With all the buzz around Artificial Intelligence (AI), filtering out the reliable and significant news from the noise gets increasingly challenging. To tackle this problem, experts from the Tooploox AI and Research team have delivered a cherry-picked list of what was most important or, at least, the most interesting to read during the month.
The year 2020 has brought changes to multiple fields, starting from our daily lives, ending with the global economy, and accelerating technological trends that were visible before, yet not so popular. The COVID-19 pandemic was both a trigger and fuel to these changes, further explored in our recent “How COVID-19 Changed VC and Startup Reality” report that covered the COVID-induced changes in a more detailed way.
But reducing the year 2020 to the pandemic alone would be an oversimplification. The end of this challenging year has delivered two answers to questions that have been haunting researchers for decades as well as one inspiring conference.
NeurIPS conference – Tooploox contributing to general AI development
The Conference on Neural Information Processing Systems (NeurIPS) is one of the top two AI and machine learning scientific conferences in the world. The technology groups researchers from all around the world together in one place.
Of course, this year’s edition was unique due to the pandemic, which induced the ban on international travel and public gatherings. Thus, the conference was held online, with the whole event being streamed through the internet. As a side effect, it made the conference more approachable for data scientists and researchers from all around the world, significantly reducing participation costs.
NeurIPS is one of the longest-running conferences on neural network development, being held since 1989 – the age of the ground-breaking novelty of the Motorola flip phone.
OpenAI-developed GPT natural language processing models deliver groundbreaking performances and tend to produce truly natural-looking texts. The third generation of GPT models deliver further improvements in delivering text and processing natural language.
The model delivers a strong performance in multiple language-related tasks, like answering questions, translations, or determining the sentiment in more complex texts, such as in a movie review. The paper received the spotlight at the conference for being the largest language model ever delivered.
Tooploox And AI Explainability
The Tooploox team, together with researchers from the Wroclaw University of Science and Technology, also participated in the conference with “UCSG-NET – Unsupervised Discovering of Constructive Solid Geometry Tree.” The research focused on building a neural network that transforms primitives (basic 3d shapes like a cube or a sphere) into the desired shape, be that a ship, a building, or an armchair – you decide.
Contrary to the popular approach of seeing an AI solution as a “black box,” this network also explains what was done with the provided shapes to build the more sophisticated object. Thus, the actions are fully explainable. The whole procedure can be tuned to get the desired effect, such as delivering instructions on how to build a particular structure using predefined shapes, like prefabricated elements or lego bricks. The paper was further described in our “Machine learning in 3D modeling – Tooploox at NeurIPS 2020” blog post.
This year’s edition also emphasized the social responsibility of AI researchers, with the need to deliver less biased and more fair models. With the growing importance of AI-based solutions, the need to ensure that there will be no hidden biases or glitches that deliver unfair advantages to one option or another will only get increasingly significant.
The short answer to a long-standing computer science puzzle
At its foundations basics, computer science, starting from Artificial Intelligence to medical equipment to the telco backbone of modern reality, is about processing an endless stream of booleans. The binary code is comparable to atoms, the basic building blocks behind every electronic device.
With most of these functions being relatively simple, only one problem delivered a challenge that boggled the minds of the most renowned mathematicians and computer scientists- the sensitivity problem. In its most basic form, the question is about the impact of flipping a single input bit on the output bit.
The question of the direct impact of a single bit on the output bit has been a thorn in the side of computer scientists for over 30 years, bringing frustration to even the brightest minds – until the end of December 2020, when Hao Huang, a mathematician from Evory University delivered a solution.
To make the matter even more frustrating – the answer fits within just two pages. It is available on Arxiv.
The 50-year challenge of protein folding solved
Deepmind, the Google-owned AI research lab, has been working on the protein folding challenge for some time and the problem finally appears to have been solved. The core of the challenge was in predicting the 3D structure of a protein from the 1D structure (a string) of the amino acid sequence consisting of it. The number of possibilities is countless – in 1969 Cyrus Levinthal noted that it would take longer than the age of the universe to enumerate all the possibilities.
Even if not impossible, the task appeared to be on the verge.
The challenge itself has been haunting researchers for 50 years, putting a strong obstacle on the further development of natural sciences and biology. The core of the problem was the ability to predict the way a particular protein would fold. The shape of the protein is closely related to its function. A properly folded protein translated from the mRNA chain is essential for the correct functionality of living organisms.
Following that, improperly folded proteins are responsible for multiple diseases and deformations, including neurodegenerative or allergies.
With the mystery of protein folding tackled, methods of delivering treatments have opened up. Further details are available in the deepmind blog post.
This news is actually from the end of November, yet it was added due to its significance and the fact that it is the first edition of the Tooploox AI and CS news review.
This digest is just the first of a series aiming to deliver a comprehensive-yet-not-overloaded summary of recent AI and CS-related events. Stay tuned for more, or read some of our latest blog posts, including:
- 7 disruptive AI Trends in 2021 to keep an eye on
- Pytorch vs. tensorflow a detailed comparison
- Augmenting AI image recognition with partial evidence
- Hypernetwork approach to generating point clouds
Stay tuned for upcoming editions!