6 Neuroscience Papers you shouldn’t miss–NeurIPS 2018
Even though NeurIPS (formerly NIPS) is flooded with deep learning papers nowadays, you can still find the occasional paper on neuroscience and the brain. Here, I’ll present some of my favorite neuroscience papers from this year’s conference and explain what they can tell us about computations in the brain and beyond.
Dendritic cortical microcircuits approximate the backpropagation algorithm (Sacramento et al.)
This is my favorite paper from this year’s conference and one of the most exciting papers I’ve read recently. The authors aim to do nothing less than explaining how the neocortex (the part of the brain where basically all intelligent functions are located) might learn new tasks. Specifically, they model a multi-layered network of pyramidal neurons, the main computational unit in the neocortex (they also add a second class of neurons, so-called somatostatin-positive interneurons, which play a role in regulating the pyramidal neurons; see image above for a sketch of the network). Their network successfully learns image classification on MNIST without requiring explicit backpropagation (which is biologically implausible). Instead, the neurons transmit error signals to lower layers through feedback connections (which are abundant in the brain). Obviously, this paper is no proof that the brain actually learns in this way, but it’s definitely a cool proposal and shows that networks in the brain might be able to learn in a similar way to machine learning models.
Task-Driven Convolutional Recurrent Models of the Visual System (Nayebi et al.)
Like the paper by Sacramento et al. presented above, this work makes use of feedback connections, which seem to play a major role in information processing in the brain. However, instead of using these feedback connections to backpropagate error signals, here they are used to improve the feedforward computation. The authors design CNNs with recurrent and feedback connections (see red arrows in the image above) and train them on ImageNet. The networks achieve results comparable to state-of-the-art ResNets but have way fewer parameters and layers. This is exciting because it shows how the brain might make use of feedback connections to achieve good performance without requiring hundreds of layers. Also, the authors show that the features of their networks can accurately predict the responses of neurons in the visual cortex.
Gradient Descent for Spiking Neural Networks (Huh & Sejnowski)
One of the biggest differences between artificial and biological neural networks is how they transmit information: While artificial networks use continuous numbers, biological neurons communicate via discrete, all-or-none signals, so-called action potentials or spikes. Obviously, computing with spikes is a problem for gradient-based optimization because these signals are not differentiable. In this paper, the authors tackle this problem and introduce a gradient descent algorithm for spiking neural networks. This allows such networks to be trained with supervised learning, just in the same way artificial neural networks are trained today.
Incorporating Context into Language Encoding Models for fMRI (Jain & Huth)
A few years ago, several papers showed that CNNs are actually pretty good at predicting the activity of neurons in the visual system of the brain (see review in Yamins & DiCarlo 2016). This work shows the equivalent thing for natural language. The authors used fMRI (functional magnetic resonance imaging) to record the brain activity of humans while they listened to a text. Then, they computed word embeddings for the same text with an LSTM language model. By comparing the two representations (via ridge regression), they are able to show that the features in the word embeddings are pretty well aligned to the neural activity. Moreover, they find that different brain areas are modeled better or worse depending on the number of context words used in the language model. It seems like low-level sensory areas (e.g. auditory cortex; AC in the image above) are better represented by word embeddings with few context words, whereas high-level areas (e.g. in the prefrontal cortex; red blobs at the top of the image above) are better represented by word embeddings with more context words. This makes intuitive sense because these high-level areas should retain a knowledge of the entire text.
Generalisation in humans and deep neural networks (Geirhos et al.)
While neural networks achieve great results in computer vision, they often fail in unpredictable ways (e.g. when some minor noise is added, like in adversarial attacks). In this paper, the authors investigate the generalization performance of CNNs by adding different kinds of noise and distortions to the input images (see examples above). Then, they compare this performance to the visual system of the brain by carrying out experiments on human subjects. Not surprisingly, they find that humans are way more robust to all kinds of image distortions. While the paper does not propose a solution to this problem, it sheds light on an important issue and quantifies how machine learning models compare against human performance.
Toddler-Inspired Visual Object Learning (Bambach et al.)
Even though this is not a hardcore-neuroscience paper per se, I still want to mention it because it’s super interesting for human learning (and also, they do experiments with toddlers in a machine learning paper, how cool is that). The authors invited children and their parents to play with a bunch of toys. At the same time, they recorded their views with head-mounted cameras (see image above). By comparing the images recorded by the children and the images recorded by their parents, they find important differences in the image data, which might help the children to learn more efficiently. They also train CNNs for object recognition on this data and find that training on the children’s image actually produces better object models than training on the adult’s images.
