Peter Zhegin
A VC in london / exABRT Venture Fund / exOzon.ru / @LeedsUniBSchool. I write about technologies that have consequences. A history buff
Key themes of the week in artificial intelligence and cognitive tech
Checking on AI hype, AI and the wider tech ecosystem, the need for wider integration. Weeks June 27th, July 4th
You may find below the list of key themes in AI and cognitive tech covered in the media during a week. Commentaries, quotes from original sources and references are included.
All good in the note below is by authors of orignial texts, all omissions, typos and misinterpretations are by the author of the note.
Themes of the week:
- To avoid disappointment it may be suggested to ask more questions to AI enthusiasts and entrepreneurs;
- AI revitalize some existing technologies and allow using them at the higher capacity (e.g. bio-optics, 3d-sensors);
- There is a challenge in making algorithms less data-hungry;
- AI requires the next level of integration of different platforms, data streams, etc. e.g. Salesforce uses data from Pinterest to help Zillow in selling houses;
- We are moving to reassessment of the value personal data possesses;
- AI’s progress on matching humans in manual jobs is stunning;
- Regulation seems to fail with AI.
AI hype?
- Lenovo is launching an cognitive tech fund [15];
- ‘…we should combat this fictionalization of the reality of AI’ and challenge AI companies by asking inconvenient questions, for example ‘…what cases the system will fail on’, ‘what the expected number of failure cases are in standard use’. It also may be suggested to pay more attention to checking research behind an AI project, even if ‘….there are many AI systems that are useful for real world applications that will likely never receive a paper — but it is a missing element that you should be aware of’ [1]
AI and the wider tech ecosystem
Examples of how the progress in hardware revitalizes concepts, that were developed decades ago and how these concepts, applied in real life empower other technologies.
- ‘The maths for deep learning was done in the 1980s and 1990s… but until now, computers were too slow for us to understand that the math worked well’ — Greg Corrado, a senior research scientist at Google [2];
- Using of bio-optics, (the use of light to look deep within the human body) was limited by the demand for ‘computing power to move from those images into a unified 3D portrayal of the body part being scanned’. Huge power is required due to fact that ‘…calculations are much more difficult because the way in which light interacts with tissue is far more complex than X-rays’. Modern hardware allows to build relevant algorithms that ‘… follows a process over and over, feeding the result from one iteration back as input of the next’. An these algorithms make application of bio-optics reasonable [13];
- Another example is how algorithms empower 3d-sensing technologies. ‘…sensors use light to measure every feature of a building’s interior, room by room and floor by floor, to create a massive data file that captures the spatial geometry of any building. The novelty is in how the Stanford process feeds the raw data file captured by the sensors into a new computer vision algorithm. The algorithm automatically identifies structural elements such as walls and columns, as well as desks, filing cabinets and other furnishings’ [4].
The issue with algorithms is that they are very data hungry.
- ‘The issue with the dominant paradigm is that it’s very, very data-hungry,’ Gary Marcus says. ‘This is proof you can learn faster. And I think that’s something people are going to think about a lot’ [3];
- There are also attempts to cut some corners in researching how the human brain works: ‘The specific approach of MEPhI scientists is that they try to understand and reproduce the principles of information processing in the human brain, assuming that it is unnecessary to reproduce all neurons and ion channels to do so’ [10];
- To allow learning from less data, researchers shift their focus from learning patterns, derived from huge data to understanding ‘programs’ behind these patterns: ‘“What we’re trying to learn is not a signature of features, or a pattern of features. We’re trying to learn a program that generates those characters’ — Joshua Tenenbaum, a co-creator of the ‘…new AI program [that] can recognize a handwritten character about as accurately as a human can, after seeing just a single example’ [3].
The reliance on data is challenging not only from technological point of view. It also requires rethinking questions of how valuable the data is, and who does own it.
- Bloomberg Beta partner James Cham— ‘I think we’ve consistently chosen comfort and better decisions, better and more valuable products, over our sense of privacy. I think that we, as a community, will realize at some point the value of our data and probably figure out ways to make that cost something for companies. Right now, a lot of companies are getting a free ride’ [8].
It seems that AI requires the next level of integration, for example, between diverse platforms and data sources.
