Inside | Real news, curated by real humans
Inside AI

Inside AI (Oct 30th, 2016)

Happy Sunday and welcome to the latest edition of Technically Sentient!  This edition is awesome with some great links, and really interesting guest commentary about whether companies should starting having an "information balance sheet" about the data they have and its value.  The newsletter is divided into sections to make it easy to skip around if you desire.  They are:  Big Idea, Must Read Links, Industry Links, Research Links, and Commentary.  Let's get to it...
-- BIG IDEA --

The most interesting thing I read this week was this article about some new research out of MIT that shows how to train a neural network in a way that they can inspect why the NN made the decision it did.  This is a huge step forward and I'm excited to see people working on this problem.  Neural networks were in the news a lot this summer over a proposed EU law that would require that any algorithm used to make a decision about a person be able to explain that decision.  Since most neural nets are not inspectable and we can't understand the "why" behind their answers, there were big concerns that neural nets might not comply with this law if it passes.  But research in this direction, while early, is very promising and is going to be a very good thing for humanity if we can start to better understand the criteria behind the underlying algorithm learned by a neural net.  Note that, this approach from MIT isn't one the type that has been worked on by other teams to make more direct inspection of a NN more feasible.  So, the creativity of the approach is cool too.  Definitely something to keep an eye on. (Also, Wired wrote just this week about the challenges of holding machine learning models accountable).
An A.I. Judge Predicts Court Cases With 79% Accuracy.  Link.
This is interesting not only because these judgements require some understanding of human morality, but also because it raises a question of whether someday robots will be better judges than humans.  And if they really are, could we accept their decisions?  

Google A.I. Evolves New Cryptography.  Link.
A very cool experiment, and while they don't go into details, it sounds like this was an evolutionary computation approach to neural networks.  Cool to see more people combining key A.I. methods as we better understand them.

IBM's Watson Is Everywhere, But What Is It?  Link.  Also, IBM is investing in Watson to drive their future growth.  Link.
These pieces are interesting because Watson is a big fixture in the A.I. ecosystem, but poorly understood by many people.

A great article about Google's attempt at Natural Language Understanding.  Link.
I wrote in a previous newsletter about Apple, Google, and Amazon and how they would compete, and I still believe Google is the best positioned to solve the NLU problem because of all their search data, which displays a level of user intent that Amazon and Apple don't have in their data.  I'm very long Google when it comes to A.I.


Yoshua Bengio launches a Deep Learning Incubator.  Link.

How Do Neural Networks See the World?  (video) Link.

Did Microsoft Just Outwit Slack?  (my Venturebeat post on the Slack - IBM Watson partnership)  Link.

Tim Cook on how Apple thinks about A.I.  Link.

The Darker Side of Machine Learning.  Link.

How to Boost Your Marketing With Artificial Intelligence.  Link.

Uber's Otto makes the world's first fully automated delivery trip.  Link.

Deep Learning:  The Data Delusion.  Link.

Andrew Ng on why A.I. is the new electricity.  (video) Link.

The State of Enterprise Machine Learning.  Link.

A summary of the 2016 World Robot Conference.  Link.

Artificial Intelligence Safety and Cybersecurity:  A Timeline of Failures.  Link.
A Review of 40 Years of Cognitive Architecture Research.  Link.
This week we have some great guest commentary by my friend and co-worker Will Murphy.  Hope you enjoy it.  

Data is the New Dollar: The Future of Value

Historically, companies have pursued the dollar as a measure of value in the pursuit of shareholder value. Cash flow, profits, and asset values are all about dollars. I think an additional set of measurements can be added to the mix around the value of data. Businesses should be measured not only via current financial measurements but also by the amount of monetizable data they can capture, consume, store, and utilize. This doesn't change the role of money; it just augments it with some new ways to value companies by looking at the value of the company's data. 

Business Is Evolving
One reason for the rising value of data is the evolving landscape of business. All businesses are starting to look like software companies. It seems to me businesses are mostly choosing one of two paths: become a software company (at least in part), or become obsolete. This is due to the growing capabilities of modern software and hardware and what we can do with the growing amount of available data. Data are utilized in almost every aspect of modern business. I think we can agree; software is taking over business.

The Rise of Information Value Statements
I think that data is so important to businesses that it will become measured and accounted for as an asset. There could even be a measure of "information flow," that is, how much new data is being collected, how is it being used, and how much new value it is generating from training new A.I. models. It could be that the corporate health checks of tomorrow are not focused purely on the financial solvency of a company but also on the creation, flow, and application of data within that organization. In fact, it will become valuable to understand a company's "Information Value Statement" as an augmentation to its Financial Statements.

