AI to Tackle Image Reading
September 11, 2017
The new frontier in analytics might just be pictures. Known to baffle even the most advanced AI systems, the ability to break pictures into recognizable parts and then use them to derive meaning has been a quest for many for some time. It appears that Disney Research in cahoots with UC Davis believe they are near a breakthrough.
Phys.org quotes Markus Gross, vice president at Disney Research, as saying,
We’ve seen tremendous progress in the ability of computers to detect and categorize objects, to understand scenes and even to write basic captions, but these capabilities have been developed largely by training computer programs with huge numbers of images that have been carefully and laboriously labeled as to their content. As computer vision applications tackle increasingly complex problems, creating these large training data sets has become a serious bottleneck.
A perfect example of the application of this is MIT attempts to use AI to share recipes and nutritional information just by viewing a picture of food. The sky is the limit when it comes to possibilities if Disney and MIT can help AI over the current hump of limitations.
Catherine Lamsfuss, September 11, 2017
Blockchain Quote to Note: The Value of Big Data as an Efficient Error Reducer
September 6, 2017
I read “Blockchains for Artificial Intelligence: From Decentralized Model Exchanges to Model Audit Trails.” The foundation of the write up is that blockchain technology can be used to bring more control to data and models. The idea is an interesting one. I spotted a passage tucked into the lower 20 percent of the article which I judged to be a quote to note. Here’s the passage I highlighted:
as you added more data — not just a bit more data but orders of magnitude more data — and kept the algorithms the same, then the error rates kept going down, by a lot. By the time the datasets were three orders of magnitude larger, error was less than 5%. In many domains, there’s a world of difference between 18% and 5%, because only the latter is good enough for real-world application. Moreover, the best-performing algorithms were the simplest; and the worst algorithm was the fanciest. Boring old perceptrons from the 1950s were beating state-of-the-art techniques.
Bayesian methods date from the 18th century and work well. Despite LaPlacian and Markovian bolt ons, the drift problem bedevils some implementations. The solution? Pump in more training data, and the centuries old techniques work like a jazzed millennial with a bundle of venture money.
Care to name a large online outfit which may find this an idea worth nudging forward? I don’t think it will be Verizon Oath or Tronc.
Stephen E Arnold, September 6, 2017
An Automatic Observer for Neural Nets
August 25, 2017
We are making progress in training AI systems through the neural net approach, but exactly how those systems make their decisions remains difficult to discern. Now, Tech Crunch reveals, “MIT CSAIL Research Offers a Fully Automated Way to Peer Inside Neural Nets.” Writer Darrell Etherington recalls that, a couple years ago, the same team of researchers described a way to understand these decisions using human reviewers. A fully automated process will be much more efficient and lead to greater understanding of what works and what doesn’t. Etherington explains:
Current deep learning techniques leave a lot of questions around how systems actually arrive at their results – the networks employ successive layers of signal processing to classify objects, translate text, or perform other functions, but we have very little means of gaining insight into how each layer of the network is doing its actual decision-making. The MIT CSAIL team’s system uses doctored neural nets that report back the strength with which every individual node responds to a given input image, and those images that generate the strongest response are then analyzed. This analysis was originally performed by Mechanical Turk workers, who would catalogue each based on specific visual concepts found in the images, but now that work has been automated, so that the classification is machine-generated. Already, the research is providing interesting insight into how neural nets operate, for example showing that a network trained to add color to black and white images ends up concentrating a significant portion of its nodes to identifying textures in the pictures.
The write-up points us to MIT’s own article on the subject for more information. We’re reminded that, because the human thought process is still largely a mystery to us, AI neural nets are based on hypothetical models that attempt to mimic ourselves. Perhaps, the piece suggests, a better understanding of such systems could inform the field of neuroscience. Sounds fair.
