NATURAL INTELLIGENCE PLATFORM
The rise of Big Data and the Internet of Things provides an opportunity to gain new insights from growing volumes and sources of data not previously available. Up until now, most organizations focused on capturing and filtering the data. The capacity and ability to truly exploit big data to create personalized customer experiences, to learn of emerging business threats and opportunities across myriad and disparate data sources, to simply connect the dots, or to find networks or similarity patterns across massive amounts of data – has remained elusive. Traditional analytics such as Artificial Intelligence applications or Business intelligence solutions continue to offer rule-based systems, which require complex engineering knowledge including rule engines, relational databases, statistical probability and search indexes. Regrettably, none of these approaches is truly intelligent. We believe there is a better way, and it begins with Natural Intelligence.
EXECUTIVE SUMMARY
As humans, we have the ability to see, analyze and understand the world around us in real-time. We assimilate countless facts, ideas and events, and make sense of our world by associating these inputs in our memory. We find similarities between people, places and things, using reasoning from the past to understand the present and to anticipate the future. In the business world, these abilities let us prepare for future outcomes to maximize profit while limiting our costs. Until now, such capabilities were limited to human beings, but the power of cognitive computing platforms is bringing computers closer to learning and reasoning like us.
Just as humans find meaning in the data around them, Saffron brings to Big Data the incredible ability of the brain to find meaning in data, but at a scale, that far exceeds human capabilities. With each new experience, Saffron relentlessly builds memories, learns from outcomes and makes new connections, giving organizations the ability to address new situations in profound and creative ways. Saffron helps every business benefit from the knowledge in their data to make more informed decisions and get ready for potential outcomes.
Organized into five sections, this paper explains how Saffron’s unique and innovative Natural Intelligence Platform works and the methods to change the signal to noise ratio in the analysis process. Saffron helps businesses discover untapped knowledge within their data and reveal connections between data points that were not visible before.
1. The Principles of Natural Intelligence
If technology is to assist us in transforming data into information, and information into actionable knowledge, then it needs to be intuitive and more natural to the way we work
2. Associative Memory and Cognitive Distance
We apply the power of Natural Intelligence in the real world. Saffron unifies data points in its associative memory and finds patterns in the data using similarity analysis.
3. Entities and Networks
Saffron’s network of networks is more natural, like independent agents that hold their own memories and express certain links to each other, depending on context.
4. Case Studies
Saffron’s track-record of successful implementation across multiple industries including manufacturing, healthcare, philanthropy and defense illustrates the ease and impact of using the Natural Intelligence platform.
5. Conclusion
Saffron has developed the technology of natural intelligence. Today, we apply our Natural Intelligence platform to address serious problems for government and corporate enterprises. However, we know that this is only the beginning of what is possible.
1. THE PRINCIPLES OF NATURAL INTELLIGENCE
The Genius of the Human Brain
Technology to address Big Data needs to be more natural to the way we work in order to assist us in transforming data into actionable knowledge. Saffron’s Natural Intelligence Platform offers a revolutionary approach, inspired by natural systems, to address the ever-growing problem of data and to help us transform data into intelligence. The company develops software founded on principles of memory-based representation and reasoning to make applications mimic the human brain.
There are many examples of associative memory from ordinary life. When we are choosing a restaurant, we intuitively integrate a variety of factors. We note how often and when we last ate there, we add in the quality of the food and service, we calculate the price and the time of day and we often adjust for the patrons (i.e. children or dog friendly). Our brains associate many of these factors to make an informed decision.
Similarly, we use associative memory to identify people from various parts of our lives. The associations contained in these relationships matter: Were we on a sports team together? What other interests or friends do we share? When or where did we know each other? Relational databases lose much of this valuable information a priori, as the restrictions of columns and fields are rigid, predefined, often due to the constraints on size, structure and management of computer maintenance. Saffron’s cognitive computing platform not only reads free text to identify relationships, it analyzes the strength of relationships based on correlation, count and context.
