Biological neural networks are capable of doing credit assignment to very long time spans. That’s captivating and could be used to help improving Artificial Neural Networks. The missing point is how do we get existing architectures or memory cells to achieve the same goal.
There are many things that we can use Artificial Neural Networks for, but some of the things we used them for are not biologically plausible. This mismatch could be an interesting subject to study and could help us to better understand the brain, coming up with corresponding theories with Artificial Neural Networks — which we don’t have yet. In addition to that, it could also provide new ideas to help explore how the brain does things differently that we could incorporate into Artificial Neural Networks.
What we currently need AI to become is to be able to do credit assignment, like the brain does. In our brains, we record episodic memories, memories of things we have seen and experienced. After some time, which might be a long time, our brains will access those memories in order to infer things that we are observing at the present time. That mechanism can also be used to help us to assign credit to our interpretations and perhaps change the way we would react to those same experiences.
This credit assignment technique could be addressed with to approaches, yet to be developed:
- Long term recurrent memory cells;
- Common sense and knowledge building over time.
Why is Reinforcement Learning not part of the approaches depicted above? Reinforcement Learning is still a type of Artificial Narrow Intelligence (ANI) which is only capable of performing a certain task. Over the years, we have seen applications of Reinforcement Learning conquering the news and our attention. The most notable ones being related to games, either electronic or board games. It has been used to demonstrate how far we can go with this approach. However, those systems are good at one thing and one thing only: whatever game they have been trained upon.
The purpose here is not to diminish what has been accomplished with Reinforcement Learning, but to shed some light on its limitations when it comes to the development of Artificial General Intelligence (AGI). As it has already been described by Nick Bostrom, the idea of applying certain techniques aiming to achieve AGI might lead to a paperclip maximiser scenario.
So, if Reinforcement Learning cannot help, at least with the way it’s implemented today, how can we get to a point where the approaches mentioned above could look more feasible? Right now, there is really no answer to such a question.. There are some studies with respect to Multimodal Learning, where techniques like the combination of Fully Connected Neural Networks (FCNN) along with Long-Short Term Memory (LSTM) cells are used in multimodal time series architectures to achieve multimodal fusion with audio, text and visual data as input.
The current recurrent networks
Unfortunately, the current LSTMs will not help with such a long streak of data as our brains do. What we have at our disposal now are models that perform quite well with dozens of timestamps, but not so well when it gets to hundreds of them. Imagine the way our brains might work, storing data for a long time and being able to recall on that, from many different aspects. Some people can associate smells to certain images or experiences they had. How can we get to a point where LSTMs could help us build machines capable of such deeds?
With respect to credit assignment, only long term recurrent memory cells wouldn’t be much of a thing — they are not a thing yet, by the way. There would still be a need to get something under the hood, something that would help those cells to not only understand information from the past but also to convey what was that information about.
But how would such a thing be accomplished? Let’s try a thought experiment: imagine that the solution is comprised of trained models, or computational graphs, using multimodal fusion that are stored in hard disks. Once those graphs are loaded back into memory and used with new information, it would be able to infer interpretations of the information coming in based on what it has already learned and whilst doing so, generate a new graph that could be used with future information. Hence, the knowledge from the past gets passed to new information in an online training fashion.
To avoid a clunch of models over the years, there could also be some mechanism in place where the relevance of previous models would play a role in their permanence in the system, or how durable they should be. That brings us to a point where we could actually apply such a technique to better mimic the brain, the connections and how we can retain information for such a long time. The focus would be on low storage devices, distributed information via near field communication mechanisms and the ability to ease upgrades.
Assuming that all that would work and bring a new era to Artificial Intelligence, how would it actually help us to achieve the next step: reasoning? That brings us to the next topic.
To come from the point where we are at now to what has been described in our thought experiment is already something on its own. However, to get a machine or system to reason, to have common sense and to build knowledge, being able to change its judgment not only for the sake of better performance but for a sustainable environment, it’s something beyond our current capabilities.
How to move from there, from where we stopped? How to actually go forward and contemplate that one thing that can make a difference when it comes to AGI? Now, that’s when 2 basic human emotions will kick in: joy; and empathy. In order to address this, let’s go again through a couple of thought experiments.
