#freshfromthelab: Pedro Costa on how artificial neural networks can enhance our understanding of human cognition.
After recently completing a master’s degree at Imperial College London, Pedro Costa came to the CBCD to discuss the findings from his thesis on cognitive processes in artificial neural networks. He is currently working on developing cognitive tasks for the TABLET (Toddler Attentional Behaviours and Learning with Touchscreens) project, here at the CBCD.
So, what exactly are ‘cognitive processes’ and ‘artificial neural networks’? I hear you ask. Time for a (very exciting, I promise) history lesson!
Human cognition (or thinking) has been studied for since the birth of psychology, with the field pioneering the method of observing people’s behavioural reactions to thought-provoking tasks. One famous example is the Stroop task, where individuals are asked to respond with the colour of the word and inhibit their drive to say the colour spelt out by the letters.
Sounds straightforward right? Have a try using the pic to the left! Tricky, isn’t it?
For most people, this task is actually very difficult – we must stop ourselves from reading the word, process the colour and resolve the conflict – and behavioural observations have consistently shown that when the colour of the word and the colour spelt out do not match, our responses are slowed. So what is our mind doing there? Figuring that out will give us an idea of how our mind functions not only when reading words written in mismatching ink but in general.
Building on these principles, the field of cognitive neuroscience examines how thought processes are implemented in the brain. For instance, researchers have used a method called NIRS (explained in Laurel’s recent post) or fMRI (as explained in our article on Dr Kodosh’s work) to assess what parts of the brain become active when a person engages in the cognitively-demanding tasks, such as the Stroop task.
So that’s cognition, but what exactly are ‘artificial neural networks’?
In recent news, you may have read about AI or Artificial Intelligence (such as The Guardian’s take on AI in human learning) and perhaps stumbled across the rather wordy term artificial neural networks (such as this article on the composition of music by neural network) Inspired by the human brain, these networks consist of layers of interconnected elements (or units) which are responsible for processing information.
The main idea is that it all revolves around whether unit A passes on information to the unit B, and this is decided according to a "weight" that increases, for instance, each time that unit A and B are active together. Why does this matter? Because this reproduces, in a simplified fashion, aspects of how neurons work together: our example would refer to the fact that neurons that fire together tend to "wire together" -- aka, the more neurons have been active together in the past, the more they tend to activate each other in the future. To read up a more scientific yet accessible account, have a look at this MIT resource, but what you can already take away is that computational models help us study aspects of cognitive processing in a very rigorous way. And one of these cognitive processings is learning... and that's relevant for parents because they have some pretty exceptional learners in their homes!
We know that babies are biologically programmed to learn things. Some skills, such as walking, talking and playing, are more easily learned than others. We also know that children differ in their learning styles. If you watched the first episode of the BBC2 programme Babies: Their Wonderful Word (unapologetic plug for our Babylab research!) you will have seen that babies have different approaches to new objects. For instance, when the babies in the programme were shown a Jack-in-the-box, some were more curious than others to explore the new object. GREAT!
So, what can artificial neural networks tell us about our babies?
If we know more about how our babies learn, we may be able to understand how to best offer information to support their development. Artificial neural networks have already paved the way to better understanding learning in adults. In 2017, game-playing AI AlphaGo was able to beat the world’s best player in the ancient Chinese board game Go (see this article from WIRED for a more detailed account). Through the help of artificial neural networks, this AI is now able to learn other games and more recently taught itself chess in the space of four hours!
Instead of games, Pedro’s work on artificial neural networks involved getting the model to complete 6 different tasks. He wanted to see how the artificial networks would work towards the desired output of these tasks, in order to better understand how the human brain processes this information.
Pedro then compared the patterns he found using artificial networks with patterns from the human brain. Although these patterns did not match up in his data, Pedro has taken the first step to developing a method of assessing the brain mechanisms involved in cognitive processes like language learning. This method could be instrumental in mapping the human brain and how it processes information. As we have discussed, this research may have important implications for understanding how our children learn and develop. Artificial neural networks could help us map the brain mechanisms involved in processes such as learning and attention, and thus enable us, parents, to provide the best quality input to enhance our children’s development. Exciting stuff!
Author: Megan Tongs