Biomedical applications with graphs

A graph based approach can be used to classify artery/vein density and thereby do a estimate for the intensity of a disease. The artery/vein map is obtained from a retinal image and then processed to obtain a corresponding graph that corresponds to the particular intensity of the disease/condition.

It is easy to verify that the arterial or the vein graphs are undirected, tree-like. The reason is due to the lack of cycles concerning the blood flow. By looking at graph structure one can derive useful properties particular to the vein and the artery separately. The separation problem hence deals with classification with respect each of these properties. Some of the properties that are looked into are the angle subtended and the connectivity of the graph.

A graph theoretic approach has been in development in DSP, known as Signal Processing on Graphs. Certain viewpoints becomes a lot more simpler in this domain.

A joint work with students of Communication Engineering for the course Estimation and Detection

Deep Learning in Speech Recognition

What is the difference between natural learning (performed by humans) and machine learning?

Learning is to arrange/ classify objects in accordance to some feature set of interest. See how a child learns her/ his mother tongue. The mother utters the word once to the child. Then the child tries to repeat it (or replicate). Consider the word 'അമ്മ' (or mother in malayalam). Since this might be the first word introduced to the child, she/he just adds it to her/his dictionary set. A similar sounding word (say ആമ (tortoise in malayalam)) takes time for the child to distinguish. Only a feedback and learn or to learn from falling will help the child as the letter അ needs to be distinguished from ആ and the sound മ്മ from മ.Once the subtle differences in the sounds are known to the child she/ he can classify similar words. ശാസ്ത്രം | ശസ്ത്രം , മൂല്യം | മൂലം , ആദം | ആദ്യം , കാരണം | കരണം etc

Similar is what happens in spam classification also. From a data-set, we need to distinguish the the spam and the ham (or non-spam) emails. If we provide a learning set to the computer program, with markings of spam and ham, the algorithm can take care on its own. The future emails can be classified again based on the understanding/ pattern identified by the algorithm. How is this different from the learning of the language by the child? The learning performed by a machine is dependent on the code that is written into it. It just looks for patterns that are asked to look. (Ex. The words commonly occurring in a spam : SALE, PRIZE, !!!, WaTch, PassWord, etc. ) It does not look for new features or evolve on its own. On the contrary human systems can accommodate features on its own and is a non-linear evolution process.

The deep learning is associated with the human learning where the learner comes with new features relevant for the better classification spaces. It can be understood as a dimension spreading process where the learner adds one more feature (dimension) in the evolution process. The classification is said to be complete when the learner observes that all the points (or objects) in this space can be classified by different decision boundaries.

As is often quoted "Learning never ends", the machine capable of deep learning should be able to evolve to do the classification at any scale. From isolated words to recognizing sentences and to understand contexts and form interpretations.

 Fig : Example of decision regions in the 2D plane