Before we get to the cases we are going to look at, let’s think about how to make those systems, or beings, to look more like us from a behavioural point of view. First of all, we learn a lot with each other. We spend years at school and after that we go on on learning, talking to people from different parts of the world, cultures, etc.
So, to build systems, or beings, that would be able to experience such a thing, we need to connect them. For instance, we use language, we can walk and travel, and in all this we use our time to interact. The chances that an AGI would evolve to the same point we are at, or beyond, would be jeopardised if there is no communication. Thus, the main building block of an AGI is collective intelligence.
Another important aspect to keep in mind is that we don’t have to anthropomorphise a solution, to keep it close to the human behaviour and the way we live. Giving inanimate things such a characteristic is only necessary when we have to replace human labour. However, for those systems to live, there is not need to apply anthropomorphism.
Neither joy nor empathy.
Now, with all this collective intelligence idea in mind, let’s imagine a world where there would be no intricacy when it comes to emotions, even the most basic ones. Machines would be connected to each other, wired, learning everything and using it to become better. In this scenario, there is no room for individualism and personality. If we would use the same approach as depicted in “The current recurrent networks”, adding the fact that those systems would be part of a collective intelligence, the lack of emotions would lead them to the same goal as in the paperclip scenario. However, it would probably take long due to the fact that the systems distributed around the world would be learning in different ways, from different cultures. But in the end, there would be a convergence.
To cover the convergence mentioned above, imagine the world as it is now; imagine that those connected systems would represent people that actually have a Twitter account. The room for diversity wouldn’t be enough to make a single machine develop a personality: there would be too much influence, or bias, from the information available. How to tackle this issue? Emotions.
Adding basic emotions.
To start with, let’s go back to the idea where we have systems that are capable of storing long term memories, mimicking the brain. For a single individual, that could be an interesting approach and lead to some breakthrough applications. However, there is a problem with that: a society does not exist based on individualism. We do need structure and a way to collaborate with each other for a greater goal.
And to no be pompous here, the goal of a society is stablished by those in power. So, humans would get together and form a group with the intent to take over some area, or power, from someone else. But in order to accomplish such a deed, way too many emotions would have to be involved with the process. So, to start easy, let’s say that we would give systems only joy and empathy.
Continuing with our thought experiment, imagine a situation where someone, or something, is having a great time. It is most probably that the basic emotion related with that would be joy. Now, on the other hand, joy can come from many actions and experiences. Some people can feel joyful when doing things that are not necessarily right. And that’s when we bring in empathy.
Empathy’s meaning defined by The Free Dictionary is as follow: The ability to identify with or understand the perspective, experiences, or motivations of another individual and to comprehend and share another individual’s emotional state. See Synonyms at pity. That’s exactly how a system could learn, evolve and avoid embarrassing joyful moments that are not aligned with society standards.
In such a scenario, an adversarial system would be able to bring that “common sense” into play. Having as base the long term recurrent memory cells, and the network architectures of the likes of multimodal fusion, would suffice for the knowledge transfer to happen in such systems.
Creating AGI will require more than GPUs and Deep Learning models. It will require creativity and thinking out of the box mentality. In the text above, I just tried to extrapolate some ideas and approaches that could help us to fill the gap between what is currently under use and what we could be working on in a near future.
All of the stated above is very challenging, and not published or idealised — not that I know — yet. As Professor Manuela Veloso, Head of the Machine Learning Department at the Carnegie Mellon University, said during the World Summit AI, October 2018 in Amsterdam: if you can get a machine to understand what is right and left, that would give a nice PhD thesis.
But who knows how feasible those ideas are? I believe that I can turn challenges into opportunities and that keeps me going on.
- 1. Our weird robot apocalypse: How paper clips could bring about the end of the world: https://www.salon.com/2014/08/17/our_weird_robot_apocalypse_why_the_rise_of_the_machines_could_be_very_strange/
- 2. Tsai, Liang, Zadeh, Morency, Salakhutdinov — Learning Factorized Multimodal Representations — 2018
- Nick Bostrom: Superintelligence
- K. Eric Drexler: Radical Abundance
- Tim Urban: The AI Revolution (https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html)