Memories Rather Than Models
Intelligence includes the ability to associate how things relate to other things. Associations, learned from experience, are not a novel concept. Aristotle believed associationism[1] defined knowledge. The birth and foundation of Psychology is the elemental study of how sensations, ideas and actions are associated to each other. Psychology was briefly interested in the metaphor of the brain as computer during the early evolution of information processing and artificial intelligence, but most consider this position is inaccurate[2]. Although the human brain is more complex than the above concept allows, we are always looking for new ways to make computers work more like our brains to assist us better.
As We May Think
Dr. Vannevar Bush was notably the first to articulate the idea of a human-like associative memory for computers in his article “As We May Think” published during July 1945 in the Atlantic Monthly[3]. Bush imagined a machine called MEMEX that would read and remember everything so it could act as a personal memory assistant. He envisioned that MEMEX would remember what he forgot or would recall additional information that might be relevant. “The human mind operates by association,” Bush suggested. “Selection by association, rather than indexing, may yet be mechanized.” Dr. Bush was intimately familiar with the development of computing during and after World War II. The view of computing during this timeframe focused on indexing rather than associating, a view and process still in place today, almost 70 years later.
Models Limit Information
Non-cognitive computers index and retrieve documents and records according to their content. Using these traditional designs, representing and remembering how everything potentially relates to everything else is not scalable and any complexity in the data is challenging.
Traditional methods address this scaling challenge by reducing the information residing in a population of data into an abstract model or set of rules. Rules place specific constraints on the data, the data structures or the relationships within the data. Rules by their nature demand reduction of available information for the sake of simplicity and in doing so also lose the actual relationships and exceptions within the data. Statistics, another traditional method, reduces all the information residing in the data down to a formula. Inevitably, whether using rules or statistics, the inexorable driver for knowledge engineering is to make the data fit into a simpler model. These models may be representative of yesterday’s normal cases and approaches, but today’s problems require knowledge of the exceptions – whether the problem is predicting an adverse event before it happens or providing enhanced diagnostic capabilities through pattern recognition. Rich and detailed information about each entity or individual is what matters and requires a different approach.
Memories Between Data and Models
The foundation of Saffron’s Natural Intelligence approach is our patented associative memory technology, which uniquely addresses the problems unresolved by traditional methods. Searching through massive data sets and reading large volumes of documents to make sense of it is difficult for humans. Saffron’s Associative MemoryBase stores information as associations between data points. As the MEMEX Bush imagined in 1945, Saffron’s associative memory technology recalls all relevant associations from experiences, cases and evidence – as they relate to a specific situation.
2. ASSOCIATIVE MEMORY AND COGNITIVE DISTANCE
The power of Saffron’s memory-based approach rests in its simplicity, inspired by natural intelligence. As humans, we learn how to relate things when we see them together or in sequence. We remember associative coincidences and reason from this experience. We make sense of new situations because they remind us of prior familiarities: we recall what each have in common, and then imagine what is missing or what is new and different. As in human experience, software should also capture knowledge from direct observation of the real world.
Memories: Representation of Coincidence Matrices
Saffron defines memory as association frequencies represented in an associative matrix. An association represents an observed coincidence between any two things. The software counts each observed coincidence and represents an association frequency. An associative matrix is a representation of many association frequencies, which we call the Saffron MemoryBase.
Figure 1 shows how a simple data table represents an associative memory. Rather than recording raw data, such as records, or documents, the associative matrix represents the cross-record and cross-document information about all the things within the data.
Assuming the table, is a subset of a larger database of order transactions, it shows only the specific transactions where ‘country:USA’ appears in Column 5. Using Saffron, we show equivalent associative memory representing the frequencies of association across all the data elements. You can easily see that on May 13 there was one order placed to HP and two orders placed to IBM. Additionally, you can easily discern the relative number of orders placed to each company in relation to the other orders. We use a form of Structured Query Language (SQL) to create a query of companies associated with country:USA on date:05/13 and which results in IBM and HP, with two and one transactions respectively. While you count these from the table, it is much easier to look at the matrix to find useful information, instead of just raw data. You might also note that John has a propensity to order from only IBM.
