fastest Rubik’s Cube robot built from commodity electronics components takes just over a second to solve a cube:

Facebook and its focus

• Facebook’s sudden decision to wind down Parse its BaaS (Backend as a Service) platform must have come as a shock to the many developers who rely upon it.  Arguably it shouldn’t have done given it’s a canonical example of “platform risk” in action.  In what is something of a bitter irony, the suggested solution is for developers to host their own version of the Parse Server. Avoiding having to deal with devops was the whole point of them using Parse in the first place:

We are committed to maintaining the backend service during the sunset period, and are providing several tools to help migrate applications to other services.  First, we’re releasing a database migration tool that lets you migrate data from your Parse app to any MongoDB database. During this migration, the Parse API will continue to operate as usual based on your new database, so this can happen without downtime. Second, we’re releasing the open source Parse Server, which lets you run most of the Parse API from your own Node.js server. Once you have your data in your own database, Parse Server lets you keep your application running without major changes in the client-side code. 

• The decision to close Parse is not so much about economics given Facebook’s recent glowing results.  It’s more about focus.  And Zuckerberg’s focus is on other areas.  A key one is Facebook’s hugely ambitious internet.org mission to connect the world.  Wired’s inside look at internet.org makes for fascinating reading.

• Another key area of focus is Facebook Messenger which stands at the cusp of the next major evolution in user software:

The evolution of software pic.twitter.com/ZnAMMxbZP1

— Andreas Constantinou (@andreascon) January 26, 2016

• Wired suggest many businesses are already leveraging Messenger as a customer service and complaint channel. Facebook clearly have a bigger goal in mind though:

Beyond customer service, it wants users to migrate to chat apps—and in particular, Messenger and WhatsApp, which it also owns—for every aspect of commerce, including discussing and tracking purchases and even buying things directly.

• AI will be a key part of their Messenger play.  This post provides an inside view a recent 24-hour Facebook AI hackathon offering a fascinating glimpse of their work culture and focus in action.

• However, there remain plenty of Facebook detractors.  The Kernel shone a baleful light on “Facebook anxiety” comparing the experience of flicking through the curated feeds of friends to that of being an inmate of the infamous Panopticon:

Researchers linked a high number of Facebook friends with feeling burdened or stressed out by the site.   our desire to lurk and unhealthily peer into the lives of those around us is built into the experience itself. …  the site [is like] an 18th century “panopticon”—a type of prison that allowed inmates to be viewed by guards at all times. “We know what all the prisoners are showing us from their jail cells,”

• And all those friends might not even have a discernable upside if you really did need support according to recent research:

On average, folks had 150 followers but said that they could only count on 4.1 of them during an “emotional crisis,” and only 13.6 ever express sympathy

• The source research is authored by Robin Dunbar.  In it he makes the assertion that being online does not change the fundamental social dynamics of his eponymous famous Number:

“as originally proposed by the social brain hypothesis, there is a cognitive constraint on the size of social networks that even the communication advantages of online media are unable to overcome. In practical terms, it may reflect the fact that real (as opposed to casual) relationships require at least occasional face-to-face interaction to maintain them.”

• Facebook detraction is still very much a fringe activity though. Most people use it without any real concern about the downside and the brand marches confidently on to its future.  It’s worth comparing and contrasting Facebook’s clarity of focus with Twitter’s recent issues.  The company seems to have struggled with product strategy and problem users as the scar tissue of various changes to its rules testify.  Joshua Topolsky in the New Yorker argues that today it stands closer to the precipice that ever before:

In Facebook’s case, the company has demonstrated its mastery of product focus and long-term commitment to user experience. … Meanwhile, a series of mediocre product changes at Twitter (such as the much-hyped but ultimately confusing Moments feature), a stagnant user base, and a massive executive brain drain have called into question whether Twitter can survive as a business.

