Xiao-Xiong Lin

@xxlin.bsky.social

61 Followers 128 Following

Computational and systems neuroscience. DeepRL Hippocampus navigation <- Data analysis / neural network modelling PFC working memory flexible cognition. xiaoxionglin.com https://www.bcf.uni-freiburg.de/about/people/lin github.com/xiaoxionglin/dSCA

place field map: white box with cross are obstacles. spatial kernel: measured by the correlation of population activity between pairs of spatial locations, averaged by the same displacement, distance, or orientation.
sparse input like in the DG->CA3 model. dense input where batch-normalization and high thresholding was removed. CA3: our CA3 model with L=64 and R=8. RandRNN: randomly initialized fixed RNN of the same state size. SSM_LegS: fixed SSM HiPPO-LegS from Gu et al. (2020) with the same state size. LSTM: trainable LSTM with matching number of parameters.
Virtual environment (19 × 19 tiles) were constructed using deepmind lab with walls randomly placed on 15 % of the tiles. Wall layouts are kept fixed for repeated trials, with an invisible reward near the bottom right. In each episode, the agent was initially placed at a random location at least 5 tiles away from the goal. The agent receives a first person perspective vision input that is processed via a visual encoder (shallow ResNet with 3 convolutional blocks; matching the SOTA in deepmind lab environment (Espeholt et al., 2018), pretrained and fixed in our experiments). These output was linearly mapped to F=16 features (FC: fully connected layer), and then sparsified using batch normalization and high thresholding (τ = 2.43), such that the percentage of activation (∼ 2.5%) matches the sparse activity of DG granular cells that project to CA3. CA3 is modeled as sequences of neurons each for a DG input feature. The activity of all CA3 neurons are then flattened and linearly mapped to the Decoder multilayer perceptron. The visual encoder is pretrained and fixed. CA3 is hard coded to isolate the effect of long range integration. The DG and decoder modules are trained.
Illustration of theta sequences observed in rodent hippocampus. In each theta cycle, R = 3 neurons are activated and the activation propagates over L = 4 theta cycles in a sequence of ℓ = L + R − 1 = 6 neurons. The theta sequences are thought to be driven by sequential inputs despite the recurrent connections in hippocampus. The recurrent connections could support generating long horizon sequential activity without sequential external inputs.