DeepMind researchers, in their recent work, have investigated the parallels between neuroscience and neural networks. Though the comparison between human neurons and human-made neurons have become obsolete for its exaggerations, researchers are now trying to reverse engineer the process and enhance the way machines are acquiring the reasoning capabilities.
DeepMind’s latest paper titled “MEMO: A Deep Network for Flexible Combination of Episodic Memories” introduced an architecture that is capable of reasoning over longer distances. This was accomplished by adding separation between stored memories and facts, and the items that comprise these facts in external memory, by making use of an adaptive retrieval mechanism, which allowed a variable number of ‘memory hops’ before the answer is produced.
Overview Of MEMO
The primary objective of this new network was to reason out if it has to continue computing or it has already got the answer for the query. To learn the number of computational steps required to effectively answer, information is collected at every step and is used to create an observation. Gated recurrent units then process that observation.
The input to the network is formed by the distance between attention weights of the current time steps and the ones at a previous time step, (both are taken after the softmax), and the number of steps taken so far as a one-hot vector.
In order to test for associative inference capabilities in neural networks, the authors used the images from the ImageNet that were embedded using a pre-trained ResNet. First, the memory content with all the possible pair associations between the items in the sequence is created. And, then they generated all the possible queries. Each query consisted of three images: the cue, the match, and the lure. The cue is an image from the sequence, as is the match. However, the lure is an image from the same memory set but a different sequence.
The queries are presented to the network as a concatenation of three image embedding vectors: the cue, the match, and the lure. If the attention was focused on the same slot of memory for too many consecutive steps, then there was no reason to keep querying the memory because the network has already settled into a fixed point.
This whole work can be summarised as follows:
- Reasoning tasks that will indicate distant relationships among elements distributed across multiple facts.
- An in-depth investigation of the memory representation that will support inferential reasoning.
- A REINFORCE loss component that will learn the optimal number of iterations required to solve a task.
- Significant empirical results on paired associative inference, shortest pathfinding, and bAbI (Weston et al., 2015).
Reaping Rewards Of Neuroscience Obsession
DeepMind’s endeavors revolved around solving the problem of intelligence in machines, as the straightforward trivial tasks for humans can be very sophisticated. Last month, DeepMind with the help of Harvard labs, analyzed dopamine cells in mice and recorded how the mice received rewards while they learn a task. They checked these recordings for the consistency of the activity of the dopamine neurons with standard temporal difference algorithm.
With MEMO, DeepMind once again knocked on the door of artificial general intelligence by carefully probing the reasoning capacity of neural networks. They have tried to capture the essence of distant relationships among elements distributed across multiple facts or memories.