The Future of Value & A.I.
A.I. will make this even more apparent as it mines more value from current and future datasets. It could be that a relatively small organization that is rich in current data and the ability to collect future data is seen as more valuable than a high-revenue but data-poor company. Additionally, the owners of these large proprietary datasets may begin to find that this data is highly sought after by other companies that are data-poor. Data will become a competitive weapon at a level we haven't seen yet. Those of us who found and run the data-driven companies of the future will need to be very aware of the strategic management of data. And, some prevalent large companies today may become even more dangerous to competitors as they have access to lots of valuable datasets and the ability to use them. So, I'm making some proposals on how to think about this value.

Measuring Value
How do we measure the value of data? Below, I've recommended seven ways that the valuation of data can be approached in the future. There may be more, if you have ideas, feel free to reach out to me.

1. Time Value
This metric measures the present value of data taking into account its value over its entire lifetime. How will the data's value change with time? Will this value grow, reduce or fluctuate over time? How likely is it to retain its value? Some data will have a limited life, and some data is evergreen/ will be valuable forever. Most data will decay and be less valuable with time.

2. Legal Constraints
This metric measures how valuable specific data is for many use cases after all legal limits on the data have been taken into account. Data may carry legal restrictions of usage. Provisions must be made to protect privacy and confidentiality, but if data is to carry a high value, it must also be able to be used for valuable future use cases. Many of the uses aren't known yet, so, a balance must be made between protecting privacy and leaving legal agreements more open as to how the data can be used to create future value.

3. Context Value
This metric relates the number of contexts in which the dataset is useful. Specific data may only be useful for some industries. The more a dataset can be used, the more overall value it will have. Some narrow-context datasets, however, could be extremely valuable, depending on the context area. An interesting challenge will be in creating data value arbitrage between contexts. The data owner will want to collect data that is lower in value but has the potential for novel uses that which will increase its value. This will allow smart people to buy data sets, with lower perceived value, and to turn them around to be used in new contexts for a profit on the difference.

4. Quality
For data to be useful, it must have a high quality in terms of accuracy, completeness, and reliability.

5. Acquisition and Storage
These metrics measure how much it costs the business to acquire the data and then store it.  The cost of acquiring data must contain the business costs of acquisition. It's not just about the technical cost to acquire, but the amount paid to collect or purchase the data in the first place. Maintenance and storage data also has costs associated. These can reduce the value of the data in question.

6. Access
This metric measures the level of access that the company has to data from a both size (amount of data) and number/ variety of data access channels. This may be driven by where the company lies in a larger multi-company value chain. Data gathered directly from customers, partners, and from sensors could all fall under this measure. A company with only one channel of data acquisition could be less valuable than one with many channels of data collection. The amount of data collected would also matter. So, it would be a mix of the measure of the size and variety of data channels available to the company.

7. A.I. Model Training Value
A.I. is the area where the exponential value from data will be unlocked. AI systems are trained by using data. The data to be used in machine learning training will be a big indicator of its value. The more the data fits the definition of good training data, the higher the value will be.

A.I. Entrepreneurship Opportunities
Another interesting thing about data valuation models - is that I want to use them to build better models to speculate where data value may lie. We are beginning the age of A.I. entrepreneurship where spotting new business use cases, new data to be captured, and new uses for old data will be both an art and a science worth a lot. Also, as I mentioned above, data arbitrage could be a big opportunity for those who will be able to take existing data that is undervalued in some contexts and re-position it in new contexts (sometimes, by simply having better A.I. models to generate better insights and patterns).

Overall, A.I. will accelerate the increasing value of some data for entrepreneurs who know how to use it. A.I. entrepreneurs will need to create new tools and models to manage how to think about this. In fact, a data-valuation company is probably a whole new category of company that needs to be launched to work on issues like this. The economics of data is a crucial way to measure and run future companies. Investors, founders, and big company executives who understand this will have a big advantage over those who don't.

You can learn more about Will here, and can contact him at if you want to respond.
Thanks again for reading.  Please send me any articles you find that you think should be included, and A.I. angel investment deals that look interesting, or just your general thoughts.  I love the feedback.   Enjoy the rest of your weekend.


-- ABOUT ME --
For new readers, I'm the co-founder and CEO of Talla, a Chatops platform that targets HR and other internal service teams.  I'm also an active angel investor. My A.I. related investments include NetraSimbeLegalRobotGreppyIsoclineHydra.aiSensay, and a few more that will be announced soon.   I live in Boston, and spend about 30% of my time in the Bay Area (Talla has a Palo Alto office) and 10% of my time in NYC.  
Inside is powered by Mailchimp. Also, we ❤️  Mailchimp.

Copyright © 2016 Inside, All rights reserved.

You're receiving this email because you are subscribed to Technically Sentient. If you don't want to receive it anymore, go ahead and unsubscribe – or just hit reply and tell us how to make it better.

Subscribe to Inside AI