Computing is racing towards creation of a data driven world faster than one could ever imagine. 90% of world’s data repository is just over two years old as 2.5 exabytes of data get added every day! This is a whopping number equivalent to the size of 530K billions of songs per day. Computing has also become pervasive through internet covering 51.7% of world population or 3.88 billion users.

Machine learning & AI is playing a disruptive role at the backdrop of such phenomenal growth of data. It is powering the creation of a world unimaginable until a few years ago in terms of its ability to understand, predict and plan every walk of life.  With the advent of faster computing power as well as convergence of big data and sensors, statistical and machine learning tools are getting enriched every day. With this, dream of the greatest mathematician of the century, David Hilbert, of a predictable world, will eventually see the light of the day!

Man-Machine interaction is getting more natural than before. With the emergence of techniques such as deep learning, convolutional & recurrent neural networks, neural beam forming as well as LSTM, there has been significant progress towards the unification of access technologies. Today, we are much closer to discovering a unified algorithm leading to the understanding of diverse complex systems. Speech recognition has seen a long standing WER barrier (mostly dealt with GMM based approaches) being broken by deep learning technology by 30% after 10 years. With all these, automated speech recognition is now knocking at the door of a commoner through devices such as Google Home, Amazon Alexa etc.  Vision is also experiencing exponential growth powered by deep learning technology leading to more accurate object recognition, image tagging and so on. In the field of medical science, clinical examination of patients is leveraging this algorithm for quicker and accurate diagnosis increasingly.

Language translation is making a sweeping progress with the help of “one shot learning”. This is making ‘many to many’ language translation, with acceptable latency, a near reality. An insight into the geometry of translation is now making an inroad into the mainstream. This is ably supported by powerful visualization techniques (e.g. Geoff Hinton’s t-SNE). All of these are helping machine learning researchers visualize as well as debug models much faster. Visualization is an extremely crucial step towards democratizing AI within the community.

ML & AI are making quick foray into unravelling the mysteries of life sciences. Application of machine learning in understanding the fundamental issues such as aging, is gaining significant momentum. Supervised learning techniques for the discovery of aging biomarkers (e.g. Yamanaka genes) are published routinely in current literature. Rich insights into cellular metabolism, impact of organic molecules on cellular organelle (e.g. cholesterol on mitochondria) can be obtained using ML techniques. There are discussions in the air to even replace radiologists as deep learning will perform the same job much faster and more accurately.

 

Bayestree and Machine Learning

Bayestree believes there is a single algorithm powering the animate and the inanimate in the universe much in line with Pedro Domingos ! This can be fractal in nature or a cascaded hierarchical sequences of regular input output blocks with functional transforms.  The goal is to discover that elemental building block which, when repeated in a hierarchy, gives rise to complex system, in much the same way as Physicists work towards GUT. The road to that path may not be easy. It requires a building deep insight into the geometry of the input including fast discovery of the embedded subspace.  We are often constrained by the representation of the predictors and inferences from learning algorithms tend to change as representation changes.  Our constant endeavour at Bayestree is to discover approaches which is agnostic to representation of the predictors in every problem we solve! This can be achieved by working on the space of probability distributions and on its manifolds, discovering metrics intrinsic to the manifold itself and recoding algorithm as function of these metrics. Differential geometry is the tool for the same. We rely on discovering the information geometry underlying the data in any problem we undertake to solve. 

Looking at problems in the manner as above has significant benefits in the way Bayestree functions. We can ‘learn from the least’ e.g. learn from small samples for prediction. We can design algorithms which are not data hungry. Additionally, we are solely focused on inventing algorithms which can explain itself or are interpretable.

The discipline above runs into the DNA of Bayestree. Gathering data is expensive, but if available, it must be utilized to the fullest. We apply these principles to design self-diagnosable objects implementing an algorithm to minimize maintenance overhead [“We do ship, but we do not (need to) maintain!”].

To make an impact into the society, we are committed to advancing the frontier of technology in the ML area. Understanding the nuances of the underlying geometry, defining tightness of geometry (e.g. measure effectiveness of translation technology), learning cascaded hierarchical systems with backreferencing from data, ability to compare geometry of multiple hierarchically cascaded systems from data will be the prime movers of our system.

 There are endless applications of all of the above if we crack the underlying geometry of the input space well. Can computer teach music to a student? Can it teach to paint? I have a lyric, can it be sung in the voice of a performer? Can we integrate vision with language? How do healthcare, molecular biology, medicines interact? The list goes on and on and on ….