Deeply woven into the fabric of the words and phrases that we use each day, is a real, measurable richness. Accessing it requires intuition, an ability to 'read between the lines', which is the key to human-level language performance.
Mainstream approaches to the 'AI-hard' problem of Natural Language Understanding (NLU) based on Deep Learning, have so far failed to deliver efficient, reliable, human-like solutions to even very basic problems. What's more, these methods will always fall short, not least because they require millions of training examples.
Using an innovative and natural approach based on Chaos Theory, Mandorla is developing a capability which allows the detection and analysis of subtle signals otherwise 'hidden' from machine language systems.
Language understanding, naturally.
Over the last ten years, dramatic developments in AI technology have opened up the field of human language for the very first time.
Using the mathematics of Fractals - which underpin everything we see around us in the natural world - Mandorla has developed an unsupervised approach to Natural Language which offers hope of deep communication between humans and machines, and through this, a chance to ensure the AI revolution is safe and beneficial for all life on earth.
We have built a brand new Natural Language Technology with real humans in mind.
We are not the same.
Machines are playing an ever greator role in making decisions that affect us all, and are starting to take an interest in the way we use language. As they do so, it's vital that their algorithms handle real Human qualities like emotion, intuition and creativity, and that they understand the small but important differences between us. If they don't, there is a real and immediate risk that these qualities will lose their value.
Current state-of-the-art approaches to Language have no way of representing complex emotions, individual differences, or of being creative, and will always be limited by the reductive, context-naive view of language that they're forced to take.
Complex Dynamical Systems.
Language is "Natural" in much the same way that wine or bread is natural: it has a complex, dynamic structure which is more than just the sum of its parts.
It's these so-called "Emergent" properties of natural systems that allow a sommelier to distinguish grape, soil and vintage by taste and smell, or a baker to know when her dough is exactly "right" for the oven. By representing a sentence, paragraph or document as a Complex Dynamical System, our approach encodes its emergent properties in a unique fingerprint - called a trajectory - which is directly comparable to all others.
This type of embedding allows our system to measure the "quality", "sense" or "feeling" of statements and reveal deeper, more useful or relevant information.
Generalisable, unsupervised, human-focussed.
Rather than re-inventing the wheel, Mandorla’s approach can be used in conjunction with all existing methods to provide a different type of insight.
Augments the State-of-the-Art
The fractal trajectories our pipeline produces have the same dimensionality no matter the size of the input text so you can use the same analytics tools to compare, for example, a sentence and a novel.
Fits All Sizes
When a novel and a sentence can be encoded in the same space, you can make meaningful interpolations between them. Ever wanted to sum up a novel in a single word? With Mandorla, maybe you can.
The word “Gold” feels different to a commodity trader than it does to an engineer. We are able to combine raw text with structured data streams to create domain specific embeddings for each individual use case.
Only Relevant Information
Building up low level word or sentence embeddings to represent paragraphs and documents is the current state of the art in NLU technology. Mandorla offers an approach to hierarchical embedding with identical compute cost at every level of recursion, and no theoretical limit to the size of the final representation. It could, for example, represent the entire internet with just a few Gigabytes of data.
One of the great unaddressed problems of machine/human communication is how to get machines to use language like humans do, with flexibility, creativity and individual style. By representing meaning at its deepest level, our embeddings provide building blocks that will allow machines to create new language forms.
Creative Use Of Language
Mandorla offers time, Cost and Data Volume savings relative to DNN based approaches.
A leading concern in the AI field today is that we are not easily able to see how algorithms make decisions or inferences from the input data. When these systems are deployed in sensitive contexts like healthcare, law and security this is a big problem. Mandorla’s approach was built with visualisation in mind, and each stage of the process is human-readable.
Mandorla technology can be applied in any domain where data can be described as a network: Security, Defence, IT Infrastructure, Social Networks, Financial Transactions, Organisational Design, Neuroscience, Communications, to name just a few.
Not Just For Text
When combined with modern Reinforcement or Deep Q Learning techniques, Mandorla can become truly autonomous - identifying gaps in its own skills and understanding, seeking out new material and new processing methods to fill it.
All text is made of meaning.
By converting the semantic information in a sentence or paragraph into a wave that varies with time, we are developing an "ECG for the mind", allowing clinicians to see, at a glance, critical, objective information about their patients' mental state following a consultation.
Stock and derivatives trading
Using real time analysis of news, social media and domain specific forums, we are developing a sentiment derived trade signal for stock, commodity, currency and derivatives markets. Our aim is to map the fractal "signature" of market sentiment to predict short, medium and long term fluctuations.
Imagine if you could say to Google - "these results are great but I want something a bit more... exciting". In order to handle this instruction, a search algorithm has to encode a representation of your personal idea of "exciting". Mandorla Technology can help with this by identifying its unique signature.
Complex dynamical systems have increasingly well understood applications in cryptography and Mandorla technology is equally valuable here. By storing data as a fractal trajectory we effectively hide it in a landscape of infinite size. Only machines with access to the exact algorithms and parameters used at the point of encryption are able to locate it, virtually eliminating the utility of brute-force attacks.
Artificial general intelligence
There is now broad agreement in the community that Strong AI will most likely emerge, not from a single source, but out of collaboration between specialised teams. We see our technology as a missing piece of the Natural Language puzzle that will allow human-level communication by AI agents.
Pioneering human technologists.
Our founding partnership is a union between the worlds of Psychology, Medicine, Information Technology and Business. Unlike most teams tackling the "AI-Hard" problem, Mandorla started with a deep understanding of the human perspective and only then went looking for mathematical representations of it.
Our approach sidesteps many of the pitfalls of conventional, machine-centred solutions and more accurately represents the real, natural structure of human language.
"Computers are incredibly fast, accurate, and stupid.
Human beings are incredibly slow, inaccurate, and brilliant.
Together they are powerful beyond imagination."
- Albert Einstein