Representational Similarity Analysis (RSA) is a versatile analysis technique that is gaining popularity in relating neuroscientific data from different measurement modalities [1–9]. In RSA, a representational dissimilarity matrix (RDM) is computed by comparing patterns of data across every possible pairs of conditions, resulting in a symmetric distance matrix with rows and columns encoding the experimental conditions. This technique is powerful in that it enables comparison of data from multiple modalities, combining the strengths of each modality in understanding brain function, in a simple, elegant manner. This feat of RSA has been coined the Representational Similarity Trick. A recent application has shown that the full spatio-temporal dynamics of visual object recognition in brain and behaviour can be characterised by combining magnetoencephalographic (MEG), functional magnetic resonance imaging (fMRI), behavioural data (collected on our behavioural platform Meadows Research ), and artificial neural network (ANN) data using this RSA trick [4,6]. This approach exemplifies the emerging interdisciplinary field of computational cognitive neuroscience, and representational similarity analysis is increasingly popular in disciplines outside of cognitive neuroscience.
The last decade has seen the development of state-of-the-art computer algorithms, namely artificial neural networks (ANN), that can learn specific representations [10] and match human performance on certain tasks [11]. These models are inspired by brains, and the use of ANNs to try and better understand human brain function is rapidly growing. Indeed, cognitive computational neuroscience (CCN) is an emerging field in the study of brain function, the aim of which is to close the gap between the fields of cognitive psychology, computational neuroscience, and artificial intelligence [12,13]. While such multidisciplinary integration is exciting, in practice it is difficult because a single research team often lacks expertise in all three fields. The versatility of RSA, which can provide representational signatures that can be compared between models, brains, behaviour, can alleviate this difficulty, and be fully exploited by the different scientific communities through our open-source, python toolbox for RSA.
My laboratory and several others around the world are at the forefront of the open-source and community driven development of algorithms underlying RSA (see our GitHub repository at http://github.com/rsagroup/pyrsa). With this proposal, we will continue to maintain the pyRSA Python library for representational similarity analysis applied to medical imaging data, behavioural data and computational models, and continue developing novel representational metrics to construct RDMs, and to compare RDMs between imaging modalities, people, species, models etc. Key to the success of this library will be to extend our existing documentation framework (https://rsa3.readthedocs.io/en/latest/), provide additional comprehensive jupyter-notebook tutorials, and provide additional compatibilities with deep learning frameworks including tensorflow and pytorch.