Memories Are Faster to Exploit
While both the data table and the associative memory can answer this simple query, note how little work is required to recall results with the associative memory vs. pulling data from the table. Saffron’s associative memory approach requires no table joins – Saffron’s SQL does not require or use a JOIN operator – nor does it require the extra overhead of the table scans to collect the frequency metric that the memory already knows. Also, note how more reading of the table is required, in contrast to the fraction of effort required by the associative memory to provide the same answer. Memories do not “search” for answers; they simply lookup and recall the answers from everything they have already memorized. Associative memory query response times accelerate as the data increases when compared to databases. When the time comes to exploit data, an associative memory requires less computation and less input/output access to the total data store – rendering the data volume stored as less relevant.
Businesses need to be able to exploit data rapidly to understand a situation. Questions such as “Who is related to whom?” and “Who is similar to whom?” are quickly answered when information is represented in an associative memory base.
Reasoning by Memory
Memories also provide better answers to other types of queries, such as “what is it?”. The associative memories in Figure 2 represent a set of observed classes for different types of animals. The memories have observed examples of mammals (a dolphin and a horse), examples of fish (a trout and a shark), and an example of a reptile (a python). The memories represent the nonlinear coincidence structure of how the attributes of each animal occur in relation with each other. The table shows a new set of attributes for a reptile, and each memory is asked to imagine how well the pattern “belongs” to its class. Reasoning from memory, this cold-blooded, egg-layer with scales is most likely some sort of a lizard.
This example is simple, and the linear structure of the attributes would suffice. Saffron focuses in more challenging areas where associative richness and detail make all the difference when the nonlinear combinations of attributes define the class. Think about applying this to personalization where a consumer might like to wear the color red, but only red ties not red shoes. Alternatively, consider an example from genomics, where the meaning of one gene expression might be highly dependent on many other gene expressions. In most domains, this is how we think. We think in patterns of attributes, how they occur together, reminding us of past cases and their categorizations, including categories of actions, and outcomes.
Memories Learn Autonomously
Traditionally, the accomplishment of data categorization is by the modeling of rules or statistics. Knowledge engineering, also called “data mining,” includes the authoring of rule bases or the parameterization and training of statistical models. Despite the approach taken, the resulting static models require significant time and effort to develop, deploy, and maintain. As the environment constantly changes, the rules and models quickly become out of date, which in turn requires yet more time and effort to reconfigure, test and redeploy. In contrast, the category results in Figure 2 show how memories observe each case, in real-time, without rules or models. Associative memories are autonomous, non-parametric, incremental learners, similar to the brain.
Memories Learn Quickly
The example in Figure 2 also shows how memory-based reasoning is not data hungry, which is a problem with data mining and the probabilistic assumptions about The Law of Large Numbers[4]. Statisticians might argue that so few examples do not provide enough statistical confidence, but increasingly, the assumptions of statistics are failing.
A good example of the failing of statistics is the drug Vioxx[5]. In 1999, Vioxx was marketed by Merck to the American people as an effective anti-pain medication with minimal side-effects, targeting arthritis sufferers in the over-65 category. As a matter of statistics, Vioxx was safe and effective, but the 8 studies (conducted with 5,400 patients) did not take into account the outliers that could suffer from adverse reactions. In 2004, a detailed FDA study proved that Vioxx had potentially deadly consequences for its patients, greatly increasing the risk of sudden cardiovascular death. This reaction most-likely resulted in the death of between 30,000-60,000 Americans since its introduction. Learning of the pending publication of this study, Merck immediately pulled the drug from the market.