Mobile Crossroads

• Apple are predicting Q2 figures that “would mark the first year-over-year decline in revenue for the company in years” and may signal that we have reached ‘peak iPhone’.  Meanwhile Jon Gruber dissects Donald Trump’s campaign pledge to “get Apple to start building their damn computers and things in this country” explaining that it is unworkable because the US simply doesn’t have the skill at scale to do it.  The only way it could is by automating the entire process but that by definition wouldn’t create any manufacturing jobs:

The U.S. can’t compete with China on wages. It can’t compete on the size of the labor force. China has had a decades-long push in its education system to train these workers; the U.S. has not. And the U.S. doesn’t have the facilities or the proximity to the Asian component manufacturers.

• In the world of Android, Samsung are being sued in the Netherlands by a consumer pressure group for a ‘lackadaisical approach‘ to Android updates. And in an interesting move, Google have announced a partnership with Movidius whose specialised Myriad “visual processing unit” may be able to support localised deep learning computation on the handset.

• Windows Phone on the other hand looks to be a dead platform walking:

“It won’t stop Microsoft producing a few handsets every year as a vanity project, but for everyone else it’s the end of the line.

Sorry, Windows Phone. 110m lifetime sales – 4.5bn iOS & Android phones sold in the same period pic.twitter.com/CO03XWhYJg

— Benedict Evans (@BenedictEvans) January 28, 2016

• Any remaining die-hards can console themselves with $150 windows on a stick or, implausibly, Windows 95 running entirely in the browser.  I just used it to play Solitaire for the first time in years!

The Internet of Things and Cloud

• The Orange Pi One is like the Raspberry Pi Zero but faster.

• Good ContentLoop article outlining the key concepts behind microservices and includes this handy graphic of the journey to date from the first to the third platform:

• TechCrunch on why successful digital transformation requires ‘total organizational commitment’.  Anything short of that is likely to result in failure at first attempt.

Startups and Hiring

• Thunderclap is a social amplification and pre-crowdfunding platform.

Thunderclap is a smart companion for any crowdfunding project. Think about it: your most loyal supporters are willing to donate money to fund your project. Those same supporters will likely want to donate their social reach to spread the word. That’s where Thunderclap comes in.

• Do Google hire programmers based on their search queries? Some interesting responses in this Quora thread.

• Every developer knows it always takes longer than you think even if you take that thought into account.  Jeff Atwood’s tweet echoes the wonder of Hofstaders Law:

In software, the first 90% of the work takes 90% of the time, the last 10% takes the other 90% of the time.

— Jeff Atwood (@codinghorror) January 29, 2016

Undercultures

• This Guardian examination of the life, faked death and subsequent confused resurrection of a notorious troll called Eris is riveting.  It shines a light on the intensity of the world of online games forums and the emotional vulnerable and often very young adolescents that inhabit them who seem to think that social norms and conventions don’t apply to them.

• Talking of which, the Daily Beast looks at the shabby antics of the Oxford Bullingdon Club a ‘dining club’ for the elite with its own code of silence that has been breaking bad for 235 years.  The Bullingdon counts among its alumni David Cameron, George Osborne, Boris Johnson and Cecil Rhodes, the Victorian colonialist whose statue will continue to survey proceedings with impunity outside Oriel College.  We’re all in it together.

• Just when you thought ‘peak beard’ had passed, evidence emerges that growing a beard may confer health benefits.

On the shortness of life

• What is the most difficult thing to learn and accept about life? Quora answers:

The fact that death is inevitable is the most commonly mentioned hard to accept fact.

• Paul Graham’s latest essay is a characteristically plain, powerful and essential read simply entitled “Life is Short“.  In it he outlines his coping strategies:

Relentlessly prune bullshit, don’t wait to do things that matter, and savor the time you have. That’s what you do when life is short.

• French writer André Gide was able to use the inevitability of his own death as a mobilising force for his creative work and saw aging as an agent of revitalisation:

The thought of death pursues me with a strange insistence. Every time I make a gesture, I calculate: how many times already? I compute: how many times more? and full of despair, I feel the turn of the year rushing toward me. And as I measure how the water is withdrawing around me, my thirst increases and I feel younger in proportion to the little time that remains to me to feel it.