Cynthia Murrell, August 25, 2017
Google and Apple Narrow Search Results
August 11, 2017
Remaining relevant means making money in technology and Google and Apple are not about to be outdone by Amazon despite it appears that may be the case. In an effort to stem the potential loss of revenue both Apple and Google are re-engineering their search capabilities to “buttress the value of traditional search.”
According to GeoMarketing, the two tech giants are approaching the same problem from different angles:
In a sense, the battle between the mobile web and apps is a proxy war between Google and Apple.
For Google,
The (Q&A box) fits right in with the current idea of getting direct, personalized responses to queries as opposed to the traditional method of showing infinite hypertext listings based on general popularity. It follows a path that Google has already taken with its search functions, including the automatic addition of the term “near me” into the search box as well as providing searchable menu listings for restaurants and direct bookings to salons and spas.
Apple is focusing on apps rather than search, but with the same end in mind.
As consumers are demanding local results and more organic answers to their search questions, search giants have to continually find ways to accommodate. As long as it results in more revenue, the infinite chase is worth it, we suppose.
Catherine Lamsfuss, August 11, 2017
Palantir Settles Discrimination Case
May 15, 2017
Does this count as irony? Palantir, who has built its data-analysis business largely on its relationships with government organizations, has a Department of Labor analysis to thank for recent charges of discrimination. No word on whether that Department used Palantir software to “sift through” the reports. Now, Business Insider tells us, “Palantir Will Shell Out $1.7 Million to Settle Claims that It Discriminated Against Asian Engineers.” Writer Julie Bort tells us that, in addition to that payout, Palantir will make job offers to eight unspecified Asians. She also explains:
The issue arose because, as a government contractor, Palantir must report its diversity statistics to the government. The Labor Department sifted through these reports and concluded that even though Palantir received a huge number of qualified Asian applicants for certain roles, it was hiring only small numbers of them. Palantir, being the big data company that it is, did its own sifting and produced a data-filled response that it said refuted the allegations and showed that in some tech titles 25%-38% of its employees were Asians. Apparently, Palantirs protestations weren’t enough on to satisfy government regulators, so the company agreed to settle.
For its part, Palantir insists on their innocence but say they settled in order to put the matter behind them. Bort notes the unusual nature of this case—according to the Equal Employment Opportunity Commission, African-Americans, Latin-Americans, and women are more underrepresented in tech fields than Asians. Is the Department of Labor making it a rule to analyze the hiring patterns of companies required to report diversity statistics? If they are consistent, there should soon be a number of such lawsuits regarding discrimination against other groups. We shall see.
Cynthia Murrell, May 15, 2017
To Make Data Analytics Sort of Work: Attention to Detail
March 10, 2017
I read “The Much-Needed Business Facet for Modern Data Integration.” The write up presents some useful information. Not many of the “go fast and break things” crowd will relate to some of the ideas and suggestions, but I found the article refreshing.
What does one do to make modern data centric activities sort of work? The answers are ones that I have found many more youthful wizards often elect to ignore.
Here they are:
- Do data preparation. Yikes. Normalization of data. I have fielded this question in the past, “Who has time for that?” Answer: Too few, gentle reader. Too few.
- Profile the data. Another gasp. In my experience it is helpful to determine what data are actually germane to the goal. Think about the polls for the recent
- Create data libraries. Good idea. But it is much more fun to just recreate data sets. Very Zen like.
- Have rules which are now explained as “data governance.” The jargon does not change the need for editorial and data guidelines.
- Take a stab at data quality. This is another way of saying, “Clean up the data.” Even whiz bang modern systems are confused with differences like I.B.M and International Business Machines or numbers with decimal points in the incorrect place.
- Get colleagues in the game. This is a good idea, but in many organizations in which I have worked “team” is spelled “my bonus.”
Useful checklist. I fear that those who color unicorns will not like the dog work which accompanies implementing the ideas. That’s what makes search and content processing so darned interesting.