In “The Cult of Statistical Significance: How the Standard Error Costs Us Jobs, Justice, and Lives,”[6] Ziliak and McCloskey describe the mistake of measuring statistical confidence rather than magnitude and importance of effects across a range of sciences and their applications.
In contrast, memory-based reasoning provides effective accuracy with few examples. Using similarity analysis, a single observation of a reptile is enough to estimate that the new case is most likely a reptile based on all the information available to date. Applied to situations more serious, Saffron’s Natural Intelligence Platform can help us create a more informed decision based on the entirety of the available information, not just the information that is statistically in the majority.
3. ENTITIES AND NETWORK
The previous examples are illustrative of associative memories, but real applications are significantly larger data sets and are far more complex. Saffron’s memory-based reasoning is much more elegant than we describe here and it addresses the scale of data far beyond these simple illustrations. Extending Bush’s vision of a MEMEX, imagine an army of personal assistants. Each assistant contains at least one associative memory that learns everything about an entity whenever it appears, whether in structured data or unstructured text. Each assistant reads, remembers and recalls all its associations to other entities (people, places, things and events) in the context of other attributes, which are not first class entities. The following figure shows a number of such associative memories, one for each entity, and how implicitly connected they are to each other.
In Figure 3, each memory represents a network of connections within its entity’s perspective, and the collection of networks is yet another level of connectivity. Other graph-based representations of entity networks are too atomic. In these atomic representations, each entity is an abstract node, and connections are context-free links with only one characteristic. Saffron’s network of networks is more natural, like independent agents that hold their own memories and express certain links to each other depending on context. The blue outlined box within the USA memory illustrates that USA is associated to Mary when the context is HP, and that USA is associated to John when the context is IBM. John’s memory knows the association to IBM when the context is USA. Given any starting point, the memory network can recall the other associated entities, in context.
Observe Associations in Context
Rather than have us search for and read one document at a time while building a matrix and collecting evidence, Saffron reads all the documents and connects all the entities automatically. With Saffron, businesses spend less time searching and more time analyzing. They have the ability to understand associations between people, places and things more quickly and visualize these associations in a variety of ways. The result is a faster access to a complete picture of the available knowledge, enabling quicker and wiser decisions, as illustrated in Figure 4.
Figure 4 shows how Saffron bridges the gap between massive data sources – structured and unstructured – and the individual desktop. The memory base constantly observes and remembers all of the Message traffic. Additional data sources feed Saffron memories by a federated search engine[7]. All unstructured data pass through entity extractors to mark the people, places and things that define the entities. Saffron builds an associative memory for each one, remembering the links, context of links, and snippets of evidence, to warrant each link across all the source data. Saffron provides a search query field, but rather than returning documents to read, the memory base returns entities in rank order of associated relevance to the query. Each entity recalls other associated entities that are relevant to the query context supported by snippet evidence for how the entity is associated to the query. Rather than spend time in researching, collecting, and reading, businesses can more quickly find and make sense of contextually relevant entity networks, and publish reports, thereby accelerating the production of intelligence decisions and possible action.
Massive Network Scalability
Saffron’s MemoryBase can contain millions of memories, which can contain many more millions of coincident attributes. Each memory represents many billions of associative “triples.” A triple is composed of three elements: the name of the entity (also called primary key), the memory of the entity and another entity or attribute paired within it.
In an implementation of this size, Saffron’s memory base requires only three 32-bit general-purpose computers. The figures above introduce the logical idea of associative memories for which Saffron uniquely provides an enterprise solution for massive scale data. Saffron has solved the associative “crossbar” problem using various patented methods for compression and partitioning. Moving beyond normal matrix compression (which requires decompression to access the compressed content), Saffron’s real-time memory allows random access to the representation, both to incrementally write new observations and quickly read the memory to answer any question. Partitioning is also required between many and within large memories (matrices). At the system level, Saffron uniquely provides near-linear, infinite scalability across multi-processor systems.
