Collection of resources dealing with confluence of DL and neuro
-
Activations of Deep CNN and visual pathway regions - how brain regions responsible for visual pathway and the alenet layers capture "similar" features. They compared confusion matrix-type matrices between inputs from diff brain regions and alexnet layers. Uses Representation dissimalirity matrices(RDM) for each layer or probe characteristic ifentification and subsequent Representation similarity analysis(RSA) for quantification of mapping similarity between each probe and each layer of AlexNet.
-
Convolutional neural network-based encoding and decoding of visual object recognition in space and time - stimulus representations from CNNs can predict MEG source activity across visual system, both in space & time. Used encoding approach instead of RSA. Most regions are predicted well by low-level layers, whereas later in time, most regions are predicted well by high-level layers. Decoding results for some subjects are "above-chance".
-
Large-Scale, High-Resolution Comparison of the Core Visual Object Recognition Behavior of Humans, Monkeys, and State-of-the-Art Deep Artificial Neural Networks - Rajalingham, DiCarlo paper
-
Adversarial Examples that Fool both Computer Vision and Time-Limited Humans
-
A Unified Theory of Early Visual Representations from retina to cortex through Anatomically Constrained deep CNNs - Surya Ganguli and Stephane Deny paper
-
The emergence of multiple retinal cell types through efficient coding of natural movies - Surya Ganguli and Stephane Deny paper
-
A novel deep learning approach for classification of EEG motor imagery signals - Used a CNN for BCI competition dataset III and IV and got good accuracy. Greedily trained CNN layer-by-layer to achieve better accuracy.
-
Deep Learning with CNN for EEG decoding and visualization - Schirrmeister paper. Used several network architecutres. Good visualization techniques using correlation as a metric to understand features and CNN layer output relationships (not completely clear - read once more). CNNs don't outperform FBCSP clearly. Extensive experiments, results and discussions - good to refer for comparison. Discusses flexibility of CNNs vs simplicity of other models. Github repo Braindecode
-
Learning Representation from EEG with Deep RCNN - ICLR 2016 paper. Projet EEG electrode position from 3D space to 2D space and then form images with red = alpha channel activity, green = beta channel activity, blue = gamma channel activity. RCNN on this data to classify coginitive-load during task. DeconvNet to visualize filters learnt.
-
Learning Robust Features using Deep Learning for Automatic Seizure Detection - Uses the 3D to 2D projection of electrodes and then RCNN to predict seizure. Occluded image analysis to look at regions of image which affect decision of network.
-
Joint Optimization of Algorithmic Suites for EEG Analysis - EMBC 2014 paper. Uses backprop to jointly optimize spatial projection matrix (initialized with regular CSP), temporal projection matrix and a classifier. Results on BCI II and III datasets.
-
Deep Feature Learning for EEG recordings - Sebastian Stober paper. Classification of listened music tracks from EEG in OpenMIIR dataset. Used supervised CNN training, then CAEs (Conv AEs) to learn common signal components and compared the reconstruction loss to PCA and ICA components. Used similarity-constraint encoding and cross-trial encoding (Hydra-nets) to learn "neural" components of the signal (noise removal) and then used these features for classification. A brief presentation of the work here - ppt.
-
Deep Learning Human Mind for Automated Visual Classification - CVPR 2017 paper. Used LSTM cells to learn time-dependent features from EEG. Features extracted from Deep CNN for the images used for regressing the EEG features. Finally the regressed features used for predicitng which image class was the subject seeing. Best classification results for 320-480 ms after stimulus presentation (the last segment of EEG).
-
BrainNetCNN: Convolutional neural networks for brain networks; towards predicting neurodevelopment - Introduced edge-2-edge, edge-2-node and node-2-graph layers instead of the normal CNN layers to leverage the graph topography. Good introduction on graph methods and CNN on graphs. Validate on synthetic as well as real data. Real data from infants (168 DTI images) scanned between 27 and 45 weeks. BrainNetCNN age prediction MAE is ~2 weeks and average prediction error of ~11% on motor and cognitive scores. Used gradients on input data to visualize important connections in the brain that are predictive of age, motor scores and cognitive scores. Importantly, did SMOTE (synthetic minority over-sampling technique) to balance and augment training set.
-
Neural Encoding and Decoding with Deep Learning for Dynamic Natural Vision
-
Deep Neural Networks for Modeling Visual Perceptual Learning - Visual Percpetion Learning framework usign DNN. Interesting paper!!
- Theories of Error Back-Propagation in the Brain
- Towards an integration of Deep Learning and Neuroscience
- Dendritic error backpropagation in deep cortical microcircuits
- Deep Networks for motor control functions
- Using neuroscience to develop artificial intelligence - Ullman paper
- Dendritic solutions to the credit assignment problem - Blake and Tim paper
- A Critique of Pure Learning: What Artificial Neural Networks can Learn from Animal Brains