Look at an example how Salesforce works:
- ‘…The bot then asks if the buyer [of a house] has a Pinterest board, which she does and is able to share with one click. The bot then uses the Pinterest API to go through the images and links them to MetaMind, Salesforce’s deep-learning platform, accessing a pre-trained classifier that categorizes the style of houses as Colonial, Contemporary, Victorian, Tudor or Greek Revival’. Categorization then helps Salesforce’s clients, e.g. Zillow, to create individualized offers [6];
- Data science tools integration also is mentioned as the first among five trends in data science platform development. Other trends are making the analytical modeling process more efficient, simplified model deployment, managing modeling and experiment results, and introducing cooperation mechanisms [12].
Impact of AI
In the case with white-collar jobs, the idea of coexistence seems to prevail, at least at the current stage of the progress:
- Bob Craig, chief information officer at BakerHostetler: ‘ROSS [AI system] is not a way to replace our attorneys — it is a supplemental tool to help them move faster, learn faster and continually improve’ [5];
- ‘In addition to producing 20 to 90 per cent time-saving on contract reviews, because Kira’s system is trained by experienced practitioners, it helps to ‘institutionalise a firm’s knowledge advantage’ by allowing junior lawyers to benefit from the skills of their senior colleagues ‘— explains Noah Waisberg, a co-founder of due diligence and contract analysis software Kira Systems [5];
- Neither Keith Goldner, a statistician who consults for NBA and NFL, nor Ben Goertzel the lead researcher in the OpenCog AI lab at Hong Kong’s Polytechnic University expect human coaches to fully disappear from the sidelines, but going forward we’re likely to see an increased partnership between coach and machine [7];
- ‘Website building platforms are experimenting with AI as part of the future of design. The goal isn’t to replace great artists, but to make everyone else into a pretty good one’ — Michael J. Coren [19]’;
- People have to leave to AI things that it performs better and to concentrate on something that they are better in. ‘In short, I would rather sacrifice code to the machine (as it is arguable far better suited) and spend the time focusing on how can leverage the machines to make us better humans’ [17];
- ‘You will be paid by how well you work with AI’ — Kevin Kelly, the founding executive editor of Wired magazine [18].
Jobs that require manual and repetitive operations are at the higher risk of being replaced by AI. Despite the fact that currently humans are better in many operations, the rate of AI progress is high.
- ‘Amazon is quick to point out that there is no move to replace workers just yet. Human workers typically pick 400 items per hour, and they won’t suffer the 16.7 percent failure rate of the Picking Challenge leader’ [9] and [11] for detail about the challenge;
- However, Delft, a picking robot, was over three times faster at picking objects than last year’s champion (100 per hour versus 30) [9].
Regulation concerns
The question is how to regulate AI, regulators are too slow and not technically savvy.
- ‘… market regulators were eons behind the complexity that precipitated the 2010 flash crash. So what can regulators do to evolve in a ML [Machine Learning] markets environment?’ [14];
- ‘Regulation According to a Euromoney Survey and report commissioned by Baker & McKenzie, out of 424 financial professionals, 76% believe that financial regulators are not up to speed on AI and 47% are not confident that their own organizations understand the risks of using AI’ [16].
Sources
[1] http://smerity.com/articles/2016/ml_not_magic.html
[2] http://www.theguardian.stfi.re/technology/2016/jun/28/google-says-machine-learning-is-the-future-so-i-tried-it-myself?sf=bbabovn&utm_content=bufferef65a&utm_medium=social&utm_source=twitter.com&utm_campaign=buffer#ab https://twitter.com/alexhern
[3] https://www.technologyreview.com/s/544376/this-ai-algorithm-learns-simple-tasks-as-fast-as-we-do/
[5] http://raconteur.net/business/firms-must-embrace-ai-or-risk-being-left-behind
[6] https://medium.com/@salesforce/the-ai-first-salesforce-customer-success-platform-6aafc43c9748
[7] http://motherboard.vice.com/read/coach-bots-nba-ai
[9] http://futurism.com/deep-learning-ai-leads-robot-to-victory-in-amazons-picking-challenge/
[10] http://phys.org/news/2016-07-science-verge-emotional.html
