Mr. Gabriel Rene reports

VERSES CHALLENGES AI INDUSTRY WITH BENCHMARK TESTS

Verses AI Inc. today provided a research road map that outlines the key milestones and benchmarks against which to measure the progress and significance of the company's research and development efforts, against conventional deep learning, for the benefit of industry, academia and the public.

"We laid out a road map that can be accessed at [our website], which we expect to use to demonstrate over the course of this year that Verses's approach to AI is able to match or exceed the performance of advanced AI models on multiple industry-standard benchmarks while using materially less data and energy," said Gabriel Rene, founder and chief financial officer of Verses.

This is notable in light of OpenAI CEO Sam Altman's recent statement that the future of AI depends on an energy breakthrough1 along with a plan to raise $7-trillion to reshape the global semiconductor industry.

Mr. Rene further stated: "The implications of meeting these benchmarks is to provide scientific evidence that Verses's approach can yield better, cheaper and faster AI that applies to a broader market opportunity, and is commercialized in our Genius platform. We have published our research road map so that both the industry and the public can track our progress."

First benchmark: classification and generation tasks

With the first benchmark, Verses intends to demonstrate the compute and sample efficiency on image classification and generation tasks such as MNIST and CIFAR; in particular, demonstrating the computational efficiency of Verses's approach over and above other modern Bayesian inference toolboxes, such as NumPyro. The company also intends to show how this approach is competitive with the computational efficiency of traditional deep learning approaches based on tools like PyTorch -- but augmented with the great sample efficiency that comes from adopting a fully Bayesian approach. The company plans to release these results demonstrating the efficient compute and improved sample efficiency of the company's approach to classification and generation tasks around the end of Q1 to Q2 2024 in open-access publications.

Second benchmark: Atari 10k challenge

With the second benchmark, the Atari 10K challenge, Verses intends to demonstrate that its approach is vastly more sample and compute efficient than other alternatives. The initial Atari benchmark challenge was introduced in 2015 and involved producing a single AI system that could meet or beat human-level performance on 26 classic Atari games. The AI model must learn directly from pixel data, using only the score as a reward signal. The initial architecture designed for this was data-heavy, using years of game play -- usually more data than a human player might ever have access to.

To address this, the Atari 100k benchmark was introduced, which restricts the amount of game play used in learning to 100,000 environment steps. Atari 100k is a good benchmark to showcase the power and sample efficiency properties of the active inference approach. The company expects to demonstrate two sources of gains in efficiency. The first comes from fast on-line learning of the world model for the game. The second comes from efficient policy estimation that does not require periodic resets of the sort used by traditional gradient-based methods, such as Q-learning.

Although the Atari 100k (two hours of game play) is the industry-leading benchmark, and Verses plans to demonstrate competitive play at the 100k benchmark, the company intends to further showcase the unique strengths of active inference-based AI, namely, rapid learning and improved sample efficiency by proposing the Atari 10k benchmark challenge (roughly 12 minutes of game play), using only raw pixel data and the score as input. The challenge is to reach human-level performance (or greater) measured on the same amount of game play. Humans can achieve competent play very quickly, but how do advanced architectures perform? Verses intends to demonstrate that our system can outperform sophisticated deep learning on the 10k benchmark-learning to play the game efficiently with little data. The company's preliminary results currently demonstrate that its agents are able to learn the dynamics of game play and score on simple games in only several thousand steps, demonstrating more efficient learning using a model that is 99-per-cent smaller in parameter size than the leading competitors, and able to train on a laptop without a large GPU infrastructure.

The company plans to share final results in Q3 2024, as well as in open-access publications.

Third benchmark: NeurIPS 2024 melting pot challenge

The previous two benchmarks cater to the strengths of deep learning approaches, for example, they often involve noiseless tasks that are completely observed (with no ambiguity) and that involve well-defined reward functions.

These benchmarks do not showcase the power of active inference. For the third benchmark, Verses intends to use the new multiagent NeurIPS melting pot challenge benchmark since the ultimate goal is to develop more naturalistic benchmarks that showcase the ability of active inference agents to deal with uncertain environments. Specifically, one of the main advantages of building active inference agents that work directly in belief space with an explicit representational structure is that it becomes possible to share beliefs between agents.

The company believes that this benchmark will showcase the benefits that active inference brings for engineering multiagent systems and align with the central ambitions of Verses AI research: to create ecosystems of AI systems.

Verses plans to share these results showcasing the unique ability of active inference agents to lay the foundations of smart multiagent systems around Q4 2024 to Q1 2025, additionally in open-access publications.

About Verses AI Inc.

Verses AI is a cognitive computing company specializing in biologically inspired distributed intelligence. The company's flagship offering, Genius, is patterned after natural systems and neuroscience. Genius can learn, adapt and interact with the world. Key features of Genius include generalizability, predictive queries, real-time adaptation and an automated computing network. Built on open standards, Genius transforms disparate data into knowledge models that foster trustworthy collaboration between humans, machines and artificial intelligence, across digital and physical domains. Imagine a smarter world that elevates human potential through innovations inspired by nature.

We seek Safe Harbor.