More information Always Wins
Although there are more intricacies and nuances to memory-based representation and reasoning than is presented here, the basic idea is simple to understand: Capture, store, and exploit more information. Saffron owns unique methods in capturing more information through compression and partitioning and in exploiting this information through nearest-neighbor reasoning. However, starting with more information will always have the advantage.
Context Matters
Perhaps the all-time classic tabular database is the old telephone book. If you know whom you want to call, it is easy. However, imagine if you met someone at a party who was looking for a new bassist for her band, and you could not remember her name. Saffron would reveal not only who was at the party, but also their names, the band’s name, and where they were playing next. It would also tell you who the last bass player was and where he was playing now. You could also find in what bands your friend has played and whether there are other bass players attending the party.
A social network graph, for even a small gang structure, is meaningless to an analyst unless the context of financial transactions, supply routes and control channels is clear within the network; all of which focus on the relevant relationships of interest. Mere membership in a group is irrelevant. In contrast, Saffron’s memory base represents a complex, conditional multi-graph. Saffron captures the context of association automatically within its network of network and uses it to exploit social information effectively. The more the network increases in size, the more this context-dependency is required to find the right links among billions and trillions of links. Simple social network graphs are abstract, reductionist models. To understand the details of relevant entities and links well requires the provision of more information and more data with the appropriate filters.
Knowing Who is Who
Saffron’s memory-based approach addresses both major aspects of entity analytics: “Who is associated to whom?” and “Who is similar to whom?”. Other approaches attempt to find similar people by finding similar names, such as by lexical or similarity. While logical, this method only leads to massive false alarm. For example, even the perfect similarity of identical names in very large databases is unlikely to identify the same person. Other highly informative attributes, such as an address, might also be included but are still limited to only a few principle attribute comparisons. Saffron uses virtually all the information about an entity to lookup other similar entities. Rather than rely on assumptions of rules to select and match on only a few variables, Saffron estimates the matching value of all the attributes in the data, whatever they might be. Furthermore, the matching value based on the entropy of the attribute is a measure of a frequency distribution. By definition of a memory, Saffron memories store such frequency distributions. In Saffron, all information for all attributes is available and quickly exploited to answer similarity-based questions, allowing you to know “Who is who?” more quickly and with higher accuracy.
The Future is About Prediction
Predictive analytics is also part of data mining, dominated by traditional statistical methods. But, to understand the earlier criticisms of statistical methods with clarity, these reductionist models tend to find a few terms that capture the most variance. We have discussed how the average of a distribution is irrelevant to the individual, outlier case. Models tend to follow Occam’s razor: the belief that a good model is the simplest model that captures the most information[8]. To be most simple, modelers tend toward attribute subset selection: to reduce dimensionality, trying to determine which data attributes among all attributes capture the most information. Rules tend to have only a few antecedent conditions and statistical models represent only a few terms, whether direct attributes, or higher order abstractions.
When the goal is to capture as much variance as possible in the fewest terms, most information is lost. Saffron compresses rather than loses information: Imagine all the pair-wise associations as formal terms. One association-term might account for a small amount of variances, but Saffron stores many millions, and many billions of such terms. Overall, Saffron’s memory base accounts for and applies far more information. Its power is similar to an emergent property; each element might be small and simple, but the combination of all the elements provides greater power at large. Furthermore, memory-based reasoning to similar past cases helps de-correlate all this information to focus on the terms that matter most for a given situation.
Saffron memories do not rely on attribute subset selection of the most informative terms. Instead, a memory retains all the attributes to select and “shrinks” them, depending on the desired context. A particular attribute may be of little value in most situations yet under another construct, it may become the most critical to inform.
