Democratizing AI: When do we elect to make AI like SaaS?
I’ll start with historical context before explaining my view on how AI should get ready to become democratized for implementation and use by any business professional. Kind of like SaaS.
Deep Learning has become synonymous with Artificial Intelligence (AI). Deep Neural Networks (DNNs) combined with powerful training and optimization techniques have led to advances in the field of AI, once Machine Learning became a common phrase. However, the concept of Neural Networks (NNs) is not new. Neural Networks date back to 1940–1950, and the idea of using back-propagation in NNs was first proposed in the 1970s. Geoferry Hinton and Yan LeCun showcased the first implementation of leveraging back-propagation, demonstrating the potential of powering AI systems with Neural Networks. However, at the time, the techniques did not scale to more practical scenarios, and even a decade ago could not outperform other classical Machine Learning techniques such as Support Vector Machines (SVMs).
Then, in 2012, “AlexNet” took the Machine Learning research community by storm. It was a showcase of a dramatic improvement in image classification tasks using deep Convolutional Neural Networks (CNNs). The authors showcased for the first time, that the deeper a Neural Network model is, the higher its performance: in other words, one can push the boundary of performance with more sophisticated models and more data. Similar trends appeared across a variety of tasks such as Speech Recognition, Text Classification, and Text Generation. As a result of this monumental achievement, the prestigious Turing Award was presented to the famed Deep Learning trio: Geoffrey Hinton, Yan LeCun and Yoshua Bengio. Other pioneers who have directly or indirectly contributed to the field are too numerous to name but include: Fei Fei Li, Zoubin Ghahramani, Jitendra Malik, Andrew Ng and Juergen Schmidhuber.
Research also led to a surge in startups and applied research groups leveraging state of the art DNN techniques. Some of the top startups and applied research groups which were among the first few or have had the most impact in AI through Deep Learning:
- Deepmind: Founded in 2010, Deepmind built and showcased a system in 2013 which outperformed humans on a variety of games, which immediately got attention from Google, thereby leading to an acquisition in 2014. Deepmind has delivered some of the scientific breakthroughs, one of the latest being AlphaFold: a grand challenge in biology of predicting protein 3D folding structure based only on the amino acid sequence .
- Geometric Intelligence: Zoubin Ghahramani and Gary Marcus founded Geometric Intelligence in 2014, a Machine Learning company leveraging Bayesian methods and Cognitive Science techniques, which was acquired by Uber in 2016 and later became Uber AI.
- OpenAI: In 2015, Sam Altman and Elon Musk founded OpenAI. Started as a non-profit organization, it transitioned from non-profit to a for-profit company after entering into a partnership with Microsoft. Most well known for Reinforcement Learning, as well as Language Modeling: GPT-3, the biggest AI model that has taken the world by storm with a variety of applications in Natural Language Processing.
- MetaMind: Richard Socher from Stanford founded MetaMind in 2014 for advancing AI research and engineering. It was later acquired by Salesforce and became Einstein AI.
- Nervana Systems: It aimed to bring unprecedented scale and simplicity to the application of brain-inspired algorithms. Founded in 2014, it was later acquired by Intel.
- Vicarious: Founded in 2010 by Dileep George and D. Scott Phoenix, Vicarious is an AI company using theorized computational principles of the brain to build software that can think and learn like a human. Backed by Founders Fund, it has raised money from digerati such as Jeff Bezos, Marc Benioff, Vinod Khosla, and Mark Zuckerberg. More recently, Vicarious has been advancing the state-of-the-art in robotics.
- Element AI: Founded by Yoshua Bengio in 2016, Element AI built AI-Powered solutions and services such as predictive modeling, image recognition and natural language processing, which led to an acquisition by ServiceNow.
Despite multiple applied research start-ups in the field of AI, the field has not been democratized easily. Most of the above-mentioned startups were either acquired or have partnered with large firms such as Microsoft (OpenAI). A key moment in NLP (and now beyond NLP) arrived when Transformer based DNNs were popularized by the work described in the seminal paper “Attention is all you need” by Vaswani et. al. This paper heralded the “ImageNet moment for NLP” for many experts in the field and led to large scale, pre-trained language models.
The beginning of making AI easily accessible to the masses arrived, as it usually does in many fields, with open source:
- Tensorflow and Keras: Google released some of the most popular Deep Learning toolkits which has laid the foundation for research and development of AI in the community.
- PyTorch: Soon after Google, Facebook soon released PyTorch which has been much easier to use and made it possible to build more complex AI models with less complicated code.
- OpenCV: It is one of the oldest AI toolkits for computer vision and image processing applications. Most of the existing computer vision projects either use OpenCV or use the ones inspired from OpenCV such as Torchvision.
- Transformers: Huggingface released the Transformers repository which provides easy access to state-of-the-art transformers based Natural Language Processing (NLP) models such as BERT and GPT. This repository has made it possible to access some of the biggest NLP models with just a few lines of code.
In addition, application specific open source AI companies like Rasa for conversational AI applications are also gaining visibility and popularity in the developer community.
But, and this is a big BUT, despite open source availability it’s been impractical to leverage open source AI without prior knowledge or expertise in Deep Learning, since the toolkits are primarily designed for developers and not for application users. There are some frameworks such as Dataiku DSS, which let developers configure Machine Learning and Deep Learning pipelines through a “No-code” interface, however, they still require some expertise and lack application interfaces. Moreover, they do not have capabilities for auto-improving with time or provide ease of self-management. Hence, Deep Learning remains the purview of big technology companies with total expertise in the field, and maybe just a few start-ups.
Democratization of AI
In the last decade terabytes and terabytes of data have been generated every day through digital channels across various applications. Unsupervised and Self-Supervised techniques can be leveraged to build AI models and applications which can extract information from historical data. The true democratization of AI would involve providing access to all application users with models that have the capability to “Self Train” and “Self Manage”, by discovering patterns from data automatically. Then it’s not about data pipelines, and ML toolkits. The AI models deal with those themselves. It just makes Deep Learning AI so much more accessible. This was the kind of revolution that SaaS brought about in traditional software applications, and made SaaS accessible to every business user.
Peter Relan, Founder YouWeb Incubator
Chandra Khatri, Chief Scientist & Head of Conversational AI at Got It AI
Note: My view is quite informed by a company I’ve incubated and I currently run, Got It AI , that is addressing key limitations of NLP and democratizing the field of Conversational AI.