As more computing power goes into machines, so do more complex algorithms that test that power. Artificial Intelligence (AI) has been met with consistent processing power limitations on its long road of development. In recent years, many of these issues have begun to be overcome, sometimes with power and other times with better algorithms. DeepMind’s MEMO is one such program that has taken the route of better algorithmic computation.
Alphabet subsidiary DeepMind has been a keystone for AI development. You may remember DeepMind’s AlphaGo winning a Go tournament against world champion Lee Sedol. Fewer may be aware of AlphaStar, which can play the popular computer game StarCraft 2! Each of these AIs, and many others, are run on artificial neural networks, connectionist systems that emulate how brains function.
Neural networks emulate brain function
Neural networks have significantly pushed AI technology forward into more advanced realms that sometimes seem like science fiction, yet there is always room for advancement! Although these advanced AIs can learn on their own, defeat gamers, win debates, and even fold proteins better than scientists, they often have the issue of being slightly slow due to calculating every possible outcome and combing over all information in a databank and calculating all potential outcomes. DeepMind’s MEMO is meant to circumvent that problem in a brilliant fashion.
MEMO has been designed with the intent of providing complex reasoning in machine learning. To perform such a task, the researchers scoured over neurological research papers to study memory in the brain. Our brains don’t just record facts; we file information as memories in different locations, such as the hippocampus. One thing that has been separating the way AI recalls information and the way that we recall memories is that neural networks haven’t been fitted with their own version of the hippocampus. With each researcher contributing in equal measure, their study posted on arXiv states:
“During our every day life we need to make several judgments that require connecting facts which were not experienced together, but acquired across experiences at different points in time. For instance, imagine walking your daughter to a coding summer camp and encountering another little girl with a woman. You might conclude that the woman is the mother of the little girl. Few weeks later, you are at a coffee shop near your house and you see the same little girl, this time with a man. Based on these two separated episodes you might infer that there is a relationship between the woman and the man.
This flexible recombination of single experiences in novel ways to infer unobserved relationships is called inferential reasoning and is supported by the hippocampus”
The power of DeepMind’s MEMO
DeepMind researchers realized that they could add an external memory to AI that functioned as the hippocampus does. Empowering AI to recall associations to data and, with that power, more quickly and readily access immediately relevant information. In other words, DeepMind’s MEMO project almost gives artificial intelligence the ability to have memories, which allows for complex reasoning skills and deeper comprehension.
“To sum up, our contributions are:
1. A new task that stresses the essence of reasoning — i.e. the appreciation of distant relationships among elements distributed across multiple facts.
2. An in depth investigation of the memory representation that support inferential reasoning, and extensions to existing memory architectures that show promising results on these reasoning tasks.
3. A REINFORCE loss component that learn the optimal number of iterations required to learn
to solve a task.4. Significant empirical results on three tasks demonstrating the effectiveness of the above two contributions: paired associative inference, shortest path finding, and bAbI (Weston et al., 2015).”
Inspired by neuroscience literature, the researchers developed a procedurally generated task called paired associative inference (PAI). Using PAI forces AI to learn abstractions to solve problems by capturing inferential reasoning the same way that you and I do. MEMO retains sets of data in memory and learns a projection paired with a mechanism that provides more flexible use of memories. Rather than taking steps to come to conclusions, MEMO is able to use “hops.”
Through a model of human associative memory called REMERGE, where the information retrieved is recirculated as a new query and the difference between retrieved content at different time steps is used to calculate if the model has settled into a fixed point, MEMO indicates whether it wishes to keep computing and querying or whether it’s able to provide an answer.
This ability of MEMO to decide whether or not it needs to continue combing through data and keep thinking or whether it already has the answer makes it significantly faster than other AI models. When put through baseline AI tests in Facebook AI Research’s bAbi suite, MEMO achieved the highest accuracy score on the PAI task and was the only architecture that successfully answered the most complex queries on longer sequences. When compared with other models, MEMO outperformed the baselines in more complex graphs by 20% and only required three data hops as opposed to the best-performing baseline model’s 10 steps.
As we move further into a world in which AI is used every single day behind common tasks most people are seemingly unaware of, just what is our next step in this evolution? How much more human will AI become? Peering through recent DeepMind research articles on dopamine, it seems as if its next venture will be in AI’s mental reward system.