Cognitive Distance and Kolmogorov Complexity
Associative memories can be divided into auto-associative and hetero-associative memories. Auto-associative memories find patterns to answer questions like, “who is similar to whom?” using entities. Hetero-associative memories recall data from distributed graphs to find patterns, and make predictions. For example, a challenge might be to learn and predict the success of movies. A typical input would be people (director, actor), budget, title (genre), and season of release; the predicted revenue would be the obvious output. A hetero-Associative Memory solves the problem by associating the input vectors (movie data) to the output vector (revenue). Saffron connects and illuminates the dots that matter using reasoning by similarity to discern the signal from the noise. This method is powered by cognitive distance, based on Kolmogorov complexity. To learn more, we invite you to read our “Reasoning by Cognitive Distance on an Associative Memory Fabric” white paper, explaining how Saffron combines the power of associative memories with cognitive distance in order to anticipate outcomes.
4. CASE STUDIES
To judge the true value of Saffron’s technology and memory base approach is to see it in use. Saffron’s early successes range from spam filtering to oil well “back kick” prediction, to national security disambiguation analysis. In 2006, we released our first commercially available applications focused on entity analytics for national security. Since that time, we have applied our memory base technology to solve a host of industry problems ranging from manufacturing to healthcare and energy.
DEFENSE
The Challenge
The United States continues to struggle the corrosive impact of illicit trafficking, especially cocaine and traffickers. This dark, highly mobile, asymmetrical target is extremely difficult to detect. A National Task Force conducts all-source fusion using information and intelligence from an extraordinary array of sources to best place its national and international detection and monitoring assets in support of US and Partner Nation Law Enforcement. Any initiative to decrease the reaction time puts the traffickers at greater risk for interdiction and arrest.
The Saffron Natural Intelligence Solution
After ingesting more than 80 data sources from various government agencies and external organizations, Saffron discovered many connections and associations among data points, in real-time, to include detecting repeating and emerging patterns about people, their associations and aliases.
As part of the process of knowing where the shipments and suppliers are moving, the Law enforcement Officers and Agents need to know the connections and associations within the trafficking world. However, simply observing the data is not sufficient to know who is who, as two identities in the data might be the same person. Sometimes, mere data entry errors such as misspelling a name cause this problem. More often than not, the problem is deeper due to name variants, either because of nicknames, or in foreign intelligence, because there are often many acceptable translations of the same name, such as for Mohammed (or Muhammad or Mohammad and many more). Such name variants are often called aliases, but the harder problem is to find intentional aliasing: when a person attempts to create a different identify with a completely different name, address, and other attributes to the extent possible. When we know little about a given individual, this challenge problem is more difficult. Generally, the task of entity resolution is finding and combining different entities into a single identity.
The Failure of Rules
Analysts developed a large and important database of foreign persons of interest. The assumed name variants and intentional aliases existed, but the database was so large that it was difficult to know where to begin searching. More interestingly, intentional aliasing is the most significant “signal” to find, which begged the question: Could a tool help find potential aliases for an analyst to investigate? An analyst cannot investigate an alias without a target, but targets exist only when there is a suspected alias. Prior rules to match US commercial data were useless in this foreign intelligence data. Name match rules generated false alarms in matching “UNKNOWN UNKNOWN” to “UNKNOWN UNKNOWN” – an exact lexical match rule. This stymied analysis when the names are often unknown. The organization tried to fix the issue by writing a rule to handle each exception. After writing many rules, they achieved no significant improvement in accuracy. The non-cogitative solution remained no better than 2% accurate the customer abandoned this effort.
The Success of Memories
Saffron took a different approach. We represented almost all known attributes for an identity, weighed each attribute by a special computation of its information value and then used the speed of associative lookup to find other similar identities based on all the information across each entity’s complete attribute description. Using cognitive distance, Saffron then asked each entity to identify which other entity was closest to it, from its perspective. It is rare that any two entities will point to each other as closest. Using letters as examples, ‘I’ might think it is most similar to ‘Y’, but ‘Y’ might think it is most similar to ‘X’. Our “reciprocal filter” of the most unusual similarities brought the variants and aliases right to the top.
