Frequently Asked Questions
What is Intraspexion®?
Intraspexion is a startup organized as a Delaware for-profit corporation, with headquarters near Seattle in the state of Washington. Intraspexion offers a patented Deep Learning software system to identify litigation (and other) risks.
What is “Deep Learning”?
Artificial Intelligence a broad term. A sub-category is Machine Learning (ML). A sub-category of ML is Neural Networks. Then there’s sub-category of Neural Networks called (formally) a “multi-layer” Neural Network, which is also called (informally) Deep Learning. As the result of academic breakthroughs beginning in 2013, such Deep Learning neural networks have enabled computers “to learn” from either images or words or both.
Take driverless vehicles and images: after training runs consisting millions of miles, a Deep Learning system can "understand" the camera and Lidar images well enough to accurately control the gas pedal, brakes, and steering wheel.
How does Intraspexion use Deep Learning?
Now take litigation and words. First, a Deep Learning system must "learn" from examples of words in a specific classification of litigation, e.g., breach of contract, workplace discrimination, fraud, etc. Once a Deep Learning model is trained and "understands" the pattern, it can assess a batch of emails an enterprise generates from yesterday (which no one assesses because even a day’s haystack is too large). Acting like a filter, the system reports to designated personnel in a corporate legal department and delivers only the tiny fraction of the emails that are related to (i.e., pattern-match with) that litigation risk classification. Those emails are the needles a human review can now assess.
Now the human reviewers make the final cut: Which emails are False Positives? Which emails are True Positives and potential “smoking guns” in a potential future lawsuit?
What’s the value of identifying a True Positive?
There are two “values.” First, a True Positive is an impetus for an internal investigation. Then, law department personnel, perhaps with the assistance of outside counsel, may use email threading and also access internal databases to decide whether a potential lawsuit is brewing. Then they can advise a control group executive and (hopefully) nip the risk in the bud. Why find the “smoking guns” after the lawsuits are filed?
What’s the second value?
The second value is this: When True Positives are identified, our system enables human reviewers to tag and store them in a database. When there’s a sufficient amount, we add them to the training set. That’s the second way the training data becomes company-specific, and why the Deep Learning system “learns” to be an even better filter.
What’s the first way the training becomes company-specific?
Our “generic” training set consists of factual allegations from a certain number of previously filed lawsuits without regard to the identity of the defendant. For each prospective client that’s interested in a specific litigation risk, we extract the factual allegations in previously filed lawsuits against that prospective client for that category of risk (which occasionally is in the hundreds of recent lawsuits) and add them to the generic training.
You mentioned patents. Where does Intraspexion stand with them?
Intraspexion is the only startup among the Top 10 companies in the U.S. with “Deep Learning Patents.” By that phrase, we mean patents with "deep learning" or "deep neural" or “multi-layer neural” terms in the legally relevant field of Claim(s). We tracked the results on a year-by-year basis, beginning with 2013.
You can see the explosion of interest in “computers that can learn” from the tracking results: 2013 (3); 2014 (3); 2015 (4); 2016 (36); and 2017 (81); and in 2018 (through September 25) (113).
Currently, for Deep Learning Patents issued from January 1, 2013 through September 25, 2018, the Top 5 Deep Learning Owners (“Assignees") are Google (25), IBM (25), Microsoft (14), Siemens Healthcare (13), and Intraspexion (7, but will soon be 8). After Intraspexion, Amazon, NEC, and Nuance are tied with 6 each.
Intraspexion is the only startup sapling among some very tall trees. Without doubt, we won’t hold 5th place for long. Still, for now, Intraspexion is the only startup in the Top 10.
Is Intraspexion's Deep Learning System just another eDiscovery Software Tool?
No. Here's the key difference: With eDiscovery, the personnel in corporate law departments are forced to look backward for custodians of documents that are potentially relevant to an already-filed specific lawsuit. Those documents are limited by the subject matter and the timeframes in the allegations. The collection is therefore limited and the defendant company can do no more than respond in a reactive way.
At Intraspexion, our system scans all of the emails from each day, and so enables the personnel in corporate law departments to review the few emails (about 1/8th of 1%) that are “related” to a specific litigation risk. They decide which emails are True Positives and can then open a pre-litigation internal investigation. They are enabled to be proactive.
Aren’t there other ways internal investigations can be triggered?
Yes, of course. Someone on the way out might say to HR personnel, “You’ll be hearing from my attorney.” Or the company may have a “tip line.”
Do you have Office 365 connectivity?
How accurate is your system?
We can only tell you about our results using the publicly available Enron dataset. For our first category of litigation risk, we chose "employment discrimination." Enron was known for fraud, not discrimination, but our system nevertheless found a few True Positive for discrimination "needles" there.
How accurate is your system now?
Initially e trained the Deep Learning Engine to "understand" English. Then we added 10,000 Enron non-discrimination emails to the unrelated set, so the model could “understand” English in the context of emails.
Then, as a proxy for the internal company emails "from yesterday,” we grabbed a held-out set of 20,401 Enron emails that our system had not previously assessed.
Result: Our Employment Discrimination "model" called out 25 emails as being "related" to discrimination. Upon human review, 21 emails were marked as False Positives, but four (4) were recognized as True Positives.
The breakthrough was the system's ability to filter 25 out of 20,401 as "related" to the discrimination, a fraction equal to 0.001225, which a little less than one-eighth of one percent. Our System said “don't bother” as to about 99.88% of the emails it processed and "take a look" at 0.12%.
Does Intraspexion store or forward any company emails?
Besides "employment discrimination," what other litigation threats (i.e., types of cases) does your system detect?
We’re an early stage startup and we followed the common advice to early stage companies: “Don’t boil the ocean.” So, currently, our system provides insight only to "employment discrimination" risks. As litigators familiar with PACER already know, there are many business-relevant litigation categories there. We can train a Deep Learning model for any such category. We’ll scale out based on customer demand.
Do you offer a cloud solution?
We are exploring "cloud" solution and have more than one provider in mind. However, based on initial customer interest, our first deployments are for on-premises solutions.
How secure is your system?
For Intraspexion’s “on-prem” solution, emails never go beyond our customer's firewall. When we offer a cloud solution as well, the security may be as good or even better.
What Technical assistance do you provide?
For installations, we provide step-by-step instructions. The instructions include both text and images.
What's your pricing structure?
Each prospective customer has a unique litigation profile. Their prevention priorities differ. As a result, we have no set pricing structure. We customize our pricing and we keep our relationship with each customer confidential.
Hmm. What about the privacy concerns of employees?
First, we don't monitor personal devices. However, where a “computer technology resource" (CTR) policy is in place, and the employee is aware of the CTR policy, and agrees to it, it's generally settled that an employee has no reasonable expectation of privacy when using a company-owned computer resource. See Holmes v. Petrovich Development, LLC, 191 CAl.App.4th 1047, 119 Cal.Rptr.3d 878 (2011) (holding that not even the attorney-client privilege can be raised against a company when a CTP policy is in place and the company monitors the emails in its own computer resources).
For the sake of customer confidence and transparency, can you tell us anything about the software you use?
Yes. We are proud to say that our indexing and visualization capabilities are powered by dtSearch®, www.dtsearch.com. We also use Global Vectors for Word Representation (GloVe) and Google's TensorFlow. Both are “open-source.” GloVe and TensorFlow are included in our system under an Apache 2.0 license, the terms of which are here and are incorporated herein in full by this reference. TensorFlow, the TensorFlow logo and any related marks are trademarks of Google Inc.