Stephen E Arnold, March 10, 2017
ScyllaDB Version 3.1 Available
March 8, 2017
According to Scylla, their latest release is currently the fastest NoSQL database. We learn about the update from SiliconAngle’s article, “ScyllaDB Revamps NoSQL Database in 1.3 Release.” To support their claim, the company points to a performance benchmark test executed by the Yahoo Cloud Serving Benchmark project. That group compared ScyllaDB to the open source Cassandra database, and found Scylla to be 4.6 times faster than a standard Cassandra cluster.
Writer Mike Wheatley elaborates on the product:
ScyllaDB’s biggest differentiator is that it’s compatible with the Apache Cassandra database APIs. As such, the creators claims that ScyllaDB can be used as a drop-in replacement for Cassandra itself, offering users the benefit of improved performance and scale that comes from the integration with a light key/value store.
The company says the new release is geared towards development teams that have struggled with Big Data projects, and claims a number of performance advantages over more traditional development approach, including:
*10X throughput of baseline Cassandra – more than 1,000,000 CQL operations per second per node
*Sub 1msec 99% latency
*10X per-node storage capacity over Cassandra
*Self-tuning database: zero configuration needed to max out hardware
*Unparalleled high availability, native multi-datacenter awareness
*Drop-in replacement for Cassandra – no additional scripts or code required”
Wheatley cites Scylla’s CTO when he points to better integration with graph databases and improved support for Thrift, Date Tiered Compaction Strategy, Large Partitions, Docker, and CQL tracing. I notice the company is hiring as of this writing. Don’t let the Tel Aviv location of Scylla’s headquarters stop from applying you if you don’t happen to live nearby—they note that their developers can work from anywhere in the world.
Cynthia Murrell, March 8, 2016
IBM and Root Access Misstep?
March 2, 2017
Maybe this is fake news? Maybe. Navigate to “Big Blue’s Big Blunder: IBM Accidentally Hands Over Root Access to Its Data Science Servers.” When I read the title, my first reaction was, “Hey, Yahoot is back in the security news.” Wrong.
According to the write up, which I assume to be exposing the “truth”:
IBM left private keys to the Docker host environment in its Data Science Experience service inside freely available containers. This potentially granted the cloud service’s users root access to the underlying container-hosting machines – and potentially to other machines in Big Blue’s Spark computing cluster. Effectively, Big Blue handed its cloud users the secrets needed to potentially commandeer and control its service’s computers.
IBM hopped to it. Two weeks after the stumble was discovered, IBM fixed the problem.
The write up includes this upbeat statement, attributed to the person using a demo account which exposed the glitch:
I think that IBM already has some amazing infosec people and a genuine commitment to protecting their services, and it’s a matter of instilling security culture and processes across their entire organization. That said, any company that has products allowing users to run untrusted code should think long and hard about their system architecture. This is not to imply that containers were poorly designed (because I don’t think they were), but more that they’re so new that best practices in their use are still being actively developed. Compare a newer-model table saw to one decades old: The new one comes stock with an abundance of safety features including emergency stopping, a riving knife, push sticks, etc, as a result of evolving culture and standards through time and understanding.
Bad news. Good news.
Let’s ask Watson about IBM security. Hold that thought, please. Watson is working on health care information. And don’t forget the March 2017 security conference sponsored by those security pros at IBM.
Stephen E Arnold, March 2, 2017
Bad Big Data? Get More Data Then
March 2, 2017
I like the idea that more is better. The idea is particularly magnetic when a company cannot figure out what it’s own, in house, proprietary data mean. Think of the legions of consultants from McKinsey and BCG telling executives what their own data “means.” Toss in the notion of a Big Data in a giant “data lake,” and you have decision makers who cannot use the information they already have.
Well, how does one fix that problem? Easy. Get more data. That sounds like a plan, particularly when the professionals struggling are in charge of figuring out if sales and marketing investments sort of pay for themselves.