Results
Saffron was able to anticipate events and deliver actionable evidence in record time, greatly improving the agencies’ efficiency. As a result, the client was able to eliminate knowledge in various silos, apply all of the data on-hand to make informed decisions. With real-time event analysis, they reduced the response time from 3 weeks to 2 hours, sped up analyst training and knowledge acquisition.
MANUFACTURING
Experience Based Decisions
The prior case studies describe entity analytics and the use of entity memories representing information of source data. Associative memories can also learn from other sources, such as human actions. Where the prior case demonstrates the value of associative memories for analytic sense-making, Saffron’s associative memories directly support manufacturing decision-making. This example demonstrates the support of supply chain efficiency.
The Challenge
A global manufacturer needed to improve the efficiency of its supply chain by reducing downtime on one of its critical product assets. The key challenge was to find the perfect time to replace the components, preferably not too prematurely or before any disaster struck.
The Saffron Natural Intelligence Solution
We ingested data from over 40 sources, including customer, technical and historical data, to learn about each asset continuously. Saffron created personalized maintenance schedules for each asset, based on the variable conditions to which they were subject. The cognitive platform then conducted real-time similarity analyses to identify patterns and related parts, vendors, logistics, shipments and performance.
Outcomes-based Decision Support
The Saffron-based solution observes by watching actions and learns from the actions taken. Saffron learns which actions are associated with each situation as well as how the component actions (e.g. vehicle, sensors, routing) are associated with each other. When a new Request for Information (RFI) arrives, Saffron recalls which actions have been associated to similar RFIs as well as ensuring that the recalled component actions make sense with each other based on experience. Saffron provides the capture and recall of human practice. It serves as a personal assistant to help recall past actions and cases and to quicken human response by suggesting relevant past experiences. Saffron also learns the associations of situations and actions to outcomes. Given a new situation to analyze, Saffron will remember all associated actions, and then based on these actions, will describe the likely outcome. Saffron recommends outcomes-based best practices – actions that are similar to past cases that also lead to best outcomes.
Results
Saffron reduced maintenance costs and improved efficiency in supply chain management, without compromising quality, safety or lives. Specifically, we increased response time from weeks to minutes, enabling the self-funding of the project in 90 days. The Return On Investment increased 10x annually, and we saw dramatic improvements in decreasing the recall/false alarm rates: Saffron resulted in a 100% recall /1% false alarm rate whereas the previous system had a 66% recall /16% false alarm rate.
ENERGY
The Challenge
Nuclear power plants in the United States experience thousands of safety-related events annually. Our client, a service company managing US plant operations, needed to extract information from years of reported events to better understand their causes and anticipate possible adverse events.
The Saffron Natural Intelligence Solution
Ingesting 10 years of data from reported events, structured and unstructured, Saffron was able to find repeating and emerging patterns across unrelated events, providing supporting evidence for root causes.
Results
Using Saffron, our client was able to take corrective actions to reduce new occurrences of safety-related events, provide improved training for employees and contractors, enhanced supervision, created workflows that are more efficient and improved quality management with manufacturers.
HEALTHCARE
The Challenge
Distinguishing between two types of heart disease — restrictive cardiomyopathy and constrictive pericarditis — is an enormous challenge. Even top cardiologists in the country provide an incorrect diagnosis of these diseases 24% of the time; less specialized Doctors get it right about 50% of the time.
Dr. Partho Sengupta, Director of Cardiac Ultrasound Research and Associate Professor of Medicine in Cardiology at The Mount Sinai Hospital, needed a way to identify disease patterns resulting from echocardiograms in order to improve diagnostics and save more lives.
The Saffron Natural Intelligence Solution
Dr. Sengupta wanted to use all the data available, not just the seven variables commonly used by his colleagues. Responding to his request, Saffron ingested data from 10,000 attributes per heartbeat and per patient, using 90 metrics in six locations of the heart, collected 20 times in a single heartbeat. No human would have been able to find a pattern in real time from this huge amount of data.
