Artificial Intelligence is helping security and medicine become more efficient through the use of pattern recognition and improved applications of neural processing locally vis-à-vis moving data offsite to a cloud server. By Dr Manan Suri
The research in facial recognition and cervical cancer.
Using Artificial Intelligence (AI) has many benefits. To simplify – Artificial Intelligence is the process of training a computer to look for patterns, build a database of information based on the recognition of these patterns, and to build meaningful and accurate outcome based on the training data.
The beauty of AI is that every incremental instance of accurately processed information becomes a part of the training data itself. Most people will argue that this is not the case, but what AI does is it processes all new information based on the initial training data and all data it has processed since then to make it more accurate. AI is not just cognitive in terms of processing data but uses machine learning to make the processing more relevant, accurate and valuable.
Dr Manan Suri and his research group at IIT-Delhi has developed multiple proprietary proof-of-concept AI techniques and applications in fields such as security (biometrics), and healthcare (augmenting diagnosis of cervical cancer from pap smears) that are powered by principles of neuromorphic computing. In some of their implementations Dr Suri has successfully used the Neuromem technology in the AI hardware pipeline.
A face recognition system is a computer application capable of identifying or verifying a person from a digital image or a video frame from a video source. One of the ways to do this is by comparing selected facial features from the image and a face database. It is typically used in security systems and can be compared to other biometrics such as fingerprint or eye iris recognition systems.
Dr Suri has been working on making the entire process of Facial Recognition and Speech Recognition, in conjunction with each other, far more accurate and efficient using NeuroMem technology and their ability to localize pattern matching. His team has worked on a multimodal authentication (person identification) system based on simultaneous recognition of face and speech data using a novel and bio-inspired approach that utilizes NeuroMem’s CM1K chip along with other electronic components.
The CM1K chip has a constant recognition time, irrespective of the size of the knowledge base, which gives massive time gains in learning and recognition over software implementations of similar methods. What this means is that the system will check for any input against the entire database simultaneously rather than use elimination and matching each set individually. The system can also be trained to retain any incorrect input and disallow any future use of this input through machine learning.
Making it relevant and secure
Compared to other biometric identification techniques such as fingerprint analysis or iris detection which require active cooperation of participants, face and speech recognition are easier as they often can be completed passively. If the database size of the number of people being included in the facial recognition and speech recognition system are large (and ever increasing), this needs a faster and more power efficient implementation of face and speech recognition algorithms.
Software implementations are impaired by the current paradigm of Von Neumann computing resulting in slower training and recognition times.
For a comprehensive analysis, Dr Suri’s team not only used existing face and speech recognition algorithms involving dimensionality reduction and feature extraction but also developed their own proprietary neuromorphic techniques. For face recognition, several complex pre-processing techniques were compared. Parallelism of CM1K chip allowed the team to have significant increase in the recognition time compared to pure software-based solutions.
Scalability of the CM1K chip and the constant recognition times as opposed to the linearly increasing recognition times of computers (w.r.t increasing size of dataset) provides a huge advantage for real time computations, making them virtually independent of the dataset size.
Using NeuroMem chip along with their proprietary pipeline, Dr Suri and his team successfully developed and tested a methodology, to some degree, the ability to speed up the process of detecting Cervical Cancer at the Pap Smear stage with very low false positives.
In the process of testing for single-cell images, the team used the DTU/Herlev pap smear database. Some classes were grouped as normal cells, and remaining classes were grouped as abnormal cells. For multi-cell images, the team had to manually construct a custom dataset with static images of the virtual slide library.
The images used were of varying magnifications: 200x, 400x, and 600x and included both conventional pap smears, and those prepared through liquid based cytology.
The hardware implementation can achieve a best case mean accuracy of 90 per cent for single-cell, and 86.67 per cent for multi-cell respectively. On average, mean recognition-time/per vector of > 40 X was obtained for both single and multi-cell use cases compared to conventional previously published software implementations. Advantages gained in time and power with comparable accuracy enable the team’s approach to be employed for rapid preliminary diagnostics on mobile/embedded platforms.
For the medical application, Dr Suri and his team worked on an ultra-fast, efficient, low-power, neuromorphic hardware-based solution for cervical cancer diagnosis. Manual classification of pap smears for cervical cancer detection is a laborious and time-consuming process. They developed a bio-inspired approach based on dedicated ASIC based training. They have been able to validate their approach on both single- and multi-cell images.
Given the number of cervical cancer suspect cases and the difference that a timely diagnosis can make, there is a strong need for development of intelligent, low-cost, portable, low-power preliminary diagnostic hardware. Progress in the field of high performance imaging hardware for smartphones has created a favorable pre-cursor for development of true portable mobile on-the-fly diagnostic systems.
Bio-inspired neuromorphic hardware or artificial neural networks (ANNs) in hardware have proven to be promising in the field of image processing, speech recognition, pattern recognition and medical diagnostics.