An autonomous car has a safety application, but is it good enough to predict whether the car will crash?
Of course, not.
The solution, an urgent need for machine learning models to reason its assertions and clear it to humans.
Artificial intelligence is said to be better than humans at doing multiple things. For example, weather prediction, diagnosis of a disease, or winning a game of chess. If we pen down the list, there are end number of things where AI beat humans, except answering questions that require logical reasoning.
Let’s assume there’s an image of certain shapes and sizes fitted in a single frame with multiple colors. Now you question the system, what is the size of the cylinder that is lying at the left side of the brown metal which is on the left of the big sphere? Perhaps, a six-year-old would be able to answer the question just by looking at the image.
But such questions are still out of hand for traditional deep learning models to understand.
Why you need more than just deep learning?
The models run by deep learning only helps you understand the relationship taking place between input and output. Be it reinforcement learning or supervised learning, the input and the output have been defined for the model to understand. This task works great for generation or classification, however, it doesn’t work well when it comes to making decisions also called abstract reasoning. You need to enable the model to reason.
Example of Abstract Reasoning
What is deep reasoning and how do you implement deep reasoning?
Deep reasoning gives you the ability to enable machines that can understand the relationship called implicit (stated) relationship between multiple things.
For instance, all herbivores animals eat plants. The giraffe is an herbivore animal. Over here, the implicit relationship is that all giraffe eats plants, but it was not explicitly (unstated) mentioned. Humans are great when it comes to differentiating the difference between implicit and explicit relationships i.e. relational reasoning. But when it comes to computers it’s a challenge, they need to have strict and explicit rules.
Is there a way we might be able to give computers the ability to reason?
Well, according to DeepMind researchers, the deep learning needs to answer questions with 96 percent accuracy – and this was achieved using three networks namely:
- Relation Network (RN): identify how different objects relate to each other
On processing the image, the network can identify the relationship between different objects. For instance, it learns the potential relationship between object pairs like a red cylinder and a blue cube.
- Long-Short-Term-Memory Network (LSTM): used to process the question
They understand the sequences clearly because they have a great memory, thus remembering the previous part of the sequence.
- Convolutional Neural Network (CNN): to process image
This network identifies every object from an image therefore they can easily identify all the features that are present in the image.
Deep reasoning takeaways
- It allows artificial intelligence to identify the abstract relationship between multiple other different things.
- This might just be another step that will take us closer to what’s called “artificial general intelligence.”
- Relation networks can be plugged into deep learning models to generate relational reasoning functionalities.
Till then, you can keep training deep learning models to achieve more than 96 percent accuracy.