Results
Saffron was able to unify a multitude of dense data points to improve diagnosis accuracy by 90%, outperforming top physicians (76%) and state-of-the-art decision trees (54%).
CONCLUSION
New technologies are required to address the explosive growth of Big Data and data analysis. Databases efficiently store data but cannot rapidly exploit it. Rule based systems require much time and effort to define, test, and maintain, yet remain highly inaccurate in a dynamic world.
It is time for a different approach.
Saffron supports decision-making with a level of information and knowledge far beyond what raw data can produce. Saffron is schema agnostic and can read everything. It connects all the data points regardless of source or structure, and returns analytic results by entity rank rather than search results by document rank. Today, many diverse organizations apply Saffron’s Natural Intelligence Platform to address serious problems for both government and corporate enterprises. An increasing number of customers, experiencing disappointing results in the use of other technologies, value the difference in our associative memory approach. However, we know that this is only the beginning of what is possible. Working together to understand new requirements for prediction and the power of memories for patterns, we are helping solving a new generation of difficult problems with memory-based technology and similarity analysis.
[1] On Memory and Reminiscence – Aristotle, 350 B.C.
[2] A History of Modern Psychology – Duane P. Schultz and Sydney Ellen Schultz, Ninth Edition, 2008
[3] “As we May Think” – Vannevar Bush, The Atlantic, July 2008
http://www.theatlantic.com/doc/194507/bush
[4] The Law of Large Numbers says that in repeated, independent trials with the same probability of success in each trial, the chance that the percentage of successes differs from the probability ‘p’ by more than a fixed amount, ‘e’ > 0, converges to zero as the number of trials goes to infinity for every positive ‘e’ – 12 July 2008
http://www.stat.berkeley.edu/~stark/Java/Html/lln.htm
[5] Timeline: The Rise and Fall of Vioxx – Snigdha Prakash and Vikki Valentine, NPR, November 2007
http://www.npr.org/templates/story/story.php?storyId=5470430
[6] The Cult of Statistical Significance: How the Standard Error Costs Us Jobs, Justice, and Lives” – Ziliak and McCloskey, 2007
[7] Federated search computer programs allow users to search multiple data sources with a single query from a single user interface. The user enters a search query in the portal interface’s search box and the query is sent to every individual database in the portal or federated search list. Access details for the individual databases must be preset in the portal by its owner. – Federated Search, Wikipedia, 2008,
http://en.wikipedia.org/wiki/Federated_search
[8] There is a widespread philosophical presumption that simplicity is a theoretical virtue. This presumption that simpler theories are preferable appears in many guises. Often it remains implicit; sometimes it is invoked as a primitive, self-evident proposition; other times it is elevated to the status of a “Principle” and labeled as such (for example, the ‘Principle of Parsimony”). However, “Occam’s Razor” is a better known name. – Simplicity, Stanford Encyclopedia of Philosophy, 29 October 2004,
http://plato.stanford.edu/entries/simplicity
ABOUT SAFFRON
About Saffron Technology
Saffron helps every business benefit from the knowledge in their data to make more informed decisions and prepare for potential outcomes. Saffron’s Natural Intelligence Platform harnesses the incredible ability of the brain to find meaning in data, but at a scale that far exceeds human capabilities. With each new experience, Saffron builds more memories, learns from outcomes and makes new connections, giving organizations the ability to address new situations in profound and creative ways. Saffron’s track record of successful implementation across multiple industries including manufacturing, healthcare, philanthropy and defense illustrates the ease and impact of using its Natural Intelligence platform. Founded in 1999, Saffron Technology is headquartered in Cary, North Carolina. For more information, please visit www.saffrontech.com.
Contact Information
info@saffrontech.com
+1.858.610.9860
Saffron Technology, Inc.
1000 CentreGreen Way, Suite 160
Cary, NC 27513