CapsNet : New AI Architecture


Last week Google’s AI researcher Prof. Geoff Hinton published two papers on what he termed as “Capsule Network” as a outcome of his decades of work in Artificial Intelligence. He declared that AI researchers have been addressing object detection in image processing wrongly. If human child can quickly learn to identify pets with a small training, AI should also be able to do so.

The Capsule network requires smaller training dataset and also showed fewer errors than existing deep learning algorithms. As per the papers, a capsule is a group of neurons whose activity vector represents the instantiation parameters of a specific type of entity such as an object or object part. Active capsules makes predictions and when multiple predictions agree, a higher level capsule become active. Such trained multi-layer capsule system achieves state-of-the-art performance on MNIST.

CapsNet has three layers as shown in the diagram. This model gives results comparable with deep NN.


Intel’s Artificial Intelligence Answer to Nvidia : Nervana Acquisition

Artificial Intelligence in general and Deep learning in specific has been one of the most important technology adopted this year. A lot of artificial intelligence – deep learning open sources as well as commercial softwares are coming up to solve problems from image processing to fraud detection. Google’s TensorFlow, Facebook’s Torch, Amazon’s DSSTNE are few important deep learning open sources released this year.

DeepLearning processing requires a different type of compute suitable for GPU/Cuda – massively in-memory processing . Earlier this year Nividia announced Tesla P-100 processor and DGX-1 a machine specifically catered towards running deep learning / artificial intelligence type of work loads. Nividia also provided a Deep Learning SDK framework for developers.

Intel’s first answer to this was Intel Xeon Phi chip with 72 core, coupled with an on-package, high-bandwidth, memory subsystem (Multi-Channel DRAM) and integrated fabric technology called Intel® Omni-Path Architecture (Intel® OPA). However, it needed to do more of Software framework and better chipset. It achieved both with the acquisition of Nervana, founded by ex-Qualcomm researcher Navin Rao.

According to an unofficial source (Re-code), this 2.5 years old Nervana has been acquired at 400million+. This acquisition will given Intel Neon as a fast DL framework and upcoming Nervana engine (ASIC chipset to be released in 2017). This acquisition

Reflections on Artificial Intelligence Based Chat Bots

Screen Shot 2016-08-08 at 3.15.58 PM

It has been years since, applications and softwares like, Apple’s SIRI and IBM’s Watson have introduced personal assistance and AI for practical use cases to consumers and enterprises. In recent months, Microsoft, Facebook, Google, Amazon, all opened up Chat frameworks or chat-based assistant.

Anyway, the entry of the giants in the area of chat bots and personal assistance has triggered a lot of startups get started in the space. Over the last few months, many startups got funded. Some of them Digital Genius,,, Ozlo, Maluuba, Arya, Mezi, talla, GrowBot,, Your.MD etc.

In 2009 we at GloMantra, where I was CTO, started working on a personal assistance and developed assistance similar to SIRI before its acquisition by Apple. We provided voice and text activated actionable recommendations on Mobile (also supported Facebook and Web App). The solution was implemented using variety of technologies like voice to text, NLP, NLU, semantic search, user personalization (interest profile). What we developed was a text message type interface with output as text and action buttons or chat plus matching-recommendations.

Now that everywhere chat based personal assistance are coming up, I was thinking, did GloMantra did the personal assistance too early and, rather, did we give up too early (ran out of money in 2013)?

A lot of technologies like Natural language processing, Intent classification, semantic search, artificial intelligence are now much better shape and available as open sources, there are a lot of possibilities that AI based chat box solving variety of problems and tasks.

Anyway, now the market is getting flooded with a variety of chat bot startups. I recently read a comment of Phil Libin, the former CEO of Evernote who is now a VC. He is quoted as, “I’ve heard 200 bot pitches over the last couple of months.”

But is this market in its early hype cycles? Having done some work in the area, getting it right to handle variety of unpredictable way of human interactions is not that easy too. So, it will be interesting to see how many do it right and survive / succeed. But it is for sure that the AI based chatbots are going to be there solving consumer to enterprise to human-machine interface use cases.