Computer scientists gain insight into neural networks
Researchers from Massachusetts Institute of Technology (MIT) have taken steps towards understanding how neural networks – generally considered “black boxes” – really work. This could help provide insight into how the human brain responds to visual images.
Neural networks learn to perform tasks by “training” using huge sets of data, in a similar way to how the human brain learns to detect patterns through experience. The networks are made up of a number of “nodes” (artificial neurons) which process information and pass it on to their neighbours through connections (artificial synapses), firing with varying strength.
Despite our reliance on neural networks to take over important jobs in the future, such as controlling self-driving cars or to analyse video footage, we understand very little about how neural networks operate.
In 2015, a team of computer vision researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) outlined a basic technique looking into the workings of a neural network trained to identify scenes from images. Now, the CSAIL team is preparing to present a fully automated version of the same system: a neural network trained to perform more than 20 computer vision tasks, including colourising black and white images and recognising objects.
The researchers adjusted the network in order to observe the strength with which individual nodes fire in response to training data. They were then able to identify the images that provokes the strongest response from each node.
“We catalogued 1,100 visual concepts - things like the colour green, or a swirly texture, or wood material, or a human face, or a bicycle wheel, or a snowy mountaintop,” said Mr David Bau, a PhD student in the CSAIL. “We drew on several data sets that other people had developed and merged them into a broadly and densely labelled data set of visual concepts.”
“It’s got many, many labels and for each label we know which pixels in which image correspond to that label.”
By tracing responses between different “layers” of the neural network – responsible for different stages of processing data – the researchers were also able to tell which specific pixels corresponded to the most powerful responses. Low-level layers would respond to simple visual properties such as colour and texture, while high-level layers would fire in response to more complex properties.
This allowed the CSAIL researchers to observe how networks devote different numbers of nods to different tasks.
Although every neural network is different, their work sheds light on how computer vision neural networks work beneath the surface, and could even be informative in the study of how the human brain is organised.
For instance, their work has some parallels with a disputed hypothesis in neuroscience called the grandmother-neuron hypothesis, which suggests that some neurons respond only to specific visual stimuli (such as your grandmother’s face). The researchers found that – for an artificial neural network – they could identify individual nodes that were tuned to particular visual concepts.
“To my eye, this is suggesting that neural networks are actually trying to approximate getting a grandmother neuron,” said Mr Bau. “They’re not trying to smear the idea of grandmother all over the place. They’re trying to assign it to a neuron. It’s this interesting hint of this structure that most people don’t believe is that simple.”