I learned that I need more data by reading “Deepening The Data Lake: How Second-Party Data Increases AI For Enterprises.” The headline introduces the amazing data lake concept along with two giant lake front developments: More data and artificial intelligence.
Buzzwords? Heck no. Just solid post millennial reasoning; for example:
there are many marketers with surprisingly sparse data, like the food marketer who does not get many website visitors or authenticated customers downloading coupons. Today, those marketers face a situation where they want to use data science to do user scoring and modeling but, because they only have enough of their own data to fill a shallow lake, they have trouble justifying the costs of scaling the approach in a way that moves the sales needle.
I like that sales needle phrase. Marketers have to justify themselves and many have only “sparse” data. I would suggest that marketers have often useless data like the number of unique clicks, but that’s only polluting the data lake.
The fix is interesting. I learned:
we can think of the marketer’s first-party data – media exposure data, email marketing data, website analytics data, etc. – being the water that fills a data lake. That data is pumped into a data management platform (pictured here as a hydroelectric dam), pumped like electricity through ad tech pipes (demand-side platforms, supply-side platforms and ad servers) and finally delivered to places where it is activated (in the town, where people live)… this infrastructure can exist with even a tiny bit of water but, at the end of the cycle, not enough electricity will be generated to create decent outcomes and sustain a data-driven approach to marketing. This is a long way of saying that the data itself, both in quality and quantity, is needed in ever-larger amounts to create the potential for better targeting and analytics.
Yep, more data.
And what about making sense of the additional data? I learned:
The data is also of extremely high provenance, and I would also be able to use that data in my own environment, where I could model it against my first-party data, such as site visitors or mobile IDs I gathered when I sponsored free Wi-Fi at the last Country Music Awards. The ability to gather and license those specific data sets and use them for modeling in a data lake is going to create massive outcomes in my addressable campaigns and give me an edge I cannot get using traditional ad network approaches with third-party segments. Moreover, the flexibility around data capture enables marketers to use highly disparate data sets, combine and normalize them with metadata – and not have to worry about mapping them to a predefined schema. The associative work happens after the query takes place. That means I don’t need a predefined schema in place for that data to become valuable – a way of saying that the inherent observational bias in traditional approaches (“country music fans love mainstream beer, so I’d better capture that”) never hinders the ability to activate against unforeseen insights.
Okay, I think I understand. No wonder companies hire outfits like blue chip consulting firms to figure out what is going on in their companies. Stated another way, insiders live in the swamp. Outsiders can put the swamp into a context and maybe implement some pollution control systems.
Stephen E Arnold, March 2, 2017
Gradescope Cuts Grading Time in Half, Makes Teachers Lives 50% More Bearable
February 8, 2017
The article titled Professors of the World, Rejoice: Gradescope Brings AI to Grading on Nvidia might more correctly be titled: TAs of the World, Rejoice! In my experience, those hapless, hardworking, underpaid individuals are the ones doing most of the grunt work on college campuses. Any grad student who has faced a stack of essays or tests when their “real work” is calling knows the pain and redundancy of grading. Gradescope is an exciting innovation that cuts the time spent grading in half. The article explains,
The AI isn’t used to directly grade the papers; rather, it turns grading into an automated, highly repeatable exercise by learning to identify and group answers, and thus treat them as batches. Using an interface similar to a photo manager, instructors ensure that the automatically suggested answer groups are correct, and then score each answer with a rubric. In this way, input from users lets the AI continually improve its future predictions.
The trickiest part of this technology was handwriting recognition, and the Berkeley team used a “recurrent neural network trained using the Tesla K40 and GEForce GTX 980 Ti GPUs.” Interestingly, the app was initially created at least partly to prevent cheating. Students have been known to alter their answers after the fact and argue a failure of grading, so a digital record of the paper is extremely useful. This might sound like the end of teachers, but in reality it is the beginning of a giant, global teacher party!
Chelsea Kerwin, February 8, 2017
-
- Subscribe to Beyond Search
Feature archive
News archive
-
