What initially got you interested in artificial intelligence?
Randomness. I was reading a book on probability when I came across a famous theorem. At the time, I naively wondered if I could apply this theorem into a natural language problem I was attempting to solve at work. As it turns out, the algorithm already existed unbeknownst to me, it was called the Naïve Bayes, a very famous and simple generative model used in classical machine learning. That theorem was Bayes theorem. I felt this coincidence was a clue, and planted a seed of curiosity to keep learning more.
You’re the CEO of Quantum Stat a company which offers solutions for Natural Language Processing. How did you find yourself in this position?
When there’s a revolution in a new technology some companies are most hesitant than others when facing the unknown. I started my company because pursuing the unknown is fun to me. I also felt it was the right time to venture into the field of NLP given all of the amazing research that has arrived in the past 2 years. The NLP community has the capacity now to achieve a lot more with a lot less given the advent of new NLP techniques that require less data to scale performance.
For readers who may not be familiar with this field, could you share with us what Natural Language Processing does?
NLP is a subfield of AI and analytics that attempts to understand natural language in text, speech or multi-modal learning (text and images/video) and computing it to the point where you are driving insight and/or providing a valuable service. Value can arrive from several angles, from information retrieval in a company’s internal file system, to classifying sentiment in the news, or a GPT-2 twitter bot that helps with your social media marketing (like the one we built couple of weeks ago).
You have a Bachelor of Arts from Hunter College in Experimental Psychology. Do you feel that understanding the human brain and human psychology is an asset when it comes to understanding and expanding the field of Natural Language Processing?
This is contrarian, but unfortunately, no. The analogy of neurons and deep neural networks is simply for illustration and instilling intuition. One can probably learn a lot more from complexity science and engineering. The difficulty with understanding how the brain works is that we are dealing with a complex system. “Intelligence” is an emergent phenomenon from the brain’s complexity interacting with its environment, and very difficult to pin down. Psychology and other social sciences, which are dependent on “reductionism” (top-down) don’t work under this complex paradigm. Here’s the intuition: imagine someone attempting to reduce the Beatle’s song “Let it Be” to the C Major scale. There’s nothing about that scale that predicts “Let it Be” will emerge from it. The same follows with someone attempting to reduce behavior to neural activity in the brain.
As it stands, because deep learning models interpolate data, the more data you feed into the model the less edge cases it will see when making an inference in the wild. This architecture “incentivizes” large datasets to be computed by models in order to increase accuracy of output. However, if we want to achieve more intelligent behavior by AI models, we need to look beyond how much data we have and more towards how we can improve the ability of model’s ability to reason more efficiently, which intuitively, shouldn’t require lots of data. From a complexity perspective, the cellular automata experiments conducted in the past century by physicists John von Neumann and Stephen Wolfram show that complexity can emerge from simple initial conditions and rules. What these conditions/rules should be with regards to AI, is what everyone’s hunting.
You recently launched the ‘Big Bad NLP Database’. What is this database and why does it matter to those in the AI industry?
This database was created for NLP developers to have a seamless access to all the pertinent datasets in the industry. This database helps to index datasets which has a nice secondary effect of being able to be queried by users. Preprocessing data takes the majority of time in the deployment pipeline, and this database attempts to mitigate this problem as much as possible. In addition, it’s a free platform for anyone regardless of whether you are an academic researcher, practitioner, or independent AI guru that wants to get up to speed with NLP data. Link
Quantum Stat currently offers end-end solutions. What are some of these solutions?
We help companies facilitate their NLP modeling pipeline by offering development at any stage. We can cover a wide range of services from data cleaning in the preprocessing stage all the way up to model server deployment in production (these services are also highlighted on our homepage). Not all AI projects come to fruition due to the unknown nature of how your specific data/project architecture works with a state-of-the-art model. Given this uncertainty, our services give companies a chance to iterate on their project at the fraction of cost of hiring a full-time ML engineer.
What recent advancement in AI do you find the most interesting?
The most important advancement of late is the transformer model, you may have heard of it: BERT, RoBERTa, ALBERT, T5 and so on. These transformer models are very appealing because they allow the researcher to achieve state-of-the-art performance with a smaller datasets. Prior to transformers, a developer would require a very large dataset to train a model from scratch. Since these transformers come pretrained on billions of words, it allows for faster iteration of AI projects and it’s what we are mostly involved with at the moment.
Is there anything else that you would like to share about Quantum Stat?
We are working on a new project dealing with financial market sentiment analysis that will be released soon. We have leveraged multiple transformers to give unprecedented insight to how financial news unfolds in real-time. Stay tuned!