What is the future of AI?

mediumThis post was originally published by Beril Sirmacek at Medium [AI]

Today I have joined to an online AI-coffee-talk where I wanted to talk about ideas which I haven’t discussed earlier. I brought the question myself and suggested some answers. The question was, what is the future of AI? and below I summarize the answers I gave in the talk. Of course, only the future will tell if they will be true or not. However, I wanted to share where I wish AI to go towards at least.

  1. Self-awareness:

The first idea is inspired from the industrial revolution 4.0 expectation, which is corresponding to this statement: machines should be aware of themselves. In other words, machines are expected to be aware of what kind of problems they are experiencing and where. We could have the same expectation from AI. I would personally expect that, AI can tell me the reason of low accuracy. Is it because of a network architecture? are there some bad labels in the data set (and which data labels are those exactly)?

2. Needing less data:

When we want to teach stars and clouds to a child, an example like below would be enough. The child does not need thousands of examples to understand what is a star and what is a cloud. I believe it is important that we gain more understanding of human vision system’s abstraction and reasoning capabilities in order to be able to develop AI systems which can learn from one or two example of a class.

3. Lifelong learning:

We have gained a lot of experiences with reinforcement learning in last few years. However, experiments are limited with short time and environment feedback. We need to know whether we can achieve lifelong learning where the AI system constantly learns new classes, new goals and actions and get better at them. We also need to know how to validate the performance of such system.

4. Imagination:

We humans do not need to experience the negative event in order to learn from it, we can simply imagine and learn. For instance, if we are hiking in a forest and see a rock in front of us, we can simply imagine what happens if we try to jump over it, if we walk around it etc. Then we can make the optimal path planning in our imagination. We don’t need to hit to the rock thousand times and fall on the ground before we learn the best action. We need to know how we can make AI imagine and learn from its imagination in order to take real-time decisions and actions.

5. AI without training:

Lately, researchers are looking for ways to design neural network architectures and finding their optimal ways even without training the network. I believe soon we will develop methods to know exactly how many layers and neurons and which weights we need, when we start using a network or to use this network to apply transfer learning in order to fine-tune it with a small data set.

6. AI without the AI that we know of:

AI does not mean that we need so many deep layers to solve every problem. Sometimes, simple but intelligent algorithms can show equal or even better performances. Knowing that training a neural network creates CO2 emissions hundreds of times more than an airflight, we need to get big responsibility of making decision to train a network or not. I belive that we must avoid training when we can. The literature might be full of burried mathematical algorithms which can come back to the stage in order to solve many problems in a simple way. Dijkstra’s algorithm could be an example to this. Dijkstra has proposed the algorithm in the field of graph theory many years ago. He was not a robotics engineer. However, the researchers found the algorithm later and they saw that it can be helpful with finding the shortest path between two points in a very quick way. Now, the algorithm is back to the labs and highly used in many robotics applications. Likewise, I believe that there are other (and many) mathematical algorithms that we need to look back and bring to life, in order to be able to solve our AI problems without the AI that we know of (without needing very deep neural networks for any problem).

Thank you for reading.

Spread the word

This post was originally published by Beril Sirmacek at Medium [AI]

Related posts