Please enable JavaScript.
Coggle requires JavaScript to display documents.
ANIE, pikachu-transparent-9 - Coggle Diagram
ANIE
RESULTS
Haxby
ANIE only 📈+ temp context
accuracy: 73.7% at 1tp - 79.2% at 20tp
F1 📈
baseline models :chart_with_downwards_trend:+ longer time windows
ANIE :crown: long-range temporal structure
Myawaki decoding
:muscle::skin-tone-2: performance
accuracy: 84.6% - 87.6%
:crown:: 10 tp
longer temp seq provide useful info for :eye: reconstruction
internal features learned
after
reducing to 2D
KNN classification
separate different stim types with over 93% accuracy
ANIE learning useful patterns in brain activity
Myawaki encoding
challenging bc predicting ks of voxels from just small bin image
longer time sequences stll :muscle::skin-tone-2: performance
R2: 0.2, pearson corr: 0.46
ARCHITECTURE
encoder: maps input data into latent space
ANIE op: models dependencies across time using ANIE
decoder: outputs class labels or reconstructed visual stimuli
INTRO
problem
brain act 📖 across space and time
difficult to model with 🗿
stim info - voxel activity: complex & distributed
relev info: mult tp
solution
deep learning framework: ANIE
goal: learn nonlocal rel directly from 🧠 signals
learn how evolve across time
DATASETS
Haxby
👁️ visual categories
faces
cats
bottles
houses
scissors
shoes
decode vox act to cassify 👁️
Miyawaki
👁️: 10x10 binary images
structured geometric patterns
random patterns
two tasks
decoding: model reconstructs 👁️ from fMRI voxel activity
encoding: model predicts vox responses from 👁️
